WEBVTT 00:00:00.000 --> 00:00:00.970 align:middle line:90% 00:00:00.970 --> 00:00:02.425 align:middle line:90% [SQUEAKING] 00:00:02.425 --> 00:00:03.880 align:middle line:90% [RUSTLING] 00:00:03.880 --> 00:00:06.305 align:middle line:90% [CLICKING] 00:00:06.305 --> 00:00:16.750 align:middle line:90% 00:00:16.750 --> 00:00:17.500 align:middle line:90% GLENN ELLISON: OK. 00:00:17.500 --> 00:00:18.833 align:middle line:90% Let me go ahead and get started. 00:00:18.833 --> 00:00:23.710 align:middle line:84% So today, continuing my Monday theory, Wednesday empiric stuff, 00:00:23.710 --> 00:00:27.610 align:middle line:84% I'm going to talk about empirics of dynamic competition. 00:00:27.610 --> 00:00:29.680 align:middle line:84% In this set of lectures, I say there's also-- 00:00:29.680 --> 00:00:33.160 align:middle line:84% there's reduced form and structural literatures 00:00:33.160 --> 00:00:35.230 align:middle line:84% on dynamic competition, and here I'm 00:00:35.230 --> 00:00:37.760 align:middle line:84% going to completely ignore the structural literature. 00:00:37.760 --> 00:00:41.350 align:middle line:84% It's just there's some technical things that come at the start 00:00:41.350 --> 00:00:44.560 align:middle line:84% of that are techniques, and so we've left that all 00:00:44.560 --> 00:00:45.590 align:middle line:90% for [14.]273. 00:00:45.590 --> 00:00:48.430 align:middle line:84% So those we do in [14.]273 in the spring, we'll see it there. 00:00:48.430 --> 00:00:51.680 align:middle line:84% But I would say that also it's one of the-- 00:00:51.680 --> 00:00:54.530 align:middle line:84% there's some neat structural stuff on dynamic competition, 00:00:54.530 --> 00:00:57.650 align:middle line:84% but it is a real challenge for structural estimation 00:00:57.650 --> 00:00:59.960 align:middle line:84% because a lot of structural estimation 00:00:59.960 --> 00:01:02.630 align:middle line:84% is built on there being a unique equilibrium 00:01:02.630 --> 00:01:05.060 align:middle line:84% and being able to invert going from the equilibrium back 00:01:05.060 --> 00:01:05.925 align:middle line:90% to the primitives. 00:01:05.925 --> 00:01:08.300 align:middle line:84% If you think about everything I said about repeated games 00:01:08.300 --> 00:01:11.870 align:middle line:84% when there are a zillion different equilibria, 00:01:11.870 --> 00:01:14.300 align:middle line:84% if the same primitives can lead to a thousand 00:01:14.300 --> 00:01:17.060 align:middle line:84% different equilibria, you can't necessarily go backwards from, 00:01:17.060 --> 00:01:20.330 align:middle line:84% here's what we observed to what the primitives are, 00:01:20.330 --> 00:01:22.530 align:middle line:84% and that causes challenges for that literature. 00:01:22.530 --> 00:01:25.710 align:middle line:84% But anyway, I will not get in there. 00:01:25.710 --> 00:01:29.120 align:middle line:84% So anyway, I'm going to talk about reduced form literature, 00:01:29.120 --> 00:01:32.780 align:middle line:84% which a lot of it is trying to provide evidence on, 00:01:32.780 --> 00:01:36.920 align:middle line:84% do the models that we've talked about previously apply 00:01:36.920 --> 00:01:39.793 align:middle line:84% and can we get anything about where they're working well 00:01:39.793 --> 00:01:41.210 align:middle line:84% and where they're not working well 00:01:41.210 --> 00:01:43.490 align:middle line:84% and what else we might want to put in them. 00:01:43.490 --> 00:01:46.190 align:middle line:84% I'm going to start with one of my favorite papers, 00:01:46.190 --> 00:01:48.320 align:middle line:84% Rob Porter's "Study of Cartel Stability: 00:01:48.320 --> 00:01:51.890 align:middle line:84% Joint Executive Committee, 1880-1886." 00:01:51.890 --> 00:01:53.900 align:middle line:84% You know, I have a number of classic papers 00:01:53.900 --> 00:01:55.650 align:middle line:84% in this course, a number of classic papers 00:01:55.650 --> 00:01:58.260 align:middle line:90% I omit from this syllabus. 00:01:58.260 --> 00:02:00.420 align:middle line:84% Porter's is a rare paper that you pick it up 00:02:00.420 --> 00:02:03.420 align:middle line:84% and it's 40 years old and somehow, 00:02:03.420 --> 00:02:06.090 align:middle line:84% 40 years later, it still just reads 00:02:06.090 --> 00:02:09.550 align:middle line:84% very much like a current paper, seems very well done. 00:02:09.550 --> 00:02:12.090 align:middle line:84% It's very thoughtful and there's hardly anything in it 00:02:12.090 --> 00:02:14.580 align:middle line:84% when you look at it and say, wow, that seems archaic, 00:02:14.580 --> 00:02:16.290 align:middle line:90% the way he's doing things here. 00:02:16.290 --> 00:02:22.902 align:middle line:84% So anyway, paper is about railroad cartel in the 1880s. 00:02:22.902 --> 00:02:24.360 align:middle line:84% Again, if you go back to the 1880s, 00:02:24.360 --> 00:02:27.510 align:middle line:84% you have to remember a lot of things had not yet been invented 00:02:27.510 --> 00:02:29.043 align:middle line:90% in 1880. 00:02:29.043 --> 00:02:30.960 align:middle line:84% The electric light bulb had not been invented. 00:02:30.960 --> 00:02:32.550 align:middle line:90% The car had not been invented. 00:02:32.550 --> 00:02:35.495 align:middle line:84% The paved road had not been invented. 00:02:35.495 --> 00:02:36.870 align:middle line:84% Obviously no need for paved roads 00:02:36.870 --> 00:02:38.490 align:middle line:90% if you hadn't invented the car. 00:02:38.490 --> 00:02:40.708 align:middle line:84% But so if you needed to transport goods, 00:02:40.708 --> 00:02:42.750 align:middle line:84% there were basically two ways to transport goods, 00:02:42.750 --> 00:02:46.200 align:middle line:84% by water or, then newly, by rail. 00:02:46.200 --> 00:02:50.280 align:middle line:84% And railroads became an incredibly important industry. 00:02:50.280 --> 00:02:52.975 align:middle line:84% If you look at the US stock market in the 1880s, 00:02:52.975 --> 00:02:55.350 align:middle line:84% I think the statistic is like railroads were roughly half 00:02:55.350 --> 00:02:57.250 align:middle line:84% the capitalization of the US stock market. 00:02:57.250 --> 00:03:00.120 align:middle line:84% So in some sense, half the firms by value in this country 00:03:00.120 --> 00:03:02.940 align:middle line:90% were railroads. 00:03:02.940 --> 00:03:05.370 align:middle line:84% One thing that was challenging for railroads 00:03:05.370 --> 00:03:08.190 align:middle line:84% is a lot of this class has been about what makes 00:03:08.190 --> 00:03:10.410 align:middle line:90% price above marginal cost. 00:03:10.410 --> 00:03:12.120 align:middle line:84% Well, if you've got a railroad, you 00:03:12.120 --> 00:03:13.890 align:middle line:84% need to have price above marginal cost 00:03:13.890 --> 00:03:17.580 align:middle line:84% because they have these enormous fixed costs of laying tracks 00:03:17.580 --> 00:03:20.400 align:middle line:84% to get from New York City to Chicago 00:03:20.400 --> 00:03:24.420 align:middle line:84% and putting down steel rails and clearing space on the whole path 00:03:24.420 --> 00:03:25.955 align:middle line:90% and acquiring the right of way. 00:03:25.955 --> 00:03:27.330 align:middle line:84% And then once you've done it, you 00:03:27.330 --> 00:03:30.560 align:middle line:84% can transport goods fairly cheaply. 00:03:30.560 --> 00:03:33.850 align:middle line:84% You can undercut the horse and cart by an enormous margin. 00:03:33.850 --> 00:03:39.130 align:middle line:84% But marginal cost of carrying the goods once you've 00:03:39.130 --> 00:03:41.120 align:middle line:84% built the railroad, are relatively low, 00:03:41.120 --> 00:03:44.770 align:middle line:84% and so if competition forces price down to marginal cost, 00:03:44.770 --> 00:03:47.500 align:middle line:84% you're never going to make the billions of dollars it cost you 00:03:47.500 --> 00:03:49.360 align:middle line:90% to build the rails themselves. 00:03:49.360 --> 00:03:51.400 align:middle line:84% So there were lots and lots of railroads who 00:03:51.400 --> 00:03:52.580 align:middle line:90% struggled to keep up prices. 00:03:52.580 --> 00:03:56.880 align:middle line:84% There were lots of railroad bankruptcies. 00:03:56.880 --> 00:04:01.630 align:middle line:84% So Porter's paper is studying a cartel. 00:04:01.630 --> 00:04:04.170 align:middle line:84% So all of these things which are not so visible, but these-- 00:04:04.170 --> 00:04:06.330 align:middle line:90% I took the old maps. 00:04:06.330 --> 00:04:10.950 align:middle line:84% These are all railroads that went from Chicago out here 00:04:10.950 --> 00:04:14.340 align:middle line:84% on the shore of lake Michigan to the East Coast. 00:04:14.340 --> 00:04:17.700 align:middle line:84% In 1880, there were three ways to get from Chicago to the East 00:04:17.700 --> 00:04:18.870 align:middle line:90% Coast. 00:04:18.870 --> 00:04:21.850 align:middle line:84% There was the New York Central railroad. 00:04:21.850 --> 00:04:23.940 align:middle line:84% This is the one owned by Vanderbilt. 00:04:23.940 --> 00:04:27.600 align:middle line:84% It went either this way or this way along the Great Lakes, 00:04:27.600 --> 00:04:29.910 align:middle line:84% through Albany, and then down to New York City 00:04:29.910 --> 00:04:31.740 align:middle line:90% to get to the East Coast. 00:04:31.740 --> 00:04:33.360 align:middle line:84% There was the Pennsylvania railroad 00:04:33.360 --> 00:04:36.000 align:middle line:84% that ran across Pennsylvania and got to Philadelphia 00:04:36.000 --> 00:04:37.890 align:middle line:84% and then went to places from Philadelphia, 00:04:37.890 --> 00:04:39.540 align:middle line:90% but got to the East Coast there. 00:04:39.540 --> 00:04:41.760 align:middle line:84% And then there was the Baltimore and Ohio railroad 00:04:41.760 --> 00:04:44.340 align:middle line:84% that started here, took a more southerly route 00:04:44.340 --> 00:04:47.400 align:middle line:84% across Pennsylvania, and ended up in Baltimore. 00:04:47.400 --> 00:04:49.155 align:middle line:84% And so from either of those three-- 00:04:49.155 --> 00:04:51.030 align:middle line:84% you could use either of those three railroads 00:04:51.030 --> 00:04:53.400 align:middle line:90% to get things to the East Coast. 00:04:53.400 --> 00:04:55.440 align:middle line:90% Jumping ahead a second. 00:04:55.440 --> 00:04:58.090 align:middle line:90% Why was this important? 00:04:58.090 --> 00:05:00.737 align:middle line:84% Railroads did carry passengers, but passenger travel 00:05:00.737 --> 00:05:02.570 align:middle line:84% was not where most of their money came from. 00:05:02.570 --> 00:05:04.695 align:middle line:84% Most of their money came from transporting freight, 00:05:04.695 --> 00:05:07.060 align:middle line:84% in particular, transporting grain. 00:05:07.060 --> 00:05:11.800 align:middle line:84% Then, as now, this is where corn grows in the United States. 00:05:11.800 --> 00:05:14.362 align:middle line:84% Corn and other grains tend to grow in the Midwest. 00:05:14.362 --> 00:05:15.820 align:middle line:84% It's much cheaper to produce things 00:05:15.820 --> 00:05:17.380 align:middle line:90% there than in New Hampshire. 00:05:17.380 --> 00:05:21.110 align:middle line:84% So all the grain was grown out here. 00:05:21.110 --> 00:05:23.410 align:middle line:84% It got to Chicago, and then you needed 00:05:23.410 --> 00:05:25.630 align:middle line:84% to get it from Chicago to the East Coast, which 00:05:25.630 --> 00:05:27.760 align:middle line:84% is where a lot of people lived in big cities 00:05:27.760 --> 00:05:29.380 align:middle line:90% and didn't have farms. 00:05:29.380 --> 00:05:32.170 align:middle line:84% And then in addition, a lot of the grain, 00:05:32.170 --> 00:05:37.360 align:middle line:84% when it reached the East Coast, actually went to Europe. 00:05:37.360 --> 00:05:40.400 align:middle line:84% So it would go by boat from the East Coast to Europe. 00:05:40.400 --> 00:05:43.060 align:middle line:84% And because it was going from the boat to Europe, 00:05:43.060 --> 00:05:46.300 align:middle line:84% in some sense, that equalized the prices at all the ports 00:05:46.300 --> 00:05:49.780 align:middle line:84% on the East Coast, because basically the European demand 00:05:49.780 --> 00:05:51.850 align:middle line:84% was big enough that roughly you didn't-- 00:05:51.850 --> 00:05:55.780 align:middle line:84% price getting to New York, Philadelphia, Baltimore 00:05:55.780 --> 00:05:56.950 align:middle line:90% was all roughly the same. 00:05:56.950 --> 00:05:59.475 align:middle line:90% 00:05:59.475 --> 00:06:01.770 align:middle line:84% So anyway, cartel had three members. 00:06:01.770 --> 00:06:02.613 align:middle line:90% What did they do? 00:06:02.613 --> 00:06:05.030 align:middle line:84% So there have been a couple of failed cartels before that, 00:06:05.030 --> 00:06:08.600 align:middle line:84% but cartels were still legal until 1887. 00:06:08.600 --> 00:06:12.710 align:middle line:84% So what the cartels did is they hired the premier cartel 00:06:12.710 --> 00:06:15.170 align:middle line:84% consultant of the day, Albert Fink. 00:06:15.170 --> 00:06:18.630 align:middle line:90% Albert Fink met with them. 00:06:18.630 --> 00:06:20.970 align:middle line:84% You know, these firms could have merged, 00:06:20.970 --> 00:06:23.987 align:middle line:84% but there are many reasons why you may not want to merge. 00:06:23.987 --> 00:06:25.570 align:middle line:84% Each one had their different railroad. 00:06:25.570 --> 00:06:27.400 align:middle line:84% They wanted to do different things with it. 00:06:27.400 --> 00:06:29.880 align:middle line:84% So the three firms continued to operate separately 00:06:29.880 --> 00:06:32.490 align:middle line:84% and they just hired a cartel consultant 00:06:32.490 --> 00:06:36.690 align:middle line:84% to arrange how their cartel was going to work. 00:06:36.690 --> 00:06:38.557 align:middle line:84% What do we know about what Albert Fink did? 00:06:38.557 --> 00:06:40.890 align:middle line:84% In some sense this was the secret to his business model, 00:06:40.890 --> 00:06:43.110 align:middle line:84% so he didn't publish exactly what it was that he 00:06:43.110 --> 00:06:45.600 align:middle line:90% did to maintain cartels. 00:06:45.600 --> 00:06:47.730 align:middle line:84% One thing we know he did with the railroad cartel 00:06:47.730 --> 00:06:51.355 align:middle line:84% is he had the option that observability was important, 00:06:51.355 --> 00:06:53.730 align:middle line:84% and he had the observation that observability is actually 00:06:53.730 --> 00:06:55.897 align:middle line:84% an easy thing to do in railroad because you can just 00:06:55.897 --> 00:06:58.230 align:middle line:90% hire minimum wage employees. 00:06:58.230 --> 00:07:00.370 align:middle line:84% Minimum wage employees were cheap in the 1880s. 00:07:00.370 --> 00:07:02.120 align:middle line:84% You could just hire minimum wage employees 00:07:02.120 --> 00:07:03.670 align:middle line:90% to count cars leaving Chicago. 00:07:03.670 --> 00:07:06.210 align:middle line:84% And so you can just see, trains have New York Central 00:07:06.210 --> 00:07:08.610 align:middle line:84% or Pennsylvania Railroad or Baltimore Ohio 00:07:08.610 --> 00:07:10.590 align:middle line:90% painted on the outside of them. 00:07:10.590 --> 00:07:13.500 align:middle line:84% You can just count how many carts full of grain 00:07:13.500 --> 00:07:16.500 align:middle line:84% they have leaving the rail yard and you know the quantities 00:07:16.500 --> 00:07:18.500 align:middle line:90% from all the firms precisely. 00:07:18.500 --> 00:07:20.770 align:middle line:84% What you don't know is whether the firms are cheating 00:07:20.770 --> 00:07:24.930 align:middle line:84% on their agreements and charging less than $0.30 per hundred 00:07:24.930 --> 00:07:27.880 align:middle line:84% pounds or whatever they've all agreed should be the rate 00:07:27.880 --> 00:07:30.110 align:middle line:84% for getting grain from Chicago to New York. 00:07:30.110 --> 00:07:33.255 align:middle line:84% So it's a situation where price is unobservable 00:07:33.255 --> 00:07:34.630 align:middle line:84% because you don't know if they've 00:07:34.630 --> 00:07:37.270 align:middle line:84% cut secret deals with people to carry their grain to the East 00:07:37.270 --> 00:07:38.750 align:middle line:90% Coast at a lower price. 00:07:38.750 --> 00:07:43.840 align:middle line:84% Quantity is clearly observable because you can just count cars. 00:07:43.840 --> 00:07:48.020 align:middle line:84% And those records that he has of his car counters survived, 00:07:48.020 --> 00:07:52.090 align:middle line:84% and so that's part of the basis of Porter's paper. 00:07:52.090 --> 00:07:52.720 align:middle line:90% OK. 00:07:52.720 --> 00:07:58.450 align:middle line:84% One other detail is the 1880s was before a lot of global 00:07:58.450 --> 00:07:59.120 align:middle line:90% warming. 00:07:59.120 --> 00:08:04.450 align:middle line:84% So it was not 68 degrees in November in the 1880s. 00:08:04.450 --> 00:08:08.650 align:middle line:84% In Newton, where I lived, there were lots of big ice harvesting 00:08:08.650 --> 00:08:11.770 align:middle line:84% businesses where people needed ice boxes or whatever, 00:08:11.770 --> 00:08:13.510 align:middle line:84% and they would, on the lakes in Newton, 00:08:13.510 --> 00:08:16.480 align:middle line:84% they would carve big foot thick chunks of ice 00:08:16.480 --> 00:08:18.490 align:middle line:90% and sell them in the winter. 00:08:18.490 --> 00:08:21.580 align:middle line:84% Obviously that doesn't work anymore. 00:08:21.580 --> 00:08:23.950 align:middle line:84% One thing that also didn't work then 00:08:23.950 --> 00:08:26.530 align:middle line:84% was you could also get grain out of Chicago 00:08:26.530 --> 00:08:31.000 align:middle line:84% by shipping it over the northern tip of Michigan down here 00:08:31.000 --> 00:08:33.490 align:middle line:84% and then through canals connecting 00:08:33.490 --> 00:08:36.610 align:middle line:84% the other great lakes out to the Saint Lawrence seaway. 00:08:36.610 --> 00:08:42.010 align:middle line:84% This was the cheaper way to get grain out of Chicago. 00:08:42.010 --> 00:08:44.410 align:middle line:84% There was a limited fleet of boats 00:08:44.410 --> 00:08:46.280 align:middle line:84% that would pass through the Great Lakes, 00:08:46.280 --> 00:08:48.940 align:middle line:84% and so in some sense, there's a fixed quantity that's 00:08:48.940 --> 00:08:50.470 align:middle line:84% going to go through the Great Lakes 00:08:50.470 --> 00:08:53.557 align:middle line:84% due to the capacity of the ships, but then that's fixed. 00:08:53.557 --> 00:08:55.390 align:middle line:84% It's always going to undercut the railroads, 00:08:55.390 --> 00:08:57.140 align:middle line:84% but it's just a limited quantity. 00:08:57.140 --> 00:09:00.940 align:middle line:84% And then importantly for this paper, pre-global warming, 00:09:00.940 --> 00:09:03.730 align:middle line:84% the Straits of Mackinaw were frozen for several months a year 00:09:03.730 --> 00:09:05.840 align:middle line:84% and were impassable for several months a year. 00:09:05.840 --> 00:09:08.350 align:middle line:84% So grains harvested in the fall, you want to get it out, 00:09:08.350 --> 00:09:11.358 align:middle line:84% but then in the winter, it's months 00:09:11.358 --> 00:09:13.150 align:middle line:84% in which the Straits of Mackinaw are closed 00:09:13.150 --> 00:09:14.580 align:middle line:84% and you can't get it out by ship. 00:09:14.580 --> 00:09:21.650 align:middle line:90% 00:09:21.650 --> 00:09:25.300 align:middle line:84% Any questions on the basics of this situation? 00:09:25.300 --> 00:09:31.550 align:middle line:84% So anyway, Porter's thought was that interested-- 00:09:31.550 --> 00:09:34.220 align:middle line:84% there's previous literature talking about price wars 00:09:34.220 --> 00:09:37.940 align:middle line:84% in this industry and failed attempts to collude. 00:09:37.940 --> 00:09:40.220 align:middle line:84% Porter's is the hypothesis that, no, 00:09:40.220 --> 00:09:42.560 align:middle line:84% this wasn't an industry with failed attempts to collude. 00:09:42.560 --> 00:09:44.810 align:middle line:84% These price wars that people were talking about 00:09:44.810 --> 00:09:47.870 align:middle line:84% could have been signs that Fink had already figured out 00:09:47.870 --> 00:09:50.060 align:middle line:84% the Green-Porter theory of collusion 00:09:50.060 --> 00:09:52.910 align:middle line:84% under imperfect observability, and what we were observing 00:09:52.910 --> 00:09:56.840 align:middle line:84% is the functioning of a Green-Porter style cartel. 00:09:56.840 --> 00:10:01.730 align:middle line:84% That's what he takes out to investigate. 00:10:01.730 --> 00:10:03.950 align:middle line:84% Starting point for his paper was then 00:10:03.950 --> 00:10:07.190 align:middle line:84% thinking about what would we observe 00:10:07.190 --> 00:10:09.440 align:middle line:84% if there was a functioning Green-Porter cartel 00:10:09.440 --> 00:10:10.560 align:middle line:90% in this industry. 00:10:10.560 --> 00:10:12.140 align:middle line:84% And so he starts with this assumption 00:10:12.140 --> 00:10:16.550 align:middle line:84% that demand takes a log linear form, log of qt, 00:10:16.550 --> 00:10:18.050 align:middle line:90% quantity demanded at-- 00:10:18.050 --> 00:10:22.760 align:middle line:84% and so these t's are going to be weeks in his in his paper. 00:10:22.760 --> 00:10:25.820 align:middle line:84% Log of quantity in a week as a constant minus 00:10:25.820 --> 00:10:29.640 align:middle line:84% alpha times the log of the price plus alpha 2 00:10:29.640 --> 00:10:32.800 align:middle line:90% times the LAKES dummy plus u1t. 00:10:32.800 --> 00:10:36.600 align:middle line:84% So this is a dummy for whether the Great Lakes are frozen, 00:10:36.600 --> 00:10:40.260 align:middle line:84% which he takes as just an exogenous shift in demand. 00:10:40.260 --> 00:10:41.250 align:middle line:90% OK. 00:10:41.250 --> 00:10:44.670 align:middle line:84% Let me say that's a better way to do it in a-- 00:10:44.670 --> 00:10:46.980 align:middle line:84% it would be better to do that in a linear demand model. 00:10:46.980 --> 00:10:50.430 align:middle line:84% In a linear demand model you would say there's some quantity, 00:10:50.430 --> 00:10:53.730 align:middle line:84% like 3,000 tons a week that disappears when the Great Lakes 00:10:53.730 --> 00:10:56.410 align:middle line:84% are open, and that we're subtracting. 00:10:56.410 --> 00:11:00.580 align:middle line:84% Why we're doing it additively in a log model is less clear, 00:11:00.580 --> 00:11:02.670 align:middle line:84% but anyway, that's what he what he's able 00:11:02.670 --> 00:11:03.680 align:middle line:90% to do in the log model. 00:11:03.680 --> 00:11:07.260 align:middle line:90% 00:11:07.260 --> 00:11:10.530 align:middle line:84% And then he says, what's the cartel going to do? 00:11:10.530 --> 00:11:12.720 align:middle line:84% I know that our basic markup formula 00:11:12.720 --> 00:11:16.890 align:middle line:84% is price minus cost over price is minus 00:11:16.890 --> 00:11:19.230 align:middle line:90% theta over the elasticity. 00:11:19.230 --> 00:11:21.870 align:middle line:84% And he's going to suppose that a cartel is doing 00:11:21.870 --> 00:11:24.340 align:middle line:90% like what we said cartels do. 00:11:24.340 --> 00:11:25.870 align:middle line:84% They set high prices for a while. 00:11:25.870 --> 00:11:28.050 align:middle line:84% They jump down, they set low prices, they go up 00:11:28.050 --> 00:11:29.530 align:middle line:90% and they set high prices. 00:11:29.530 --> 00:11:31.660 align:middle line:90% They do this. 00:11:31.660 --> 00:11:33.650 align:middle line:84% And he has two different mark ups. 00:11:33.650 --> 00:11:37.510 align:middle line:84% He has theta equals theta c in cooperative periods and theta 00:11:37.510 --> 00:11:40.160 align:middle line:84% equals theta w in price war periods. 00:11:40.160 --> 00:11:43.360 align:middle line:84% And you might think that if you're setting the monopoly 00:11:43.360 --> 00:11:46.450 align:middle line:84% price in the high periods, you do price minus cost over price 00:11:46.450 --> 00:11:49.400 align:middle line:84% equals one over the elasticity, so theta would be one. 00:11:49.400 --> 00:11:52.450 align:middle line:84% And in the price war phase, if you have go to pricing at cost, 00:11:52.450 --> 00:11:57.460 align:middle line:84% it's 0 over the elasticity so theta would be zero. 00:11:57.460 --> 00:12:00.280 align:middle line:84% That's not exactly what an optimal cartel does, 00:12:00.280 --> 00:12:03.150 align:middle line:84% though, because if you think about the first order condition, 00:12:03.150 --> 00:12:06.190 align:middle line:84% d pi dp is equal to zero when you're setting prices 00:12:06.190 --> 00:12:07.850 align:middle line:90% at the monopoly price. 00:12:07.850 --> 00:12:12.340 align:middle line:84% And so because d pi dp is zero, cutting prices slightly 00:12:12.340 --> 00:12:14.890 align:middle line:84% makes it easier to deter collusion 00:12:14.890 --> 00:12:18.260 align:middle line:84% and has no first order effect on your per period profits. 00:12:18.260 --> 00:12:19.930 align:middle line:84% So in an optimal Green-Porter cartel, 00:12:19.930 --> 00:12:22.660 align:middle line:84% you always want theta somewhat less than one 00:12:22.660 --> 00:12:25.330 align:middle line:84% because it just makes it easier to deter colluding 00:12:25.330 --> 00:12:30.850 align:middle line:84% and doesn't do anything to your profits to first order. 00:12:30.850 --> 00:12:35.180 align:middle line:84% And again, the theta equals theta w during price wars. 00:12:35.180 --> 00:12:37.010 align:middle line:84% I said that simplest Green-Porter model 00:12:37.010 --> 00:12:39.540 align:middle line:84% you would want theta equals zero, you just price at cost. 00:12:39.540 --> 00:12:43.760 align:middle line:84% But a [INAUDIBLE] might note you could go to theta below cost 00:12:43.760 --> 00:12:44.990 align:middle line:90% or you could-- 00:12:44.990 --> 00:12:47.695 align:middle line:84% maybe you don't go all the way to cost anyway. 00:12:47.695 --> 00:12:49.148 align:middle line:90% But we think theta is going to-- 00:12:49.148 --> 00:12:50.940 align:middle line:84% this one may be a little bit less than one, 00:12:50.940 --> 00:12:56.400 align:middle line:84% this one may be about zero would be the natural assumption. 00:12:56.400 --> 00:12:58.290 align:middle line:84% And then he says, well, OK, let me-- 00:12:58.290 --> 00:12:59.190 align:middle line:90% I've got a demand. 00:12:59.190 --> 00:13:01.850 align:middle line:90% I've assumed a demand curve. 00:13:01.850 --> 00:13:05.460 align:middle line:84% Let me also assume a functional form for marginal costs. 00:13:05.460 --> 00:13:07.730 align:middle line:90% He picks this one. 00:13:07.730 --> 00:13:14.390 align:middle line:84% So marginal cost is a constant times quantity to the beta. 00:13:14.390 --> 00:13:15.890 align:middle line:84% If there are increasing costs, you 00:13:15.890 --> 00:13:18.530 align:middle line:84% think beta is going to be slightly bigger than one. 00:13:18.530 --> 00:13:20.690 align:middle line:84% The more quantity we have to carry in a month, 00:13:20.690 --> 00:13:22.800 align:middle line:90% the higher is our marginal cost. 00:13:22.800 --> 00:13:25.100 align:middle line:84% So our quantity is being slightly higher than zero. 00:13:25.100 --> 00:13:28.430 align:middle line:84% So beta equals zero would be constant marginal costs that 00:13:28.430 --> 00:13:29.600 align:middle line:90% are independent of quantity. 00:13:29.600 --> 00:13:32.150 align:middle line:84% Beta slightly greater than zero is marginal cost 00:13:32.150 --> 00:13:34.160 align:middle line:90% increased with quantity. 00:13:34.160 --> 00:13:37.880 align:middle line:84% And then a random shock in every period. 00:13:37.880 --> 00:13:42.650 align:middle line:84% The random shock could reflect fuel costs, labor costs, 00:13:42.650 --> 00:13:46.640 align:middle line:84% difficulty of getting-- your equipment not being in place 00:13:46.640 --> 00:13:49.570 align:middle line:90% to carry a load, needing to-- 00:13:49.570 --> 00:13:50.817 align:middle line:90% whatever. 00:13:50.817 --> 00:13:52.650 align:middle line:84% You may have different quality of equipment, 00:13:52.650 --> 00:13:54.980 align:middle line:84% and so you have some weeks you have to go to cheap 00:13:54.980 --> 00:13:59.230 align:middle line:84% to run equipment is available, some weeks it's not. 00:13:59.230 --> 00:14:00.820 align:middle line:84% And so he says, what would happen 00:14:00.820 --> 00:14:02.685 align:middle line:90% if we had those marginal costs? 00:14:02.685 --> 00:14:04.060 align:middle line:84% Well, you would get this equation 00:14:04.060 --> 00:14:08.790 align:middle line:84% price minus marginal cost over price equals this. 00:14:08.790 --> 00:14:11.650 align:middle line:84% And so then you just solve that equation for price. 00:14:11.650 --> 00:14:15.120 align:middle line:84% I put price over to this side and I get price times theta 00:14:15.120 --> 00:14:16.720 align:middle line:90% over alpha 1. 00:14:16.720 --> 00:14:21.880 align:middle line:84% So I have price times 1 minus theta over alpha 1 equals this, 00:14:21.880 --> 00:14:24.900 align:middle line:84% and so I get price equals 1 minus theta over alpha 00:14:24.900 --> 00:14:25.835 align:middle line:90% 1 times that. 00:14:25.835 --> 00:14:29.950 align:middle line:90% 00:14:29.950 --> 00:14:32.138 align:middle line:84% OK, so this is how a firm would price, again, 00:14:32.138 --> 00:14:34.305 align:middle line:84% using different thetas in the two different periods. 00:14:34.305 --> 00:14:42.050 align:middle line:90% 00:14:42.050 --> 00:14:42.550 align:middle line:90% OK. 00:14:42.550 --> 00:14:44.980 align:middle line:84% This may look like you've just got a big mess, 00:14:44.980 --> 00:14:49.210 align:middle line:84% but then take logs of both sides, and what you get 00:14:49.210 --> 00:14:52.560 align:middle line:90% is log of p on the left side. 00:14:52.560 --> 00:14:56.070 align:middle line:84% The right side, I get some terms that-- 00:14:56.070 --> 00:14:59.310 align:middle line:84% let me just do this first for your inner price war. 00:14:59.310 --> 00:15:01.320 align:middle line:90% So this is theta w. 00:15:01.320 --> 00:15:08.490 align:middle line:84% I first get some terms that I would get these terms, 00:15:08.490 --> 00:15:14.970 align:middle line:84% log gamma 0 minus log of 1 minus theta over alpha 1, 00:15:14.970 --> 00:15:17.543 align:middle line:84% and then I just get beta 1 log q. 00:15:17.543 --> 00:15:22.040 align:middle line:90% 00:15:22.040 --> 00:15:24.890 align:middle line:84% And then I get log of this is just u2. 00:15:24.890 --> 00:15:32.010 align:middle line:90% 00:15:32.010 --> 00:15:34.650 align:middle line:84% And so then what Porter says is that-- so this 00:15:34.650 --> 00:15:38.730 align:middle line:84% is what I would get in a price war phase. 00:15:38.730 --> 00:15:41.100 align:middle line:84% I would get that if you had a price war. 00:15:41.100 --> 00:15:43.950 align:middle line:84% And then if you were in the cooperative phase, 00:15:43.950 --> 00:15:52.050 align:middle line:84% I'd get log gamma 0 minus log 1 minus theta c over alpha 1 00:15:52.050 --> 00:15:57.810 align:middle line:90% plus beta 1 log q plus u2t. 00:15:57.810 --> 00:16:00.790 align:middle line:84% That's what you'd get if you were cooperating. 00:16:00.790 --> 00:16:03.790 align:middle line:84% And so what Porter notes is that this is just a linear demand 00:16:03.790 --> 00:16:08.200 align:middle line:84% curve where log of pt would be a constant, 00:16:08.200 --> 00:16:11.770 align:middle line:84% where the constant is this thing you get in the price war phase, 00:16:11.770 --> 00:16:17.410 align:middle line:84% plus beta 1 log qt plus u2t, plus another constant times 00:16:17.410 --> 00:16:20.350 align:middle line:84% an indicator for whether you're in the cooperative phase. 00:16:20.350 --> 00:16:24.640 align:middle line:84% And the cooperative phase just shifts you from this to this 00:16:24.640 --> 00:16:28.690 align:middle line:84% by adding log 1 minus theta w over alpha 1 00:16:28.690 --> 00:16:33.140 align:middle line:84% and subtracting 1 minus log theta c over alpha 1. 00:16:33.140 --> 00:16:39.360 align:middle line:84% So what he says is that if you had this-- 00:16:39.360 --> 00:16:40.630 align:middle line:90% again, two assumptions. 00:16:40.630 --> 00:16:44.020 align:middle line:84% You have this log linear form for demand 00:16:44.020 --> 00:16:48.190 align:middle line:84% and you have this particular form for marginal costs, 00:16:48.190 --> 00:16:53.120 align:middle line:84% then this is the appropriate supply equation. 00:16:53.120 --> 00:16:56.710 align:middle line:84% And it's going to also be linear. 00:16:56.710 --> 00:16:58.180 align:middle line:84% And so then the neat thing he's got 00:16:58.180 --> 00:17:01.000 align:middle line:84% is that basically this assumption of Green-Porter 00:17:01.000 --> 00:17:06.369 align:middle line:84% or cartel plus log linear demand gives us just a simple two 00:17:06.369 --> 00:17:08.800 align:middle line:90% by two supply demand system. 00:17:08.800 --> 00:17:11.960 align:middle line:84% So demand is what he assumed it is. 00:17:11.960 --> 00:17:14.800 align:middle line:84% Supply is what's derived like that. 00:17:14.800 --> 00:17:19.480 align:middle line:84% And if you think about the way that system is working, 00:17:19.480 --> 00:17:21.010 align:middle line:90% I think Tobias-- 00:17:21.010 --> 00:17:22.810 align:middle line:84% or you've seen before, you're trying 00:17:22.810 --> 00:17:24.109 align:middle line:90% to estimate supply and demand. 00:17:24.109 --> 00:17:25.099 align:middle line:90% You need an instrument. 00:17:25.099 --> 00:17:27.948 align:middle line:84% You need a variable that's excluded from demand to estimate 00:17:27.948 --> 00:17:30.490 align:middle line:84% the demand curve and you need a variable excluded from supply 00:17:30.490 --> 00:17:32.070 align:middle line:90% to estimate the supply curve. 00:17:32.070 --> 00:17:37.180 align:middle line:84% And so what he gets is that we have a supply demand system. 00:17:37.180 --> 00:17:39.700 align:middle line:84% The indicator for whether we're in a cooperative phase 00:17:39.700 --> 00:17:43.160 align:middle line:84% or a price war phase doesn't affect demand. 00:17:43.160 --> 00:17:45.890 align:middle line:84% This is a very common thing to do in IO. 00:17:45.890 --> 00:17:47.080 align:middle line:90% It is a great assumption. 00:17:47.080 --> 00:17:50.770 align:middle line:84% You're looking for what affects demand 00:17:50.770 --> 00:17:52.750 align:middle line:84% without affecting price, what affects pricing 00:17:52.750 --> 00:17:55.780 align:middle line:84% without affecting demand, whether the firm 00:17:55.780 --> 00:17:58.850 align:middle line:84% behavior determines their prices but it isn't directly 00:17:58.850 --> 00:18:01.680 align:middle line:84% affecting demand, other than through the price change. 00:18:01.680 --> 00:18:04.190 align:middle line:84% So if the firms are colluding or not colluding, 00:18:04.190 --> 00:18:06.410 align:middle line:84% that affect shifts price, but that 00:18:06.410 --> 00:18:08.430 align:middle line:84% doesn't have an independent effect on demand. 00:18:08.430 --> 00:18:10.805 align:middle line:84% And then what does he have that has an independent effect 00:18:10.805 --> 00:18:12.330 align:middle line:90% on demand but not on price? 00:18:12.330 --> 00:18:14.370 align:middle line:90% The Great Lakes being frozen. 00:18:14.370 --> 00:18:17.180 align:middle line:84% So if you take this view that the Great Lakes are just 00:18:17.180 --> 00:18:19.160 align:middle line:84% pulling some amount of traffic out, 00:18:19.160 --> 00:18:20.660 align:middle line:84% but then once they've pulled it out, 00:18:20.660 --> 00:18:23.077 align:middle line:84% the remaining traffic is just the same residual demand you 00:18:23.077 --> 00:18:26.780 align:middle line:84% would have been facing before, it's not affecting your-- 00:18:26.780 --> 00:18:29.140 align:middle line:84% in this model, it's not affecting your pricing. 00:18:29.140 --> 00:18:31.800 align:middle line:90% 00:18:31.800 --> 00:18:34.750 align:middle line:84% So anyway, we can use i as the instrument for price 00:18:34.750 --> 00:18:36.420 align:middle line:84% in the demand equation, we can use lakes 00:18:36.420 --> 00:18:39.300 align:middle line:84% as the instrument for quantity in the supply equation, 00:18:39.300 --> 00:18:40.625 align:middle line:90% and estimate supply and demand. 00:18:40.625 --> 00:18:47.920 align:middle line:90% 00:18:47.920 --> 00:18:50.050 align:middle line:84% Actually, as I said, this argument 00:18:50.050 --> 00:18:52.460 align:middle line:84% is better if you use linear demand curves. 00:18:52.460 --> 00:18:56.350 align:middle line:84% So this week's problem set, it's a largely empirical problem 00:18:56.350 --> 00:18:59.170 align:middle line:90% set if you haven't seen it yet. 00:18:59.170 --> 00:19:02.410 align:middle line:84% I'm giving you the data from Porter's actual paper 00:19:02.410 --> 00:19:04.210 align:middle line:84% and having you pretend you're Rob Porter 00:19:04.210 --> 00:19:06.670 align:middle line:84% and you think a linear demand specification would be better 00:19:06.670 --> 00:19:09.290 align:middle line:84% than a log linear demand specification, 00:19:09.290 --> 00:19:11.170 align:middle line:84% because in the linear demand specification, 00:19:11.170 --> 00:19:14.110 align:middle line:84% subtracting the LAKES thing is a better thing to do 00:19:14.110 --> 00:19:17.770 align:middle line:84% and it's asking you to redo his paper had you 00:19:17.770 --> 00:19:21.790 align:middle line:84% wanted to assume linear demand instead of the log demand. 00:19:21.790 --> 00:19:25.730 align:middle line:84% Or at least discuss how you would estimate the model. 00:19:25.730 --> 00:19:26.230 align:middle line:90% OK. 00:19:26.230 --> 00:19:28.180 align:middle line:90% Anyway, so Porter's paper. 00:19:28.180 --> 00:19:32.270 align:middle line:84% Again, big data set for 1880 is a small data set for today. 00:19:32.270 --> 00:19:36.670 align:middle line:90% It's 328 weekly observations. 00:19:36.670 --> 00:19:39.250 align:middle line:84% What you've got in the data set here 00:19:39.250 --> 00:19:43.370 align:middle line:84% is total shipments of grain for the week, the official price. 00:19:43.370 --> 00:19:47.510 align:middle line:84% So Albert Fink himself announced what the price was. 00:19:47.510 --> 00:19:50.470 align:middle line:84% So the price is $0.30 per hundred pounds. 00:19:50.470 --> 00:19:52.120 align:middle line:84% That was part of his responsibility 00:19:52.120 --> 00:19:55.575 align:middle line:84% was just to tell the firms what price to charge every week. 00:19:55.575 --> 00:19:58.270 align:middle line:90% 00:19:58.270 --> 00:20:00.700 align:middle line:84% LAKES as a dummy equal to one if the Great Lakes 00:20:00.700 --> 00:20:02.300 align:middle line:90% were open for shipping. 00:20:02.300 --> 00:20:04.720 align:middle line:84% So someone recorded the dates at which these lakes 00:20:04.720 --> 00:20:07.240 align:middle line:90% froze and unfroze every year. 00:20:07.240 --> 00:20:12.490 align:middle line:84% And then when railroads are half the economy, there were-- 00:20:12.490 --> 00:20:14.240 align:middle line:84% if you read the Wall Street Journal-- 00:20:14.240 --> 00:20:16.240 align:middle line:84% I don't know if the Wall Street Journal existed, 00:20:16.240 --> 00:20:18.460 align:middle line:84% but half of the articles would be on railroads 00:20:18.460 --> 00:20:21.080 align:middle line:84% because they were half of the-- half of the economy. 00:20:21.080 --> 00:20:23.800 align:middle line:84% And so there's plenty of newspapers-- 00:20:23.800 --> 00:20:28.040 align:middle line:84% there's actually a newspaper he uses called-- 00:20:28.040 --> 00:20:31.600 align:middle line:90% I think it's Railway Review. 00:20:31.600 --> 00:20:34.180 align:middle line:84% So besides being daily newspapers on business, 00:20:34.180 --> 00:20:37.220 align:middle line:84% there were daily newspapers on the railroad industry. 00:20:37.220 --> 00:20:39.700 align:middle line:84% And so you could buy a newspaper called the Railway Review 00:20:39.700 --> 00:20:42.460 align:middle line:90% and read today's railroad news. 00:20:42.460 --> 00:20:45.873 align:middle line:84% Harvard Business School library has copies of Railway Review. 00:20:45.873 --> 00:20:48.040 align:middle line:84% I don't know if you can get in there with an MIT ID, 00:20:48.040 --> 00:20:48.710 align:middle line:90% I think not. 00:20:48.710 --> 00:20:50.740 align:middle line:84% But anyway, when I was a graduate student, 00:20:50.740 --> 00:20:53.890 align:middle line:84% I did go there at one point and look at the old railway 00:20:53.890 --> 00:20:55.945 align:middle line:90% reviews there in Baker Library. 00:20:55.945 --> 00:21:00.080 align:middle line:90% 00:21:00.080 --> 00:21:04.180 align:middle line:84% Anyway, he recorded whether there were stories 00:21:04.180 --> 00:21:06.070 align:middle line:84% in the Railway Review indicating that a price 00:21:06.070 --> 00:21:07.630 align:middle line:84% war is occurring in the industry, 00:21:07.630 --> 00:21:09.670 align:middle line:90% in the railroad industry. 00:21:09.670 --> 00:21:13.300 align:middle line:84% And so if the newspaper had stories about price wars going 00:21:13.300 --> 00:21:15.580 align:middle line:84% on in the railroad industry, then he 00:21:15.580 --> 00:21:17.980 align:middle line:84% would report it as being one, otherwise 00:21:17.980 --> 00:21:19.760 align:middle line:84% he would record it as being zero. 00:21:19.760 --> 00:21:23.560 align:middle line:84% And so in some sense he used this as his cooperative phase 00:21:23.560 --> 00:21:28.810 align:middle line:84% is no story in Railway Review about price war. 00:21:28.810 --> 00:21:31.690 align:middle line:84% Price war phase is, Railway Review says price war going 00:21:31.690 --> 00:21:34.920 align:middle line:90% on in the JEC cartel this week. 00:21:34.920 --> 00:21:39.290 align:middle line:90% 00:21:39.290 --> 00:21:43.490 align:middle line:84% So we get two ways to estimate this model. 00:21:43.490 --> 00:21:46.910 align:middle line:84% One, the simplest one, is just assume 00:21:46.910 --> 00:21:50.330 align:middle line:84% that i is observe-- the cooperation indicator's 00:21:50.330 --> 00:21:52.490 align:middle line:84% observable and just use the PO variable 00:21:52.490 --> 00:21:54.630 align:middle line:84% that he got from the newspaper to do this. 00:21:54.630 --> 00:21:58.870 align:middle line:84% You then just have a standard simultaneous equation model. 00:21:58.870 --> 00:22:01.330 align:middle line:84% The other approach that he uses is 00:22:01.330 --> 00:22:05.290 align:middle line:90% to assume that It is unobserved. 00:22:05.290 --> 00:22:09.280 align:middle line:84% So it's an unobserved random variable with some mean gamma, 00:22:09.280 --> 00:22:11.530 align:middle line:84% which is to be estimated, and then you 00:22:11.530 --> 00:22:16.000 align:middle line:84% have a simultaneous version of a switching regressions model. 00:22:16.000 --> 00:22:18.550 align:middle line:84% Have switching regressions model entirely disappeared 00:22:18.550 --> 00:22:21.730 align:middle line:90% from econometric sequence? 00:22:21.730 --> 00:22:22.330 align:middle line:90% Yeah? 00:22:22.330 --> 00:22:22.880 align:middle line:90% OK. 00:22:22.880 --> 00:22:26.860 align:middle line:84% So OK, how does this switching regressions model work? 00:22:26.860 --> 00:22:29.650 align:middle line:90% So suppose you have the-- 00:22:29.650 --> 00:22:36.430 align:middle line:84% obviously our standard linear model, 00:22:36.430 --> 00:22:40.750 align:middle line:84% like y equals x beta plus epsilon, I put x here. 00:22:40.750 --> 00:22:49.050 align:middle line:84% There's some relationship Like y equals x times beta. 00:22:49.050 --> 00:22:52.570 align:middle line:84% And what you expect to see is a cloud of points like that. 00:22:52.570 --> 00:22:53.070 align:middle line:90% OK? 00:22:53.070 --> 00:22:55.530 align:middle line:84% So you've got y equals a function of x. 00:22:55.530 --> 00:22:56.720 align:middle line:90% That's what you observe. 00:22:56.720 --> 00:23:01.280 align:middle line:84% You can fit a line by OLS going through that. 00:23:01.280 --> 00:23:02.990 align:middle line:90% We do that all the time. 00:23:02.990 --> 00:23:04.490 align:middle line:90% Suppose you have the following. 00:23:04.490 --> 00:23:07.590 align:middle line:84% Suppose there are two different relationships. 00:23:07.590 --> 00:23:12.080 align:middle line:84% There's y equals x beta plus epsilon, 00:23:12.080 --> 00:23:14.520 align:middle line:84% and there's also another relationship, 00:23:14.520 --> 00:23:16.473 align:middle line:84% y equals x beta prime plus epsilon. 00:23:16.473 --> 00:23:19.040 align:middle line:90% 00:23:19.040 --> 00:23:24.560 align:middle line:84% And this happens with probability 1 minus gamma, 00:23:24.560 --> 00:23:28.850 align:middle line:84% and this happens with probability gamma. 00:23:28.850 --> 00:23:29.350 align:middle line:90% OK? 00:23:29.350 --> 00:23:33.670 align:middle line:84% So suppose there are actually two different relationships. 00:23:33.670 --> 00:23:37.167 align:middle line:84% For instance, if you were doing demand for ice cream on a beach, 00:23:37.167 --> 00:23:38.750 align:middle line:84% there are two different relationships, 00:23:38.750 --> 00:23:42.280 align:middle line:84% demand when it's rainy and demand when it's not rainy. 00:23:42.280 --> 00:23:44.423 align:middle line:84% You may not have raininess, but you 00:23:44.423 --> 00:23:46.090 align:middle line:84% know that there's going to be one demand 00:23:46.090 --> 00:23:48.633 align:middle line:84% curve on rainy days, one demand curve on non rainy days, 00:23:48.633 --> 00:23:50.050 align:middle line:84% and they're going to be different. 00:23:50.050 --> 00:23:56.230 align:middle line:84% And it may be that this is the demand curve on the non-rainy 00:23:56.230 --> 00:23:57.657 align:middle line:90% days as a function of price-- 00:23:57.657 --> 00:23:59.240 align:middle line:84% I guess the price should be positive-- 00:23:59.240 --> 00:24:00.910 align:middle line:84% and then on the sunny hot days, there's 00:24:00.910 --> 00:24:06.400 align:middle line:84% some other relationship that looks like this. 00:24:06.400 --> 00:24:09.060 align:middle line:84% and you might ask yourself, OK, suppose it is just-- 00:24:09.060 --> 00:24:11.590 align:middle line:84% suppose there's one x, but then there's also a constant, 00:24:11.590 --> 00:24:13.320 align:middle line:84% and suppose that it really is just 00:24:13.320 --> 00:24:16.410 align:middle line:84% this is the y equals x prime beta plus epsilon, 00:24:16.410 --> 00:24:19.200 align:middle line:84% and the constant is higher on the sunny days and demand 00:24:19.200 --> 00:24:20.500 align:middle line:90% looks like that. 00:24:20.500 --> 00:24:21.000 align:middle line:90% OK. 00:24:21.000 --> 00:24:24.920 align:middle line:84% Question is, can you estimate that? 00:24:24.920 --> 00:24:25.850 align:middle line:90% OK? 00:24:25.850 --> 00:24:28.520 align:middle line:84% First answer is without any additional assumptions 00:24:28.520 --> 00:24:32.510 align:middle line:84% on-- if you just go with a standard expectation of epsilon 00:24:32.510 --> 00:24:38.410 align:middle line:84% given x equals zero assumption, you can't estimate that. 00:24:38.410 --> 00:24:40.060 align:middle line:84% And the reason you can't estimate 00:24:40.060 --> 00:24:42.550 align:middle line:84% that is you might think, well, can't I just 00:24:42.550 --> 00:24:46.450 align:middle line:84% draw two best fit lines and ask what combination of two lines 00:24:46.450 --> 00:24:50.750 align:middle line:84% minimizes the sum of squared residuals from the closer line, 00:24:50.750 --> 00:24:54.670 align:middle line:84% and isn't that an efficient estimator of this model? 00:24:54.670 --> 00:24:58.750 align:middle line:84% But the problem is you can't tell this model apart 00:24:58.750 --> 00:25:02.560 align:middle line:84% from a model where there's one line 00:25:02.560 --> 00:25:06.100 align:middle line:84% and the epsilons have this distribution. 00:25:06.100 --> 00:25:08.920 align:middle line:84% If you're only doing mean zero errors, 00:25:08.920 --> 00:25:14.620 align:middle line:84% you can't tell two lines with this distribution of errors 00:25:14.620 --> 00:25:19.570 align:middle line:84% apart from one line with this distribution of errors. 00:25:19.570 --> 00:25:23.620 align:middle line:84% And so without some assumption beyond expectation 00:25:23.620 --> 00:25:25.540 align:middle line:84% of epsilon given x equals zero, you 00:25:25.540 --> 00:25:28.040 align:middle line:84% can't estimate the switching regressions model. 00:25:28.040 --> 00:25:34.820 align:middle line:84% But if you're willing to go epsilon normal zero sigma 00:25:34.820 --> 00:25:35.930 align:middle line:90% squared. 00:25:35.930 --> 00:25:37.880 align:middle line:84% If you're willing to add that to your model 00:25:37.880 --> 00:25:41.240 align:middle line:84% as an identifying assumption, if you're willing to add this, 00:25:41.240 --> 00:25:43.790 align:middle line:84% then yeah, the simple intuition that I can estimate this one 00:25:43.790 --> 00:25:47.630 align:middle line:84% by maximum likelihood and just saying what combination 00:25:47.630 --> 00:25:50.330 align:middle line:90% of two lines and sigma. 00:25:50.330 --> 00:25:55.900 align:middle line:84% So what beta, beta prime, sigma, and lambda best fit the data, 00:25:55.900 --> 00:25:58.250 align:middle line:90% I can do that. 00:25:58.250 --> 00:26:01.170 align:middle line:84% So you need functional forms to estimate switching regressions, 00:26:01.170 --> 00:26:03.890 align:middle line:84% but if you're willing to impose functional forms, 00:26:03.890 --> 00:26:07.205 align:middle line:84% then you can estimate this and you can recover the-- 00:26:07.205 --> 00:26:11.170 align:middle line:90% 00:26:11.170 --> 00:26:18.330 align:middle line:84% you don't ex-post know which equation generated every data 00:26:18.330 --> 00:26:18.830 align:middle line:90% point. 00:26:18.830 --> 00:26:21.850 align:middle line:84% So there's a data point right here. 00:26:21.850 --> 00:26:23.620 align:middle line:84% You don't know for sure whether it 00:26:23.620 --> 00:26:25.510 align:middle line:84% was this line with an error of zero 00:26:25.510 --> 00:26:28.670 align:middle line:84% or this line with a very huge positive error, 00:26:28.670 --> 00:26:31.000 align:middle line:84% but you do get a Bayesian posterior 00:26:31.000 --> 00:26:34.750 align:middle line:84% that, with high probability, it was this line that generated it. 00:26:34.750 --> 00:26:39.190 align:middle line:84% So what you get is you get this model, and then ex-post, 00:26:39.190 --> 00:26:43.420 align:middle line:84% you get a probabilistic Bayesian posterior on which state 00:26:43.420 --> 00:26:45.460 align:middle line:90% generated every data point. 00:26:45.460 --> 00:26:46.090 align:middle line:90% OK? 00:26:46.090 --> 00:26:50.020 align:middle line:84% Anyway, so what Porter does is just basically 00:26:50.020 --> 00:26:53.680 align:middle line:84% do the simultaneous equation version of a switching 00:26:53.680 --> 00:26:54.980 align:middle line:90% regressions model. 00:26:54.980 --> 00:26:56.620 align:middle line:84% And he says that we have these two 00:26:56.620 --> 00:27:00.700 align:middle line:90% models that generated the data. 00:27:00.700 --> 00:27:02.920 align:middle line:84% Every period is either generated by this 00:27:02.920 --> 00:27:08.590 align:middle line:84% with I equals 0 or this with I equals 1, but I is unobserved. 00:27:08.590 --> 00:27:13.150 align:middle line:84% And all I know is that there's probability-- 00:27:13.150 --> 00:27:21.110 align:middle line:84% there's some probability lambda that I equals 1 00:27:21.110 --> 00:27:22.740 align:middle line:90% or that It equals 1. 00:27:22.740 --> 00:27:24.860 align:middle line:84% But I'm going to treat It as an unobservable 00:27:24.860 --> 00:27:29.030 align:middle line:84% and fit the data by a combination of two 00:27:29.030 --> 00:27:32.630 align:middle line:84% simultaneous equation systems, one simultaneous equation system 00:27:32.630 --> 00:27:34.730 align:middle line:84% with I equals zero, one simultaneous equation 00:27:34.730 --> 00:27:36.240 align:middle line:90% system with I equals one. 00:27:36.240 --> 00:27:38.090 align:middle line:90% And in every period explain-- 00:27:38.090 --> 00:27:42.170 align:middle line:84% again, it's every period you get a ptqt pair, 00:27:42.170 --> 00:27:45.230 align:middle line:84% you just ask which model best explains ptqt, 00:27:45.230 --> 00:27:47.760 align:middle line:84% and is it the model I equals zero or I equals one, 00:27:47.760 --> 00:27:50.400 align:middle line:84% but I'm going to do this by maximum likelihood. 00:27:50.400 --> 00:27:53.380 align:middle line:90% 00:27:53.380 --> 00:27:55.780 align:middle line:84% I think that's, in some sense, the preferred estimation 00:27:55.780 --> 00:27:59.080 align:middle line:84% in his model is to say that I don't want to trust the Railway 00:27:59.080 --> 00:28:01.120 align:middle line:84% Review data and I'm going to do this 00:28:01.120 --> 00:28:04.095 align:middle line:84% with the unobserved It's and say what fits the data better. 00:28:04.095 --> 00:28:08.660 align:middle line:90% 00:28:08.660 --> 00:28:14.870 align:middle line:84% I think this idea of explicit unobservables 00:28:14.870 --> 00:28:16.820 align:middle line:84% and putting the unobservables in your model 00:28:16.820 --> 00:28:20.930 align:middle line:84% and making some assumption about the random process they follow, 00:28:20.930 --> 00:28:21.650 align:middle line:90% huge. 00:28:21.650 --> 00:28:23.120 align:middle line:84% This has also been something that's 00:28:23.120 --> 00:28:25.910 align:middle line:84% done all over the place in IO these days 00:28:25.910 --> 00:28:28.460 align:middle line:84% of taking an unobservable, putting it in there, 00:28:28.460 --> 00:28:30.710 align:middle line:84% modeling the unobservable, and then seeing what you 00:28:30.710 --> 00:28:35.130 align:middle line:90% can do with that unobservable. 00:28:35.130 --> 00:28:35.670 align:middle line:90% OK. 00:28:35.670 --> 00:28:37.445 align:middle line:90% What do his estimates look like? 00:28:37.445 --> 00:28:40.430 align:middle line:90% 00:28:40.430 --> 00:28:44.810 align:middle line:84% The first thing is that prices are 00:28:44.810 --> 00:28:47.690 align:middle line:84% dramatically higher in the cooperative phase 00:28:47.690 --> 00:28:49.500 align:middle line:90% than the non-cooperative phase. 00:28:49.500 --> 00:28:57.170 align:middle line:84% If you use the PO dummy, prices are e to the 0.38 times as high 00:28:57.170 --> 00:29:00.320 align:middle line:84% when you are in the cooperative phase. 00:29:00.320 --> 00:29:05.810 align:middle line:84% When you use the PN dummy, prices are e to the 0.545 times 00:29:05.810 --> 00:29:07.220 align:middle line:90% as high. 00:29:07.220 --> 00:29:10.200 align:middle line:84% So what is e to the 0.545, like 80% or something? 00:29:10.200 --> 00:29:12.200 align:middle line:90% So prices are much, much higher. 00:29:12.200 --> 00:29:14.090 align:middle line:90% Here's a simple table. 00:29:14.090 --> 00:29:16.760 align:middle line:84% Again, we don't know PN, but we've got 00:29:16.760 --> 00:29:19.640 align:middle line:90% PN star, which is what's our-- 00:29:19.640 --> 00:29:21.860 align:middle line:84% classify them ex-post based on which 00:29:21.860 --> 00:29:23.750 align:middle line:84% do I think it is after seeing the price. 00:29:23.750 --> 00:29:27.800 align:middle line:84% And it's like the average price is 17in in the cooperative 00:29:27.800 --> 00:29:31.090 align:middle line:84% phase, 28 in the non-cooperative phase. 00:29:31.090 --> 00:29:31.590 align:middle line:90% Sorry. 00:29:31.590 --> 00:29:34.580 align:middle line:84% 28 in the cooperative, 16 in the price war. 00:29:34.580 --> 00:29:37.470 align:middle line:84% So that's roughly the magnitude of how much collusion 00:29:37.470 --> 00:29:38.323 align:middle line:90% is affecting prices. 00:29:38.323 --> 00:29:41.370 align:middle line:90% 00:29:41.370 --> 00:29:43.890 align:middle line:84% The model with the switching regressions, 00:29:43.890 --> 00:29:47.460 align:middle line:84% assuming that the PO is misreported, 00:29:47.460 --> 00:29:50.140 align:middle line:84% fits dramatically better than the one using PO. 00:29:50.140 --> 00:29:53.610 align:middle line:84% The r squared goes from 0.32 to 0.86. 00:29:53.610 --> 00:29:56.940 align:middle line:84% So it looks like the model with the unobserved cooperative phase 00:29:56.940 --> 00:30:00.240 align:middle line:90% is really important. 00:30:00.240 --> 00:30:02.940 align:middle line:84% The model tells you what percentage of weeks 00:30:02.940 --> 00:30:07.140 align:middle line:84% are in the price war phase, and it says that about 28% 00:30:07.140 --> 00:30:11.410 align:middle line:84% of the weeks are in the price war phase in the data. 00:30:11.410 --> 00:30:13.720 align:middle line:84% And if you just look at the raw data, 00:30:13.720 --> 00:30:17.120 align:middle line:84% I guess the beginning of the first price war, 00:30:17.120 --> 00:30:21.970 align:middle line:84% you see prices going on this at $0.40, 36, 35, 00:30:21.970 --> 00:30:24.550 align:middle line:84% $0.30 per hundred pounds, 35, 35. 00:30:24.550 --> 00:30:26.740 align:middle line:90% And then it drops to 16. 00:30:26.740 --> 00:30:29.950 align:middle line:84% Again, this is not some cost shock 00:30:29.950 --> 00:30:32.350 align:middle line:84% that you suddenly became able to transport grain 00:30:32.350 --> 00:30:34.930 align:middle line:84% from Chicago to East Coast at half the price as 00:30:34.930 --> 00:30:36.290 align:middle line:90% of one week ago. 00:30:36.290 --> 00:30:38.710 align:middle line:84% It does look like just dramatic evidence 00:30:38.710 --> 00:30:40.080 align:middle line:90% that this is a price war. 00:30:40.080 --> 00:30:42.890 align:middle line:84% They price really low for a long time. 00:30:42.890 --> 00:30:45.190 align:middle line:84% We do think they would be losing money getting grain 00:30:45.190 --> 00:30:47.200 align:middle line:84% to the US, the East Coast at $0.12, 00:30:47.200 --> 00:30:48.966 align:middle line:84% and then they jump back up afterwards. 00:30:48.966 --> 00:30:55.940 align:middle line:90% 00:30:55.940 --> 00:30:58.170 align:middle line:84% I didn't put that other key test in. 00:30:58.170 --> 00:31:00.800 align:middle line:84% Obviously, the key test he asks of the Green-Porter model 00:31:00.800 --> 00:31:03.770 align:middle line:84% is, does a model with two supply curves 00:31:03.770 --> 00:31:06.360 align:middle line:84% fit better than a model with one supply curve? 00:31:06.360 --> 00:31:11.160 align:middle line:84% The static Nash story here would be that we have this model 00:31:11.160 --> 00:31:15.480 align:middle line:84% and there is no I. The I is just always-- 00:31:15.480 --> 00:31:18.310 align:middle line:84% they're just always behaving the same in every period. 00:31:18.310 --> 00:31:21.390 align:middle line:84% And so if you test these two different behaviors versus one 00:31:21.390 --> 00:31:23.610 align:middle line:84% behavior model, I guess you can think 00:31:23.610 --> 00:31:32.060 align:middle line:84% of that as like testing lambda is not in either zero or one. 00:31:32.060 --> 00:31:33.890 align:middle line:84% Lambda equals zero or lambda equals 00:31:33.890 --> 00:31:37.380 align:middle line:84% one is the model with one regime instead of two. 00:31:37.380 --> 00:31:40.100 align:middle line:84% And can you tell whether lambda is significantly 00:31:40.100 --> 00:31:41.850 align:middle line:90% different from zero or one? 00:31:41.850 --> 00:31:45.150 align:middle line:90% 00:31:45.150 --> 00:31:48.180 align:middle line:84% I don't have an estimate of lambda in here, do I? 00:31:48.180 --> 00:31:52.350 align:middle line:84% OK, anyway-- He provides tests in the paper and saying it's 00:31:52.350 --> 00:31:55.260 align:middle line:84% tremendously significant that two supply curves fit 00:31:55.260 --> 00:31:57.420 align:middle line:90% better than one supply curve. 00:31:57.420 --> 00:31:59.370 align:middle line:90% You know, it's not surprising. 00:31:59.370 --> 00:32:02.010 align:middle line:84% It's things like this are really hard to fit with one supply 00:32:02.010 --> 00:32:05.040 align:middle line:84% curve, whereas with two supply curves, 00:32:05.040 --> 00:32:07.850 align:middle line:90% you fit that much better. 00:32:07.850 --> 00:32:08.350 align:middle line:90% OK. 00:32:08.350 --> 00:32:12.020 align:middle line:84% So next I get to go to one of my own papers. 00:32:12.020 --> 00:32:16.650 align:middle line:84% Still including this, just sentimental reasons. 00:32:16.650 --> 00:32:19.650 align:middle line:84% So this was actually when I was a graduate student here. 00:32:19.650 --> 00:32:21.190 align:middle line:90% There was no second year paper. 00:32:21.190 --> 00:32:26.160 align:middle line:84% So 14.192 was not invented until the year 2019 or whatever. 00:32:26.160 --> 00:32:29.760 align:middle line:84% What did exist before then was there was both a history paper 00:32:29.760 --> 00:32:31.830 align:middle line:90% and an econometrics paper. 00:32:31.830 --> 00:32:35.810 align:middle line:84% And one of the great bonuses was for-- 00:32:35.810 --> 00:32:38.280 align:middle line:84% what everyone tried to think of as getting 00:32:38.280 --> 00:32:40.697 align:middle line:84% through the program with the least amount of work possible 00:32:40.697 --> 00:32:43.200 align:middle line:84% is could I write a history paper that's also my econometrics 00:32:43.200 --> 00:32:43.720 align:middle line:90% paper? 00:32:43.720 --> 00:32:45.275 align:middle line:90% Because you then got to-- 00:32:45.275 --> 00:32:46.650 align:middle line:84% there was no rule against handing 00:32:46.650 --> 00:32:47.940 align:middle line:84% in the same paper for your history paper 00:32:47.940 --> 00:32:49.480 align:middle line:90% and your econometrics paper. 00:32:49.480 --> 00:32:53.550 align:middle line:84% So as a student in this program, I-- 00:32:53.550 --> 00:32:55.810 align:middle line:84% oh, this railroad cartel data is from 1880. 00:32:55.810 --> 00:32:57.700 align:middle line:84% That qualifies as making it a history paper. 00:32:57.700 --> 00:32:59.670 align:middle line:84% So anyway, so this was my history paper 00:32:59.670 --> 00:33:01.620 align:middle line:90% and my econometrics paper. 00:33:01.620 --> 00:33:03.900 align:middle line:90% But anyway, and I guess I also-- 00:33:03.900 --> 00:33:06.330 align:middle line:84% just as combinations of you don't 00:33:06.330 --> 00:33:09.400 align:middle line:84% have to be all that creative and original to write papers. 00:33:09.400 --> 00:33:12.570 align:middle line:84% And a lot of the time it works to just make small improvements 00:33:12.570 --> 00:33:15.982 align:middle line:84% on the existing literature and just-- 00:33:15.982 --> 00:33:18.440 align:middle line:84% you start with something that's in the existing literature. 00:33:18.440 --> 00:33:19.580 align:middle line:84% You make a small improvement of it. 00:33:19.580 --> 00:33:21.195 align:middle line:90% You know you're at the frontier. 00:33:21.195 --> 00:33:22.570 align:middle line:84% Whereas if you start from scratch 00:33:22.570 --> 00:33:23.740 align:middle line:84% and do something really creative, 00:33:23.740 --> 00:33:25.532 align:middle line:84% you can end up with a worse version of what 00:33:25.532 --> 00:33:26.720 align:middle line:90% someone else did before. 00:33:26.720 --> 00:33:29.530 align:middle line:84% And so anyway, so this was my paper looking 00:33:29.530 --> 00:33:33.730 align:middle line:84% at-- thinking about what more could I say about Green-Porter. 00:33:33.730 --> 00:33:39.130 align:middle line:84% And the motivating idea here was that Porter 00:33:39.130 --> 00:33:40.690 align:middle line:84% shows that there are two regimes, 00:33:40.690 --> 00:33:44.330 align:middle line:84% and two regimes explain the data better than one. 00:33:44.330 --> 00:33:46.990 align:middle line:84% But I think the content of the Green-Porter theory 00:33:46.990 --> 00:33:50.990 align:middle line:84% is not just that, but price wars occur 00:33:50.990 --> 00:33:53.690 align:middle line:84% when there are suspicious demand realizations and demand 00:33:53.690 --> 00:33:56.300 align:middle line:84% realizations that look like someone cheated on the cartel 00:33:56.300 --> 00:34:00.295 align:middle line:84% agreement, and then when there is a demand realization that 00:34:00.295 --> 00:34:02.420 align:middle line:84% looks like someone cheated on the cartel agreement, 00:34:02.420 --> 00:34:05.062 align:middle line:84% then you should be more likely to trigger a price war. 00:34:05.062 --> 00:34:06.770 align:middle line:84% And what we think these price wars should 00:34:06.770 --> 00:34:09.540 align:middle line:84% look like is you're pricing high, you go down, 00:34:09.540 --> 00:34:11.540 align:middle line:84% you have a price war for some number of periods, 00:34:11.540 --> 00:34:14.010 align:middle line:84% and then you go back to pricing high again. 00:34:14.010 --> 00:34:16.010 align:middle line:84% So it's both the price wars are supposed 00:34:16.010 --> 00:34:18.110 align:middle line:84% to be continuous periods and they're 00:34:18.110 --> 00:34:21.020 align:middle line:84% supposed to come after some demand pattern that 00:34:21.020 --> 00:34:22.770 align:middle line:84% looks like someone might have cheated. 00:34:22.770 --> 00:34:25.139 align:middle line:84% And so that's what I'm trying to test here. 00:34:25.139 --> 00:34:31.070 align:middle line:84% So basic exercise is just re-estimate Porter's model. 00:34:31.070 --> 00:34:36.350 align:middle line:84% I do make a few changes which reflect this being the 1990s 00:34:36.350 --> 00:34:40.100 align:middle line:84% instead of the 1980s and computers were better. 00:34:40.100 --> 00:34:45.409 align:middle line:84% One is paper allows for serial correlation in demand. 00:34:45.409 --> 00:34:48.260 align:middle line:84% The demand in Porter is tremendously serially correlated 00:34:48.260 --> 00:34:53.340 align:middle line:84% if you look at it, so he didn't allow serial correlation. 00:34:53.340 --> 00:34:56.340 align:middle line:84% Supply is just like in Porter, with there 00:34:56.340 --> 00:34:59.090 align:middle line:90% being an unobservable It. 00:34:59.090 --> 00:35:02.210 align:middle line:84% But then the main economic idea is 00:35:02.210 --> 00:35:05.270 align:middle line:84% to let the regime transitions follow a first order Markov 00:35:05.270 --> 00:35:09.830 align:middle line:84% process, where the probability of cooperation at t plus 1, 00:35:09.830 --> 00:35:11.780 align:middle line:84% conditional on whether you are cooperating 00:35:11.780 --> 00:35:15.200 align:middle line:84% at t and some other observable variables wt, 00:35:15.200 --> 00:35:20.520 align:middle line:84% is going to be e to the gamma wt over 1 plus e to the gamma wt. 00:35:20.520 --> 00:35:23.970 align:middle line:84% So these variables wt are going to affect 00:35:23.970 --> 00:35:26.010 align:middle line:84% the probability of the Markov transitions 00:35:26.010 --> 00:35:29.280 align:middle line:84% from cooperating to price war or from price war 00:35:29.280 --> 00:35:30.456 align:middle line:90% back to cooperation. 00:35:30.456 --> 00:35:33.480 align:middle line:90% 00:35:33.480 --> 00:35:36.190 align:middle line:84% And what are the variables I put in here? 00:35:36.190 --> 00:35:38.370 align:middle line:84% So obviously first version to estimate 00:35:38.370 --> 00:35:40.090 align:middle line:90% is just let w be constant. 00:35:40.090 --> 00:35:44.040 align:middle line:84% This is Porter's model where the regimes are i.i.d. 00:35:44.040 --> 00:35:48.600 align:middle line:84% If there is nothing in w other than a constant, 00:35:48.600 --> 00:35:51.980 align:middle line:90% this is Porter's model. 00:35:51.980 --> 00:35:54.780 align:middle line:90% If you put It in here-- 00:35:54.780 --> 00:35:58.820 align:middle line:84% so you get the cooperation indicator affect-- 00:35:58.820 --> 00:36:01.370 align:middle line:84% you let It affect the probability that It plus 1 00:36:01.370 --> 00:36:02.340 align:middle line:90% is one. 00:36:02.340 --> 00:36:05.600 align:middle line:84% Then you have first order Markov structure on price wars, 00:36:05.600 --> 00:36:09.945 align:middle line:84% and you can see whether price wars are continuous. 00:36:09.945 --> 00:36:12.320 align:middle line:84% Because if price wars are continuous periods, what you'll 00:36:12.320 --> 00:36:14.420 align:middle line:84% see is if you're cooperating, there's 00:36:14.420 --> 00:36:17.132 align:middle line:84% a 90% chance you're going to cooperate. 00:36:17.132 --> 00:36:18.590 align:middle line:84% Once you're in a price war, there's 00:36:18.590 --> 00:36:21.980 align:middle line:84% a 90% chance you're going to be in a price war, 00:36:21.980 --> 00:36:26.840 align:middle line:84% so the price wars become long, continuous periods. 00:36:26.840 --> 00:36:29.870 align:middle line:84% And then the primary interest in specifications 00:36:29.870 --> 00:36:32.690 align:middle line:84% that also include variables that look like suspicious demand 00:36:32.690 --> 00:36:34.860 align:middle line:90% patterns. 00:36:34.860 --> 00:36:37.120 align:middle line:90% What is it that Fink was doing? 00:36:37.120 --> 00:36:39.040 align:middle line:84% Well, Fink could have done many, many things. 00:36:39.040 --> 00:36:40.498 align:middle line:84% There are lots of things that could 00:36:40.498 --> 00:36:42.210 align:middle line:84% make you think that somebody was cheating 00:36:42.210 --> 00:36:45.480 align:middle line:84% or could make the firms think someone was cheating. 00:36:45.480 --> 00:36:49.080 align:middle line:84% One would be whether any firm had an unusually large market 00:36:49.080 --> 00:36:50.190 align:middle line:90% share. 00:36:50.190 --> 00:36:52.920 align:middle line:84% This isn't in the Green-Porter model, but in the JEC, 00:36:52.920 --> 00:36:56.380 align:middle line:84% there were guys counting cars and counting cars by railroad. 00:36:56.380 --> 00:36:58.800 align:middle line:84% And so if you observe the Pennsylvania Railroad has 00:36:58.800 --> 00:37:01.500 align:middle line:84% 50% more demand than everybody else this week, 00:37:01.500 --> 00:37:04.085 align:middle line:84% it could be that all the grain shippers were actually 00:37:04.085 --> 00:37:05.460 align:middle line:84% trying to get it to Philadelphia, 00:37:05.460 --> 00:37:08.900 align:middle line:84% but it could be the Pennsylvania railroad was cheating. 00:37:08.900 --> 00:37:11.590 align:middle line:84% So if one firm had an unusually large market share, 00:37:11.590 --> 00:37:14.920 align:middle line:84% that might make you trigger a price war. 00:37:14.920 --> 00:37:17.920 align:middle line:84% Another thing would be if any firm had an unusually small 00:37:17.920 --> 00:37:19.602 align:middle line:90% market share. 00:37:19.602 --> 00:37:21.310 align:middle line:84% Firms could worry that the other two guys 00:37:21.310 --> 00:37:22.970 align:middle line:90% are both cheating against me. 00:37:22.970 --> 00:37:27.790 align:middle line:84% And so if I see that instead of being 30/30/30 or 30/30/40 00:37:27.790 --> 00:37:33.817 align:middle line:84% or whatever, it's 40/50/10, the firm that's getting the 10 could 00:37:33.817 --> 00:37:36.150 align:middle line:84% worry that the other two are both cheating on the cartel 00:37:36.150 --> 00:37:36.855 align:middle line:90% agreement. 00:37:36.855 --> 00:37:38.730 align:middle line:84% And so you might want to have a system where, 00:37:38.730 --> 00:37:41.250 align:middle line:84% if anybody has an unusually small market share, 00:37:41.250 --> 00:37:43.762 align:middle line:90% then you start a price war. 00:37:43.762 --> 00:37:46.220 align:middle line:84% Another thing you could also do is just if aggregate demand 00:37:46.220 --> 00:37:48.140 align:middle line:90% is unusually high. 00:37:48.140 --> 00:37:51.270 align:middle line:84% This would be more like the Green-Porter textbook model, 00:37:51.270 --> 00:37:55.070 align:middle line:84% but if suddenly aggregate demand is really high, 00:37:55.070 --> 00:37:57.398 align:middle line:84% then you think that someone out there must be cheating, 00:37:57.398 --> 00:37:59.940 align:middle line:84% therefore we're going to be more likely to start a price war. 00:37:59.940 --> 00:38:03.200 align:middle line:84% And obviously you could do many combinations of these. 00:38:03.200 --> 00:38:07.520 align:middle line:84% What I do in the paper is just put in several of them 00:38:07.520 --> 00:38:10.520 align:middle line:90% and see what's there. 00:38:10.520 --> 00:38:14.790 align:middle line:84% First is the standard model without the w's. 00:38:14.790 --> 00:38:18.440 align:middle line:84% So standard model without the w's, there's 00:38:18.440 --> 00:38:20.036 align:middle line:90% strong serial correlation. 00:38:20.036 --> 00:38:22.560 align:middle line:90% 00:38:22.560 --> 00:38:29.290 align:middle line:84% The effect of collusion, or the cooperative phase, on pricing 00:38:29.290 --> 00:38:31.210 align:middle line:84% is even larger than Porter by a little bit. 00:38:31.210 --> 00:38:34.540 align:middle line:90% It's 0.6 instead of 0.5. 00:38:34.540 --> 00:38:36.310 align:middle line:84% But what does change when you account 00:38:36.310 --> 00:38:40.300 align:middle line:84% for the serial correlation is the demand elasticity, 00:38:40.300 --> 00:38:43.550 align:middle line:84% which Porter had at 0.8, and it was a puzzle. 00:38:43.550 --> 00:38:46.270 align:middle line:84% Why would the cartel be pricing where the demand elasticity 00:38:46.270 --> 00:38:47.090 align:middle line:90% was 0.8? 00:38:47.090 --> 00:38:49.690 align:middle line:84% Don't we know that you normally price in the elastic part 00:38:49.690 --> 00:38:50.890 align:middle line:90% of the demand curve? 00:38:50.890 --> 00:38:52.870 align:middle line:84% Once you account for the serial correlation, 00:38:52.870 --> 00:38:56.770 align:middle line:84% that mystery goes away and I now say that the demand elasticity 00:38:56.770 --> 00:39:00.220 align:middle line:90% was minus 1.8, not minus 0.8. 00:39:00.220 --> 00:39:06.510 align:middle line:84% And so if you think about what the implied theta by the data 00:39:06.510 --> 00:39:10.470 align:middle line:84% is, the implied theta from how they're pricing-- 00:39:10.470 --> 00:39:14.070 align:middle line:84% again, you have price minus cost over price 00:39:14.070 --> 00:39:17.830 align:middle line:84% equals theta over the elasticity. 00:39:17.830 --> 00:39:25.240 align:middle line:84% So now you plug in 1.8 here instead of 0.8 00:39:25.240 --> 00:39:28.480 align:middle line:84% and you say, what's the theta that's causing this? 00:39:28.480 --> 00:39:31.330 align:middle line:84% Now I'm getting that the theta is fairly-- 00:39:31.330 --> 00:39:33.730 align:middle line:84% they're using a theta that's pretty close to 1. 00:39:33.730 --> 00:39:36.110 align:middle line:84% It looks like the theta is like 0.85. 00:39:36.110 --> 00:39:42.880 align:middle line:84% So again, you're not-- if you assume that theta c equals zero 00:39:42.880 --> 00:39:44.090 align:middle line:90% would imply that theta-- 00:39:44.090 --> 00:39:44.590 align:middle line:90% sorry. 00:39:44.590 --> 00:39:49.720 align:middle line:84% Theta w equals zero would imply that theta c is about 0.85. 00:39:49.720 --> 00:39:52.690 align:middle line:84% So with this new demand elasticity, 00:39:52.690 --> 00:39:55.750 align:middle line:84% it looks like it's not just a Green-Porter cartel having 00:39:55.750 --> 00:39:56.350 align:middle line:90% two regimes. 00:39:56.350 --> 00:39:59.110 align:middle line:84% It's two regimes where one is close to cost 00:39:59.110 --> 00:40:01.490 align:middle line:84% and the other is very close to the monopoly price. 00:40:01.490 --> 00:40:03.100 align:middle line:84% So it looks like they are aggressively 00:40:03.100 --> 00:40:07.856 align:middle line:84% trying to get nearly the full monopoly profits. 00:40:07.856 --> 00:40:13.580 align:middle line:84% And if you look at this, the probability of a price war-- 00:40:13.580 --> 00:40:16.550 align:middle line:84% the probability of cooperating when you're cooperating today 00:40:16.550 --> 00:40:22.020 align:middle line:84% is e to the 3.6 over 1 plus e to the 3.6. 00:40:22.020 --> 00:40:24.020 align:middle line:84% So if you're cooperating today, the probability 00:40:24.020 --> 00:40:26.340 align:middle line:84% of cooperating tomorrow is very high. 00:40:26.340 --> 00:40:28.760 align:middle line:84% If you're in a price war today, the probability 00:40:28.760 --> 00:40:34.520 align:middle line:84% of cooperating tomorrow is e to the minus 2.6 over 1 plus 00:40:34.520 --> 00:40:36.147 align:middle line:90% e to the minus 2.6. 00:40:36.147 --> 00:40:37.730 align:middle line:84% So that says if you're in a price war, 00:40:37.730 --> 00:40:39.397 align:middle line:84% there's very low probability that you're 00:40:39.397 --> 00:40:41.300 align:middle line:84% going to go back to cooperating next period. 00:40:41.300 --> 00:40:43.220 align:middle line:84% So clearly that these numbers are different 00:40:43.220 --> 00:40:47.930 align:middle line:84% mean that these price wars are continuous periods in the data. 00:40:47.930 --> 00:40:50.690 align:middle line:84% Also it makes you reclassify some of the periods. 00:40:50.690 --> 00:40:54.260 align:middle line:84% In Porter's data, you can be like, OK, price war, price war, 00:40:54.260 --> 00:40:56.210 align:middle line:84% price war, price war, cooperation, price war, 00:40:56.210 --> 00:40:57.680 align:middle line:90% price war, price war. 00:40:57.680 --> 00:41:00.368 align:middle line:84% And then with putting in this variable that 00:41:00.368 --> 00:41:02.660 align:middle line:84% reclassifies them and says, no, that's a price war too. 00:41:02.660 --> 00:41:05.580 align:middle line:84% It's just a price war that has a big epsilon attached to it. 00:41:05.580 --> 00:41:10.100 align:middle line:90% 00:41:10.100 --> 00:41:12.060 align:middle line:90% OK. 00:41:12.060 --> 00:41:12.560 align:middle line:90% OK. 00:41:12.560 --> 00:41:14.018 align:middle line:84% And then what about these variables 00:41:14.018 --> 00:41:16.460 align:middle line:90% for things causing price wars? 00:41:16.460 --> 00:41:19.680 align:middle line:84% Some of them are kind of significant 00:41:19.680 --> 00:41:24.380 align:middle line:84% if you do at 5% level significant one-sided tail. 00:41:24.380 --> 00:41:29.580 align:middle line:84% So some firm having an unusually large market share, 00:41:29.580 --> 00:41:34.310 align:middle line:84% it seems like that makes price wars more likely to start. 00:41:34.310 --> 00:41:40.790 align:middle line:84% Also, VT, the aggregate demand residual being high, 00:41:40.790 --> 00:41:44.010 align:middle line:84% also makes it look like price wars are going to start. 00:41:44.010 --> 00:41:46.680 align:middle line:84% So if think about this is a model of Albert Fink, 00:41:46.680 --> 00:41:51.920 align:middle line:84% it does seem like if aggregate demand is unusually high 00:41:51.920 --> 00:41:56.000 align:middle line:84% or one firm has an unusually high market share, 00:41:56.000 --> 00:41:58.340 align:middle line:84% that makes it more likely that he's declaring price war 00:41:58.340 --> 00:41:59.155 align:middle line:90% starts tomorrow. 00:41:59.155 --> 00:42:03.430 align:middle line:90% 00:42:03.430 --> 00:42:05.800 align:middle line:84% Significance is just marginal, though, 00:42:05.800 --> 00:42:08.320 align:middle line:84% and the triggers are not really powerful enough 00:42:08.320 --> 00:42:09.560 align:middle line:90% to deter deviation. 00:42:09.560 --> 00:42:11.530 align:middle line:84% So if these were the magnitudes of the triggers 00:42:11.530 --> 00:42:14.212 align:middle line:84% that Fink was using and you were advising the New York Central 00:42:14.212 --> 00:42:16.670 align:middle line:84% Railroad, you'd be like, yeah, you guys should be cheating. 00:42:16.670 --> 00:42:19.377 align:middle line:84% Because if you cheat, you gain 10% market share, 00:42:19.377 --> 00:42:20.960 align:middle line:84% you're going to earn a lot more money, 00:42:20.960 --> 00:42:24.410 align:middle line:84% and the probability of price war only goes up by a few percent 00:42:24.410 --> 00:42:27.350 align:middle line:84% so you'd be better off cheating on this agreement. 00:42:27.350 --> 00:42:30.550 align:middle line:84% So obviously, what's going on is-- 00:42:30.550 --> 00:42:31.970 align:middle line:90% this leaves multiple hypotheses. 00:42:31.970 --> 00:42:35.680 align:middle line:84% So one is that Fink got it wrong and his triggers weren't 00:42:35.680 --> 00:42:36.590 align:middle line:90% sensitive enough. 00:42:36.590 --> 00:42:38.590 align:middle line:84% He was too hesitant to trigger price wars, 00:42:38.590 --> 00:42:41.440 align:middle line:84% and therefore, firms did have an incentive to cheat 00:42:41.440 --> 00:42:43.270 align:middle line:90% and firms did cheat. 00:42:43.270 --> 00:42:46.933 align:middle line:84% The other hypothesis is that he had a trigger that's 00:42:46.933 --> 00:42:49.600 align:middle line:84% imperfectly correlated with all the variables I've put in there, 00:42:49.600 --> 00:42:51.267 align:middle line:84% and because it's imperfectly correlated, 00:42:51.267 --> 00:42:53.290 align:middle line:84% there's attenuation bias, there's 00:42:53.290 --> 00:42:57.168 align:middle line:84% a stronger trigger out there, I just didn't find it. 00:42:57.168 --> 00:42:59.710 align:middle line:84% Unfortunately, this is the kind of thing where you might say, 00:42:59.710 --> 00:43:01.918 align:middle line:84% could I do some machine learning thing and figure out 00:43:01.918 --> 00:43:04.070 align:middle line:84% exactly what Albert Fink was doing? 00:43:04.070 --> 00:43:07.110 align:middle line:84% There are between 8 and 11 price wars. 00:43:07.110 --> 00:43:10.010 align:middle line:84% So in some sense, you only have 8 or 11 data points. 00:43:10.010 --> 00:43:13.340 align:middle line:84% There's only so much you can do with sophisticated statistics 00:43:13.340 --> 00:43:15.880 align:middle line:84% when you're trying to figure out when a price war started. 00:43:15.880 --> 00:43:20.200 align:middle line:90% 00:43:20.200 --> 00:43:20.740 align:middle line:90% OK. 00:43:20.740 --> 00:43:21.825 align:middle line:90% Any questions on that? 00:43:21.825 --> 00:43:26.580 align:middle line:90% 00:43:26.580 --> 00:43:28.540 align:middle line:84% Anyway, so other things in the paper. 00:43:28.540 --> 00:43:31.090 align:middle line:84% I do have something looking at Rotemberg-Saloner effects. 00:43:31.090 --> 00:43:34.030 align:middle line:90% I won't talk about that. 00:43:34.030 --> 00:43:38.470 align:middle line:84% The paper asks the question in the back part, 00:43:38.470 --> 00:43:40.670 align:middle line:84% are there secret price cuts, and does it 00:43:40.670 --> 00:43:44.270 align:middle line:84% look like secret price cuts were being given? 00:43:44.270 --> 00:43:49.730 align:middle line:84% And the way it does this is to add extra unobserved demand 00:43:49.730 --> 00:43:53.150 align:middle line:84% regimes, saying that if firms are giving secret price cuts, 00:43:53.150 --> 00:43:55.850 align:middle line:84% demand is going to look unusually high relative to what 00:43:55.850 --> 00:43:57.540 align:middle line:90% demand normally looks like. 00:43:57.540 --> 00:43:59.090 align:middle line:84% Can we estimate a hidden demand-- 00:43:59.090 --> 00:44:01.360 align:middle line:84% a hidden regime model to say, does it 00:44:01.360 --> 00:44:03.110 align:middle line:84% look like there's some fraction of periods 00:44:03.110 --> 00:44:06.033 align:middle line:84% in which people are cheating and demand is unusually high? 00:44:06.033 --> 00:44:07.700 align:middle line:84% And it provides some suggestive evidence 00:44:07.700 --> 00:44:10.242 align:middle line:84% that it does look like that's probably what was going on here 00:44:10.242 --> 00:44:13.580 align:middle line:84% historically is that there periods where 00:44:13.580 --> 00:44:15.620 align:middle line:84% demand is suspiciously high, making 00:44:15.620 --> 00:44:18.380 align:middle line:90% you think someone is cheating. 00:44:18.380 --> 00:44:19.670 align:middle line:90% And so we think the-- 00:44:19.670 --> 00:44:22.657 align:middle line:84% it aims toward that Fink didn't have it quite right 00:44:22.657 --> 00:44:24.865 align:middle line:84% and the firms were doing some cheating on each other. 00:44:24.865 --> 00:44:27.640 align:middle line:90% 00:44:27.640 --> 00:44:29.590 align:middle line:84% But obviously, if you think about where did I 00:44:29.590 --> 00:44:32.080 align:middle line:84% miss an enormous opportunity, there 00:44:32.080 --> 00:44:34.060 align:middle line:84% are papers written subsequent to mine 00:44:34.060 --> 00:44:36.460 align:middle line:84% that now have thousands and thousands of citations that 00:44:36.460 --> 00:44:38.950 align:middle line:84% say, well, what you could do in a situation 00:44:38.950 --> 00:44:44.463 align:middle line:84% like this is instead of estimating the trigger 00:44:44.463 --> 00:44:46.130 align:middle line:84% and saying, is the trigger strong enough 00:44:46.130 --> 00:44:49.290 align:middle line:90% to cause collusion? 00:44:49.290 --> 00:44:51.350 align:middle line:84% Maybe what you could do is impose 00:44:51.350 --> 00:44:52.850 align:middle line:84% that the trigger is strong enough 00:44:52.850 --> 00:44:56.420 align:middle line:84% to stop firms from colluding, and then estimate the model 00:44:56.420 --> 00:44:59.317 align:middle line:84% and see if you can find the trigger based on that, 00:44:59.317 --> 00:45:00.900 align:middle line:84% allowing some extra measurement error. 00:45:00.900 --> 00:45:04.520 align:middle line:84% So it could be that the price wars start one week earlier 00:45:04.520 --> 00:45:08.420 align:middle line:84% than I think they start because some firm had an enormous market 00:45:08.420 --> 00:45:10.140 align:middle line:90% share one week earlier. 00:45:10.140 --> 00:45:11.750 align:middle line:84% The reported price is still high, 00:45:11.750 --> 00:45:14.480 align:middle line:84% but maybe the price war actually started one week earlier 00:45:14.480 --> 00:45:17.600 align:middle line:84% with a mismeasured-- there was an enormous cheating 00:45:17.600 --> 00:45:19.230 align:middle line:90% incident for two weeks earlier. 00:45:19.230 --> 00:45:20.790 align:middle line:90% They start the price war. 00:45:20.790 --> 00:45:22.490 align:middle line:84% The price war only shows up in my data, 00:45:22.490 --> 00:45:27.980 align:middle line:84% which uses posted prices one week later, therefore I miss it. 00:45:27.980 --> 00:45:31.190 align:middle line:84% So anyway, I think this is, in some sense, what 00:45:31.190 --> 00:45:34.100 align:middle line:84% the structural literature has done is to say, 00:45:34.100 --> 00:45:36.350 align:middle line:84% if we think that the firms are playing an equilibrium, 00:45:36.350 --> 00:45:39.260 align:middle line:84% we should impose that an equilibrium is being played 00:45:39.260 --> 00:45:41.843 align:middle line:84% and make that an extra constraint on the model 00:45:41.843 --> 00:45:43.676 align:middle line:84% and estimate the model with that constraint. 00:45:43.676 --> 00:45:46.650 align:middle line:90% 00:45:46.650 --> 00:45:49.920 align:middle line:84% I didn't think to do that at the time, but in some sense, 00:45:49.920 --> 00:45:51.720 align:middle line:84% it is a very natural thing to say 00:45:51.720 --> 00:45:53.920 align:middle line:84% is let's impose that constraint, see what happens, 00:45:53.920 --> 00:45:56.128 align:middle line:84% and is the model now able to explain the data better. 00:45:56.128 --> 00:45:59.800 align:middle line:90% 00:45:59.800 --> 00:46:00.310 align:middle line:90% Yeah. 00:46:00.310 --> 00:46:05.560 align:middle line:84% So for instance, in this week, it's very clear 00:46:05.560 --> 00:46:07.810 align:middle line:90% where this price war starts. 00:46:07.810 --> 00:46:10.390 align:middle line:90% The price war starts here. 00:46:10.390 --> 00:46:12.670 align:middle line:84% But it could be that in this week 00:46:12.670 --> 00:46:15.070 align:middle line:84% the firms were already cheating on the agreement 00:46:15.070 --> 00:46:17.770 align:middle line:84% and they were actually pricing here, 00:46:17.770 --> 00:46:20.382 align:middle line:84% and we just have a mismeasured price because the prices 00:46:20.382 --> 00:46:22.090 align:middle line:84% in this data set are the prices that they 00:46:22.090 --> 00:46:23.840 align:middle line:84% were supposed to be charging, not the ones 00:46:23.840 --> 00:46:25.840 align:middle line:90% they actually charged. 00:46:25.840 --> 00:46:30.290 align:middle line:84% Next thing I want to get to is this 00:46:30.290 --> 00:46:32.690 align:middle line:84% is something I had the benefit of seeing with my career. 00:46:32.690 --> 00:46:33.620 align:middle line:90% It's interesting. 00:46:33.620 --> 00:46:35.078 align:middle line:84% I read Mass Control for many years, 00:46:35.078 --> 00:46:37.037 align:middle line:84% and Mass Control would point out that, be like, 00:46:37.037 --> 00:46:38.900 align:middle line:84% these Edgeworth cycles are possible. 00:46:38.900 --> 00:46:40.370 align:middle line:84% It was always thought of, at least 00:46:40.370 --> 00:46:45.830 align:middle line:84% seen as this is an interesting possibility 00:46:45.830 --> 00:46:48.500 align:middle line:84% that, in models with small numbers of firms, 00:46:48.500 --> 00:46:50.210 align:middle line:84% firms can collude and they could collude 00:46:50.210 --> 00:46:55.460 align:middle line:84% not on constant prices like in the standard repeated game 00:46:55.460 --> 00:46:57.980 align:middle line:84% equilibrium, but they might collude-- 00:46:57.980 --> 00:47:00.020 align:middle line:84% collusion might lead to these Edgeworth cycles 00:47:00.020 --> 00:47:01.850 align:middle line:90% where price goes like that. 00:47:01.850 --> 00:47:04.400 align:middle line:84% And I think it was always at least initially thought 00:47:04.400 --> 00:47:08.033 align:middle line:84% of as a intellectual curiosity that collusive models might 00:47:08.033 --> 00:47:09.950 align:middle line:84% actually more naturally lead to cycling prices 00:47:09.950 --> 00:47:11.600 align:middle line:90% rather than high prices. 00:47:11.600 --> 00:47:13.730 align:middle line:84% And then this is a view that the world actually 00:47:13.730 --> 00:47:17.480 align:middle line:84% has changed as high frequency price data came into-- 00:47:17.480 --> 00:47:21.280 align:middle line:84% started getting collected because, again, if I ask you, 00:47:21.280 --> 00:47:23.030 align:middle line:84% like, what are the prices at gas stations, 00:47:23.030 --> 00:47:25.490 align:middle line:84% and what do prices at gas stations do? 00:47:25.490 --> 00:47:27.740 align:middle line:84% You know, you may know that like, yeah, sometimes gas 00:47:27.740 --> 00:47:32.790 align:middle line:84% is 3.69, sometimes it's 3.79, sometimes it's 3.59. 00:47:32.790 --> 00:47:34.470 align:middle line:84% I've never really noticed a pattern 00:47:34.470 --> 00:47:35.700 align:middle line:84% other than knowing that sometimes I 00:47:35.700 --> 00:47:37.283 align:middle line:84% get a good deal on gas and sometimes I 00:47:37.283 --> 00:47:39.450 align:middle line:90% get a bad deal on gas. 00:47:39.450 --> 00:47:43.620 align:middle line:84% And so actually, Mike Noel was a student sitting in this class 00:47:43.620 --> 00:47:46.740 align:middle line:84% where you guys are sitting in the early 2000s, 00:47:46.740 --> 00:47:51.690 align:middle line:84% and he decided he would actually investigate the Edgeworth cycle 00:47:51.690 --> 00:47:52.680 align:middle line:90% hypothesis. 00:47:52.680 --> 00:47:56.730 align:middle line:84% At least had some notion that it seemed like-- 00:47:56.730 --> 00:47:59.130 align:middle line:84% he had the view that in Toronto, where 00:47:59.130 --> 00:48:04.890 align:middle line:84% I grew up and went to college, I think prices did cycle, 00:48:04.890 --> 00:48:07.060 align:middle line:84% or at least I know they changed a lot 00:48:07.060 --> 00:48:09.840 align:middle line:84% and I wonder if they are cycling. 00:48:09.840 --> 00:48:13.350 align:middle line:84% And so what Mike did was he still 00:48:13.350 --> 00:48:15.120 align:middle line:84% had a girlfriend who was in Toronto when 00:48:15.120 --> 00:48:17.580 align:middle line:84% he was a PhD student, and he got her 00:48:17.580 --> 00:48:19.650 align:middle line:90% to collect a data set for him. 00:48:19.650 --> 00:48:29.280 align:middle line:84% And the way she did this was she had a, like, 25 or 30 minute car 00:48:29.280 --> 00:48:34.740 align:middle line:84% commute in Toronto, and she stuck a voice recorder 00:48:34.740 --> 00:48:36.820 align:middle line:90% in her pocket or whatever. 00:48:36.820 --> 00:48:42.060 align:middle line:84% And as she passed gas stations on the way to and from work, 00:48:42.060 --> 00:48:45.390 align:middle line:84% she just read to herself what prices they were charging, 00:48:45.390 --> 00:48:48.390 align:middle line:90% and then transcribed them. 00:48:48.390 --> 00:48:52.890 align:middle line:84% And so what she got was a twice weekly data 00:48:52.890 --> 00:48:57.930 align:middle line:84% set of prices of gas stations in Toronto 00:48:57.930 --> 00:48:59.770 align:middle line:90% which are located along-- 00:48:59.770 --> 00:49:04.150 align:middle line:84% it was like map of Toronto, she works here, she lives here, 00:49:04.150 --> 00:49:06.520 align:middle line:84% she drives this route every single day. 00:49:06.520 --> 00:49:10.980 align:middle line:84% And so you get 14 gas stations along that route 00:49:10.980 --> 00:49:13.530 align:middle line:84% and you get the twice daily prices. 00:49:13.530 --> 00:49:17.970 align:middle line:84% And this is one of those ones where just the immediate paper, 00:49:17.970 --> 00:49:24.040 align:middle line:84% this is like-- so this is a graph of her twice daily data. 00:49:24.040 --> 00:49:28.950 align:middle line:84% And what you see is that this was the wholesale price 00:49:28.950 --> 00:49:32.340 align:middle line:84% of gasoline in Toronto if you just 00:49:32.340 --> 00:49:36.060 align:middle line:84% bought it wholesale at whatever those big enormous fuel 00:49:36.060 --> 00:49:38.910 align:middle line:84% tanks are that the pump trucks back up 00:49:38.910 --> 00:49:40.750 align:middle line:90% to take it to the gas stations. 00:49:40.750 --> 00:49:45.660 align:middle line:84% And this is her data on the retail price of gasoline. 00:49:45.660 --> 00:49:47.850 align:middle line:84% And I just think just that one graph just 00:49:47.850 --> 00:49:51.450 align:middle line:84% makes it enormously clear that this is not 00:49:51.450 --> 00:49:54.750 align:middle line:84% a cost-based explanation for why gas prices change. 00:49:54.750 --> 00:49:57.780 align:middle line:84% Clearly, these guys are playing-- 00:49:57.780 --> 00:49:59.970 align:middle line:84% somehow the gas stations in Toronto 00:49:59.970 --> 00:50:04.040 align:middle line:84% are somehow playing some kind of Edgeworth cycle equilibrium, 00:50:04.040 --> 00:50:06.540 align:middle line:84% and they're playing some kind of Edgeworth cycle equilibrium 00:50:06.540 --> 00:50:11.333 align:middle line:84% where the prices just jump up really high and then go down. 00:50:11.333 --> 00:50:13.500 align:middle line:84% The firms at various points are really losing money, 00:50:13.500 --> 00:50:18.420 align:middle line:84% pumping gas, and then the prices jump up and become very large, 00:50:18.420 --> 00:50:22.242 align:middle line:84% and then they just keep cycling like that. 00:50:22.242 --> 00:50:24.700 align:middle line:84% These cycles, you might think, oh, are those weekly cycles? 00:50:24.700 --> 00:50:28.815 align:middle line:84% The frequencies there are roughly a week long, 00:50:28.815 --> 00:50:30.940 align:middle line:84% but they're not synchronized with days of the week. 00:50:30.940 --> 00:50:32.690 align:middle line:84% Sometimes they're six days long, sometimes 00:50:32.690 --> 00:50:33.850 align:middle line:90% they're eight days long. 00:50:33.850 --> 00:50:35.680 align:middle line:84% Over the course of the day, the cycles 00:50:35.680 --> 00:50:37.190 align:middle line:90% shift relative to the week. 00:50:37.190 --> 00:50:39.280 align:middle line:84% So it's some weeks the big jump comes 00:50:39.280 --> 00:50:41.530 align:middle line:84% on Mondays, some it's Wednesday, some parts of the day 00:50:41.530 --> 00:50:42.520 align:middle line:90% it's Friday. 00:50:42.520 --> 00:50:46.285 align:middle line:84% But anyway, prices are just following that cycling pattern. 00:50:46.285 --> 00:50:48.160 align:middle line:84% And actually, I find this kind of remarkable, 00:50:48.160 --> 00:50:55.320 align:middle line:84% because if you think about it, somehow, 00:50:55.320 --> 00:50:58.620 align:middle line:84% there's a interesting network geography question here of, 00:50:58.620 --> 00:51:00.180 align:middle line:90% how do they even do this? 00:51:00.180 --> 00:51:03.850 align:middle line:84% I get that, OK, this station raises its price 00:51:03.850 --> 00:51:05.970 align:middle line:84% and so then the two next to it raise their prices 00:51:05.970 --> 00:51:07.928 align:middle line:84% and undercut it the next day, and then this one 00:51:07.928 --> 00:51:12.780 align:middle line:84% undercuts this, and then these two undercut these ones 00:51:12.780 --> 00:51:15.700 align:middle line:84% and then the even ones undercut the odd ones or whatever. 00:51:15.700 --> 00:51:17.922 align:middle line:84% But somehow there are lots of chains of gas stations. 00:51:17.922 --> 00:51:19.380 align:middle line:84% You could also have driven this way 00:51:19.380 --> 00:51:21.690 align:middle line:84% and you would have gas stations here. 00:51:21.690 --> 00:51:25.860 align:middle line:84% How do you keep the cycles in sync so that-- 00:51:25.860 --> 00:51:27.360 align:middle line:84% you might think like, I go this way 00:51:27.360 --> 00:51:28.780 align:middle line:84% and I pass an even number of stations. 00:51:28.780 --> 00:51:30.360 align:middle line:84% I go this way, I pass an odd number of stations. 00:51:30.360 --> 00:51:31.600 align:middle line:90% Who's undercutting whom? 00:51:31.600 --> 00:51:34.710 align:middle line:84% But somehow the whole city geography 00:51:34.710 --> 00:51:37.950 align:middle line:84% must be such that the city hangs together 00:51:37.950 --> 00:51:43.940 align:middle line:90% and cycles together with itself. 00:51:43.940 --> 00:51:47.570 align:middle line:84% And actually, someone has another paper on Canada more 00:51:47.570 --> 00:51:51.290 align:middle line:84% generally about high frequency data, and it turned out that-- 00:51:51.290 --> 00:51:52.370 align:middle line:90% and I don't know why-- 00:51:52.370 --> 00:51:57.860 align:middle line:84% some cities in Canada do this and some do not. 00:51:57.860 --> 00:52:00.710 align:middle line:84% Some cities have these cycling prices, 00:52:00.710 --> 00:52:02.635 align:middle line:84% some do not have these cycling prices. 00:52:02.635 --> 00:52:05.360 align:middle line:90% 00:52:05.360 --> 00:52:07.165 align:middle line:84% And again, it was largely unknown 00:52:07.165 --> 00:52:09.290 align:middle line:84% until people started gathering high frequency data, 00:52:09.290 --> 00:52:12.680 align:middle line:84% because you use data at a one week frequency, 00:52:12.680 --> 00:52:16.015 align:middle line:84% you don't see this because you miss the periods in between. 00:52:16.015 --> 00:52:21.240 align:middle line:90% 00:52:21.240 --> 00:52:24.720 align:middle line:84% Another subsequent paper on this, Zhongmin Wang 00:52:24.720 --> 00:52:28.920 align:middle line:84% wrote a paper about gasoline pricing and-- yeah? 00:52:28.920 --> 00:52:31.470 align:middle line:84% AUDIENCE: If we believe like that foresight story, 00:52:31.470 --> 00:52:33.810 align:middle line:84% wouldn't we believe that these cycles would only 00:52:33.810 --> 00:52:37.860 align:middle line:84% occur with gas stations that are close to other gas stations? 00:52:37.860 --> 00:52:41.143 align:middle line:84% So if you have gas stations out away from other stations, 00:52:41.143 --> 00:52:42.810 align:middle line:84% they would have no incentive to do this. 00:52:42.810 --> 00:52:43.930 align:middle line:90% GLENN ELLISON: To do this. 00:52:43.930 --> 00:52:44.430 align:middle line:90% Yeah. 00:52:44.430 --> 00:52:46.410 align:middle line:90% So obviously, this is-- 00:52:46.410 --> 00:52:46.950 align:middle line:90% yeah. 00:52:46.950 --> 00:52:49.253 align:middle line:90% This data is urban Toronto. 00:52:49.253 --> 00:52:51.670 align:middle line:84% Toronto is a city of I don't know how many million people, 00:52:51.670 --> 00:52:54.510 align:middle line:84% so I assume that there are gas stations every-- 00:52:54.510 --> 00:52:56.553 align:middle line:84% I assume every gas station in urban Toronto 00:52:56.553 --> 00:52:58.720 align:middle line:84% is close to some other gas station in urban Toronto. 00:52:58.720 --> 00:53:00.422 align:middle line:84% And so it could be something about, 00:53:00.422 --> 00:53:02.130 align:middle line:84% yeah, the density needs to be high enough 00:53:02.130 --> 00:53:06.270 align:middle line:84% to get the things to cycle, and once you get islands 00:53:06.270 --> 00:53:07.980 align:middle line:84% without gas stations, the islands 00:53:07.980 --> 00:53:10.030 align:middle line:84% break up the cycling in the other places. 00:53:10.030 --> 00:53:12.130 align:middle line:90% Yeah. 00:53:12.130 --> 00:53:15.560 align:middle line:84% You need to think of there being price competition 00:53:15.560 --> 00:53:17.960 align:middle line:84% between adjacent stations and every station's 00:53:17.960 --> 00:53:20.173 align:middle line:84% adjacent to another one to have this to work somehow. 00:53:20.173 --> 00:53:27.630 align:middle line:90% 00:53:27.630 --> 00:53:30.240 align:middle line:84% I think these graphs are the average prices in her data set 00:53:30.240 --> 00:53:32.530 align:middle line:84% across all 14 stations or something like that, 00:53:32.530 --> 00:53:35.502 align:middle line:84% so it is like a synchronous-- at least in her commute, 00:53:35.502 --> 00:53:37.460 align:middle line:84% it's synchronous for the length of her commute. 00:53:37.460 --> 00:53:40.590 align:middle line:90% 00:53:40.590 --> 00:53:41.090 align:middle line:90% OK. 00:53:41.090 --> 00:53:44.470 align:middle line:84% So anyway, Zhongmin Wang wrote a paper then 00:53:44.470 --> 00:53:48.120 align:middle line:84% about gasoline cycling in Perth, Australia. 00:53:48.120 --> 00:53:52.682 align:middle line:84% Perth is also nice because Perth is all the way over on the West 00:53:52.682 --> 00:53:54.890 align:middle line:84% Coast of Australia where there's almost nothing else, 00:53:54.890 --> 00:53:59.030 align:middle line:84% and so Perth is like an isolated city all by itself. 00:53:59.030 --> 00:54:00.280 align:middle line:90% There's also only two-- 00:54:00.280 --> 00:54:02.770 align:middle line:84% there's actually one gasoline refinery in Perth, 00:54:02.770 --> 00:54:04.330 align:middle line:90% so everyone has-- 00:54:04.330 --> 00:54:06.472 align:middle line:84% you can import gasoline, but other than other 00:54:06.472 --> 00:54:07.930 align:middle line:84% than importing gasoline, you're all 00:54:07.930 --> 00:54:10.450 align:middle line:84% buying it from the other refinery, I believe. 00:54:10.450 --> 00:54:16.020 align:middle line:84% Anyway, and Perth had this problem of prices 00:54:16.020 --> 00:54:17.880 align:middle line:90% there were known to cycle. 00:54:17.880 --> 00:54:19.560 align:middle line:84% And prices were known to cycle, and this 00:54:19.560 --> 00:54:22.660 align:middle line:84% was seen as something that was very bad for consumers. 00:54:22.660 --> 00:54:25.102 align:middle line:90% And so the Australian-- 00:54:25.102 --> 00:54:26.310 align:middle line:90% I don't know, whatever it is. 00:54:26.310 --> 00:54:29.460 align:middle line:84% Australian Competition Authority or somebody in 2001 00:54:29.460 --> 00:54:32.970 align:middle line:84% passed a regulation to try to stop price cycling. 00:54:32.970 --> 00:54:36.870 align:middle line:84% And what they did to try to stop price cycling was a law, 00:54:36.870 --> 00:54:40.230 align:middle line:84% and the law said that gasoline stations can only change 00:54:40.230 --> 00:54:42.980 align:middle line:90% their price once a day at most. 00:54:42.980 --> 00:54:46.470 align:middle line:84% Every gasoline station must, by 2:00 PM every day, 00:54:46.470 --> 00:54:49.290 align:middle line:84% report to the government what price they will charge the next 00:54:49.290 --> 00:54:52.860 align:middle line:84% day, and they must charge that price the next day no matter 00:54:52.860 --> 00:54:53.730 align:middle line:90% what. 00:54:53.730 --> 00:54:56.935 align:middle line:84% So every day by 2:00 PM you had to submit to a government 00:54:56.935 --> 00:54:59.310 align:middle line:84% website what price you were going to charge the next day, 00:54:59.310 --> 00:55:00.900 align:middle line:90% and then your price was fixed. 00:55:00.900 --> 00:55:04.050 align:middle line:84% And they were thinking that, OK, this is going to stop the-- 00:55:04.050 --> 00:55:06.480 align:middle line:84% this could stop the price cycling because in these price 00:55:06.480 --> 00:55:09.940 align:middle line:84% cycles, you go from pricing low to pricing $0.20 above everybody 00:55:09.940 --> 00:55:10.440 align:middle line:90% else. 00:55:10.440 --> 00:55:14.050 align:middle line:84% If you do that, no one is going to buy from you. 00:55:14.050 --> 00:55:15.780 align:middle line:84% And then if we do this daily thing, 00:55:15.780 --> 00:55:18.418 align:middle line:84% firms are going to be left with zero demand for a full day. 00:55:18.418 --> 00:55:19.960 align:middle line:84% They're not going to want to do that. 00:55:19.960 --> 00:55:21.420 align:middle line:84% This is going to disrupt this price 00:55:21.420 --> 00:55:23.070 align:middle line:84% cycling which we think is keeping 00:55:23.070 --> 00:55:25.748 align:middle line:90% prices above marginal cost. 00:55:25.748 --> 00:55:27.040 align:middle line:90% Anyway, that was what they did. 00:55:27.040 --> 00:55:30.120 align:middle line:84% So anyway, and what Wang does is gather data 00:55:30.120 --> 00:55:33.933 align:middle line:84% before and after the gas price change. 00:55:33.933 --> 00:55:35.850 align:middle line:84% He had another very clever thing in his paper, 00:55:35.850 --> 00:55:38.790 align:middle line:84% which is how do I get high frequency data on the prices 00:55:38.790 --> 00:55:41.010 align:middle line:90% that gas stations are charging? 00:55:41.010 --> 00:55:44.340 align:middle line:84% Can I go to every gas station in the city of Perth 00:55:44.340 --> 00:55:46.020 align:middle line:84% and say, could you give me your price 00:55:46.020 --> 00:55:48.690 align:middle line:84% change with the exact minute by minute time 00:55:48.690 --> 00:55:51.910 align:middle line:84% at which you changed your data, you change your prices? 00:55:51.910 --> 00:55:54.222 align:middle line:84% And the answer is obviously, no, you can't do that. 00:55:54.222 --> 00:55:55.805 align:middle line:84% The stations won't give you the price. 00:55:55.805 --> 00:55:59.760 align:middle line:90% 00:55:59.760 --> 00:56:02.920 align:middle line:84% He didn't have a girlfriend in Perth. 00:56:02.920 --> 00:56:05.050 align:middle line:84% You couldn't drive around to all the gas stations, 00:56:05.050 --> 00:56:07.100 align:middle line:84% and how frequently can you do that anyway? 00:56:07.100 --> 00:56:09.600 align:middle line:84% But what he did is he made a deal with a credit card company 00:56:09.600 --> 00:56:12.700 align:middle line:84% and realized that credit card companies have 00:56:12.700 --> 00:56:15.868 align:middle line:84% people dip their credit cards at all the gas stations. 00:56:15.868 --> 00:56:17.410 align:middle line:84% And so the credit card companies know 00:56:17.410 --> 00:56:19.553 align:middle line:84% what the price was because it actually gets 00:56:19.553 --> 00:56:21.220 align:middle line:84% reported back to the credit card company 00:56:21.220 --> 00:56:26.920 align:middle line:84% that they bought 8.367 gallons or whatever, liters, at 12.99 00:56:26.920 --> 00:56:27.710 align:middle line:90% a liter. 00:56:27.710 --> 00:56:30.010 align:middle line:84% And so he made a deal with one credit card company 00:56:30.010 --> 00:56:33.620 align:middle line:84% to give him gasoline prices in Perth. 00:56:33.620 --> 00:56:36.980 align:middle line:84% Every time that they recorded a swipe in one of their cards. 00:56:36.980 --> 00:56:39.640 align:middle line:84% You don't need that big of a market share in credit cards 00:56:39.640 --> 00:56:43.540 align:middle line:84% to be basically getting most of the gas stations pretty often. 00:56:43.540 --> 00:56:46.505 align:middle line:84% Anyway, so we have these, again, very neat figures. 00:56:46.505 --> 00:56:48.255 align:middle line:84% So anyway, did I say these things already? 00:56:48.255 --> 00:56:50.840 align:middle line:90% 00:56:50.840 --> 00:56:51.340 align:middle line:90% Yep. 00:56:51.340 --> 00:56:52.590 align:middle line:90% Edgeworth cycles were present. 00:56:52.590 --> 00:56:54.000 align:middle line:90% Regulatory change. 00:56:54.000 --> 00:56:56.710 align:middle line:84% Yeah, daily prices for after the change you 00:56:56.710 --> 00:56:58.700 align:middle line:84% could get from the government website, 00:56:58.700 --> 00:57:00.610 align:middle line:84% but he got data from a credit card company 00:57:00.610 --> 00:57:02.200 align:middle line:90% for six months before. 00:57:02.200 --> 00:57:05.380 align:middle line:84% Oh, and wholesale prices you could also 00:57:05.380 --> 00:57:13.780 align:middle line:84% get because the single refinery that refined gasoline-- 00:57:13.780 --> 00:57:15.910 align:middle line:84% how do you have a market with a single refinery? 00:57:15.910 --> 00:57:17.350 align:middle line:84% What the single refinery did is it 00:57:17.350 --> 00:57:19.960 align:middle line:84% committed to what price they would charge to not exploit 00:57:19.960 --> 00:57:23.140 align:middle line:84% people, which was the price of gasoline in Singapore 00:57:23.140 --> 00:57:28.300 align:middle line:84% plus 12.6 cents a liter or something like that. 00:57:28.300 --> 00:57:30.460 align:middle line:84% By observing the wholesale price in Singapore, 00:57:30.460 --> 00:57:35.010 align:middle line:84% you actually know the wholesale price in Perth. 00:57:35.010 --> 00:57:36.548 align:middle line:90% Anyway, so here we go. 00:57:36.548 --> 00:57:39.090 align:middle line:84% This is, again, the pre-data that he got from the credit card 00:57:39.090 --> 00:57:39.940 align:middle line:90% company. 00:57:39.940 --> 00:57:44.730 align:middle line:84% This is what gasoline prices did in Perth, Australia 00:57:44.730 --> 00:57:46.140 align:middle line:90% before the law. 00:57:46.140 --> 00:57:47.760 align:middle line:84% And again, you can see there's just 00:57:47.760 --> 00:57:50.340 align:middle line:84% this striking pattern of low price, 00:57:50.340 --> 00:57:54.690 align:middle line:84% jump up, charge high prices, prices drop down. 00:57:54.690 --> 00:57:56.550 align:middle line:90% These things are a week apart. 00:57:56.550 --> 00:57:58.780 align:middle line:84% So the cycles are generally about a week long, 00:57:58.780 --> 00:58:00.810 align:middle line:84% but you can see they're not exactly week long. 00:58:00.810 --> 00:58:04.020 align:middle line:84% This one here, this one looks like it's almost nine days 00:58:04.020 --> 00:58:04.990 align:middle line:90% for that cycle. 00:58:04.990 --> 00:58:06.190 align:middle line:90% This cycle is shorter. 00:58:06.190 --> 00:58:08.190 align:middle line:90% That one's, like, six days. 00:58:08.190 --> 00:58:12.360 align:middle line:84% And you get these just massive one day increases 00:58:12.360 --> 00:58:14.730 align:middle line:90% in gasoline prices. 00:58:14.730 --> 00:58:17.260 align:middle line:84% These prices are-- these are cents per liter. 00:58:17.260 --> 00:58:18.700 align:middle line:84% So if you can't see it, this is 5. 00:58:18.700 --> 00:58:21.730 align:middle line:84% So that's, like, $0.30 a gallon of gas. 00:58:21.730 --> 00:58:24.840 align:middle line:84% So it's like the price goes from 3.39 a gallon to 3.69 00:58:24.840 --> 00:58:28.720 align:middle line:84% a gallon in the space of an hour or a few hours, 00:58:28.720 --> 00:58:32.400 align:middle line:84% and then it spends a week going back down to 3.39 00:58:32.400 --> 00:58:35.770 align:middle line:84% and then jumps back up to 3.69 again. 00:58:35.770 --> 00:58:37.060 align:middle line:90% OK. 00:58:37.060 --> 00:58:39.670 align:middle line:84% Anyway, so what does this paper say? 00:58:39.670 --> 00:58:40.437 align:middle line:90% OK, they do this. 00:58:40.437 --> 00:58:41.270 align:middle line:90% They put in the law. 00:58:41.270 --> 00:58:42.520 align:middle line:90% What happens? 00:58:42.520 --> 00:58:44.470 align:middle line:84% I didn't put in all the graphs, but shortly 00:58:44.470 --> 00:58:48.010 align:middle line:84% after the law is passed, you get a bunch of chaos. 00:58:48.010 --> 00:58:50.560 align:middle line:90% The price cycles disappear. 00:58:50.560 --> 00:58:51.988 align:middle line:90% The price cycles disappear. 00:58:51.988 --> 00:58:54.530 align:middle line:84% The firms look like they don't quite know what they're doing. 00:58:54.530 --> 00:58:56.930 align:middle line:84% They're finding a new repeated game strategy. 00:58:56.930 --> 00:59:00.040 align:middle line:84% And then within three or four months 00:59:00.040 --> 00:59:04.180 align:middle line:84% after the law has gone into effect, boom, 00:59:04.180 --> 00:59:06.460 align:middle line:84% there's the daily price cycle data. 00:59:06.460 --> 00:59:11.257 align:middle line:84% So whatever learning of how to play a repeated game strategy 00:59:11.257 --> 00:59:13.840 align:middle line:84% and how to match these things up in this new environment where 00:59:13.840 --> 00:59:17.530 align:middle line:84% you can only change your prices once a day and are stuck with 00:59:17.530 --> 00:59:22.300 align:middle line:84% whatever you said you were pricing at 2:00 PM, 00:59:22.300 --> 00:59:23.670 align:middle line:90% price cycles return. 00:59:23.670 --> 00:59:28.260 align:middle line:90% 00:59:28.260 --> 00:59:31.750 align:middle line:84% This period, he's got the wholesale price data in there, 00:59:31.750 --> 00:59:35.130 align:middle line:84% and you can see at some times they get fairly 00:59:35.130 --> 00:59:37.500 align:middle line:90% close to the wholesale price. 00:59:37.500 --> 00:59:41.010 align:middle line:84% And again, I don't know what the-- 00:59:41.010 --> 00:59:43.260 align:middle line:84% I'm not sure if this wholesale price includes shipment 00:59:43.260 --> 00:59:44.160 align:middle line:90% to the gas-- 00:59:44.160 --> 00:59:48.310 align:middle line:84% trucking fuel from the wholesale terminal to the gas station. 00:59:48.310 --> 00:59:50.880 align:middle line:84% If it doesn't, then these prices could 00:59:50.880 --> 00:59:53.190 align:middle line:84% be falling very close to their actual delivered price 00:59:53.190 --> 00:59:54.480 align:middle line:90% of the gasoline. 00:59:54.480 --> 00:59:59.577 align:middle line:84% But what you are getting is this clearly more 00:59:59.577 --> 01:00:02.160 align:middle line:84% than doubling of margins where you have a periods of time when 01:00:02.160 --> 01:00:05.190 align:middle line:84% you have a very small margin, which could even be zero 01:00:05.190 --> 01:00:08.490 align:middle line:84% depending on what the delivery charges are, and then prices 01:00:08.490 --> 01:00:13.110 align:middle line:84% jump up very high and then they drift back down. 01:00:13.110 --> 01:00:18.810 align:middle line:84% So whatever learning or whatever intuiting of repeated game 01:00:18.810 --> 01:00:21.750 align:middle line:84% strategies does not stop gasoline prices 01:00:21.750 --> 01:00:22.790 align:middle line:90% from cycling here. 01:00:22.790 --> 01:00:27.030 align:middle line:90% 01:00:27.030 --> 01:00:30.330 align:middle line:84% What he also says-- he just provides other observations that 01:00:30.330 --> 01:00:34.830 align:middle line:84% might help you think about where this cycling comes from. 01:00:34.830 --> 01:00:36.930 align:middle line:84% One is that before the regulation, 01:00:36.930 --> 01:00:39.180 align:middle line:84% these cycles had this very distinctive pattern where 01:00:39.180 --> 01:00:42.030 align:middle line:84% it was BP who was always leading the cycle. 01:00:42.030 --> 01:00:45.780 align:middle line:84% And there'd be some time between 11:00 AM and 2:00 PM 01:00:45.780 --> 01:00:48.720 align:middle line:84% on a Tuesday, Wednesday, or Thursday where BP would raise 01:00:48.720 --> 01:00:52.230 align:middle line:90% its price by $0.30 a gallon. 01:00:52.230 --> 01:00:54.630 align:middle line:84% And then Caltex, which was the biggest firm, 01:00:54.630 --> 01:00:56.820 align:middle line:84% would follow BP about two or three hours. 01:00:56.820 --> 01:00:59.220 align:middle line:84% Within two or three hours of BP raising the prices 01:00:59.220 --> 01:01:01.590 align:middle line:84% at all their stations, Caltex would bump all theirs 01:01:01.590 --> 01:01:06.390 align:middle line:84% back up to where BP's was or slightly under BP's price. 01:01:06.390 --> 01:01:08.850 align:middle line:84% And then another hour later, Shell 01:01:08.850 --> 01:01:10.960 align:middle line:90% would be up there joining them. 01:01:10.960 --> 01:01:13.773 align:middle line:84% And then the others, the other small independent stations, 01:01:13.773 --> 01:01:15.690 align:middle line:84% might take a day or two to catch up and notice 01:01:15.690 --> 01:01:17.340 align:middle line:84% all the others had gone up to the high prices 01:01:17.340 --> 01:01:18.125 align:middle line:90% and they'd do it. 01:01:18.125 --> 01:01:21.330 align:middle line:90% 01:01:21.330 --> 01:01:23.430 align:middle line:84% After this pattern, something that 01:01:23.430 --> 01:01:26.190 align:middle line:84% seems like it disrupted this was that it seemed like BP 01:01:26.190 --> 01:01:28.010 align:middle line:84% was no longer willing to do this. 01:01:28.010 --> 01:01:31.010 align:middle line:90% 01:01:31.010 --> 01:01:33.320 align:middle line:84% When BP could raise its price and then 01:01:33.320 --> 01:01:35.780 align:middle line:84% have Caltex follow two hours later, 01:01:35.780 --> 01:01:38.150 align:middle line:84% it seemed like it was happy to raise its price 01:01:38.150 --> 01:01:40.790 align:middle line:84% and sell very little gas for a few hours. 01:01:40.790 --> 01:01:45.410 align:middle line:84% It seemed unwilling to do this when 01:01:45.410 --> 01:01:49.130 align:middle line:84% it would be selling nothing for a full day and it didn't know-- 01:01:49.130 --> 01:01:49.760 align:middle line:90% maybe two days. 01:01:49.760 --> 01:01:52.790 align:middle line:84% Maybe Caltex isn't going to follow you the next day. 01:01:52.790 --> 01:01:54.680 align:middle line:84% Anyway, the pattern of leadership 01:01:54.680 --> 01:02:00.335 align:middle line:84% changes, and so now it seems like BP and Caltex 01:02:00.335 --> 01:02:02.640 align:middle line:84% share the role of being the price leaders. 01:02:02.640 --> 01:02:08.330 align:middle line:84% So maybe 40% of the price cycles are triggered by BP 01:02:08.330 --> 01:02:11.330 align:middle line:84% raising prices, 40% are triggered by Caltex 01:02:11.330 --> 01:02:14.330 align:middle line:84% raising prices, and also Shell is sometimes doing it, 01:02:14.330 --> 01:02:16.300 align:middle line:84% but again, doing it less often than the others. 01:02:16.300 --> 01:02:20.112 align:middle line:90% 01:02:20.112 --> 01:02:22.070 align:middle line:84% He argues that it does look like-- you remember 01:02:22.070 --> 01:02:24.028 align:middle line:84% I talked about the way these cycles work is you 01:02:24.028 --> 01:02:26.832 align:middle line:84% go down here, and then at the bottom in the mass control 01:02:26.832 --> 01:02:28.790 align:middle line:84% model, you have this mixed strategy equilibrium 01:02:28.790 --> 01:02:31.280 align:middle line:84% where the firms mix between staying here and getting 01:02:31.280 --> 01:02:35.090 align:middle line:84% a low profit and going here and getting no profit. 01:02:35.090 --> 01:02:41.150 align:middle line:84% He says a model with independent mixing by BP and Caltex 01:02:41.150 --> 01:02:43.470 align:middle line:84% does look like it explains what's going on. 01:02:43.470 --> 01:02:47.570 align:middle line:84% It's like they're each mixing, and with probability 0.5, 01:02:47.570 --> 01:02:50.600 align:middle line:84% they go up, and with probability 0.5, they stay down. 01:02:50.600 --> 01:02:54.590 align:middle line:84% He argues that it looks like independent mixing because you-- 01:02:54.590 --> 01:02:57.260 align:middle line:84% imagine if they were mixing 50/50 independently of each 01:02:57.260 --> 01:03:00.200 align:middle line:84% other, what you should see is a quarter of the days nobody goes 01:03:00.200 --> 01:03:03.860 align:middle line:84% up, a quarter of the days, one of the two goes up. 01:03:03.860 --> 01:03:05.568 align:middle line:84% A quarter of the days, BP and not Caltex. 01:03:05.568 --> 01:03:07.152 align:middle line:84% A quarter of the days, Caltex, not BP. 01:03:07.152 --> 01:03:09.170 align:middle line:84% And then a quarter of the time, they both go up. 01:03:09.170 --> 01:03:11.300 align:middle line:84% Looking at the frequency with which they both go up 01:03:11.300 --> 01:03:13.410 align:middle line:84% and none go up and one but not the other, 01:03:13.410 --> 01:03:16.850 align:middle line:84% it does look like they're playing independent mixtures 01:03:16.850 --> 01:03:19.140 align:middle line:84% down at the bottom, and that's the thing they've 01:03:19.140 --> 01:03:21.420 align:middle line:90% settled themselves into. 01:03:21.420 --> 01:03:25.860 align:middle line:84% And basically, if you want, did this law affect average markups 01:03:25.860 --> 01:03:27.040 align:middle line:90% over wholesale price? 01:03:27.040 --> 01:03:28.560 align:middle line:90% This law did not. 01:03:28.560 --> 01:03:31.950 align:middle line:84% So whatever you thought that this was an anti-market power 01:03:31.950 --> 01:03:34.842 align:middle line:90% law, this did not work. 01:03:34.842 --> 01:03:37.050 align:middle line:84% And you know, it did not work despite the fact also-- 01:03:37.050 --> 01:03:38.508 align:middle line:84% I guess another part of the law was 01:03:38.508 --> 01:03:42.905 align:middle line:84% that you have to report your price, 01:03:42.905 --> 01:03:44.280 align:middle line:84% and those prices are being posted 01:03:44.280 --> 01:03:47.190 align:middle line:84% by the Australian government on the internet 01:03:47.190 --> 01:03:49.170 align:middle line:84% so that anyone wanting to buy gas now 01:03:49.170 --> 01:03:51.798 align:middle line:84% has the perfect gas app because everyone's 01:03:51.798 --> 01:03:54.090 align:middle line:84% required by law to charge the same price they announced 01:03:54.090 --> 01:03:55.113 align:middle line:90% at 2 o'clock yesterday. 01:03:55.113 --> 01:03:57.030 align:middle line:84% So you can just get on your app in the evening 01:03:57.030 --> 01:04:00.960 align:middle line:84% and say, these are tomorrow's prices for every gas 01:04:00.960 --> 01:04:02.880 align:middle line:84% station in the country or these are 01:04:02.880 --> 01:04:05.570 align:middle line:84% the current prices of every gas station in the country. 01:04:05.570 --> 01:04:07.570 align:middle line:84% In the city, these are what they're going to be. 01:04:07.570 --> 01:04:10.320 align:middle line:84% Plan your trip to go to the cheapest gas stations, 01:04:10.320 --> 01:04:12.465 align:middle line:90% cause more price sensitivity. 01:04:12.465 --> 01:04:13.760 align:middle line:90% That didn't happen either. 01:04:13.760 --> 01:04:19.790 align:middle line:90% 01:04:19.790 --> 01:04:21.040 align:middle line:90% Any questions? 01:04:21.040 --> 01:04:24.360 align:middle line:84% And, you know, in some cases, maybe why shouldn't it happen? 01:04:24.360 --> 01:04:26.820 align:middle line:84% Mass control is with perfect competition. 01:04:26.820 --> 01:04:29.735 align:middle line:84% With infinitely elastic consumers, you get these cycles. 01:04:29.735 --> 01:04:32.400 align:middle line:90% 01:04:32.400 --> 01:04:34.320 align:middle line:84% You make people closer to the mass control 01:04:34.320 --> 01:04:36.300 align:middle line:84% ideal of infinitely elastic, you still 01:04:36.300 --> 01:04:38.940 align:middle line:90% get the undercutting process. 01:04:38.940 --> 01:04:41.460 align:middle line:84% So it's not at all clear that that ought-- 01:04:41.460 --> 01:04:42.960 align:middle line:90% again, in a static sense, yes. 01:04:42.960 --> 01:04:45.390 align:middle line:84% In a static model, make people more price sensitive, 01:04:45.390 --> 01:04:46.920 align:middle line:90% prices go down. 01:04:46.920 --> 01:04:48.580 align:middle line:84% But make you more price sensitive, 01:04:48.580 --> 01:04:50.940 align:middle line:84% maybe it just makes the mass control thing even more 01:04:50.940 --> 01:04:52.523 align:middle line:84% of an equilibrium than it already was. 01:04:52.523 --> 01:04:58.490 align:middle line:90% 01:04:58.490 --> 01:05:00.410 align:middle line:90% AUDIENCE: [INAUDIBLE] 01:05:00.410 --> 01:05:02.525 align:middle line:84% GLENN ELLISON: Yeah, so I haven't seen-- 01:05:02.525 --> 01:05:07.340 align:middle line:90% 01:05:07.340 --> 01:05:09.230 align:middle line:90% you know, I don't know. 01:05:09.230 --> 01:05:12.800 align:middle line:90% You guys probably don't go. 01:05:12.800 --> 01:05:14.660 align:middle line:90% Leon Musolff was giving the-- 01:05:14.660 --> 01:05:17.720 align:middle line:90% 01:05:17.720 --> 01:05:22.010 align:middle line:84% he was giving the Harvard MIT IO workshop talk. 01:05:22.010 --> 01:05:25.280 align:middle line:84% You know, he was just doing data on pricing 01:05:25.280 --> 01:05:28.310 align:middle line:90% of goods sold on Amazon. 01:05:28.310 --> 01:05:31.890 align:middle line:84% So he has very high frequency data of prices sold on Amazon. 01:05:31.890 --> 01:05:35.030 align:middle line:84% It sounded like he was saying in his Amazon data 01:05:35.030 --> 01:05:36.920 align:middle line:84% it looked like there were Edgeworth cycles. 01:05:36.920 --> 01:05:39.420 align:middle line:84% And he has an Edgeworth cycle paper that he's writing on it, 01:05:39.420 --> 01:05:42.950 align:middle line:84% but he was describing it as maybe it looks like 10% to 20% 01:05:42.950 --> 01:05:44.670 align:middle line:84% of the goods being sold on Amazon 01:05:44.670 --> 01:05:46.220 align:middle line:84% he's finding Edgeworth price cycles, 01:05:46.220 --> 01:05:48.500 align:middle line:84% and the other 80% he's finding things 01:05:48.500 --> 01:05:52.310 align:middle line:84% that look more like stable pricing. 01:05:52.310 --> 01:05:53.990 align:middle line:84% This is like people selling umbrellas 01:05:53.990 --> 01:05:56.810 align:middle line:84% on Amazon or people selling whatever. 01:05:56.810 --> 01:05:59.640 align:middle line:84% He does a lot of fashion categories. 01:05:59.640 --> 01:06:02.310 align:middle line:90% So anyway, I think-- 01:06:02.310 --> 01:06:03.040 align:middle line:90% I don't know. 01:06:03.040 --> 01:06:05.970 align:middle line:84% I haven't seen papers about high frequency data on rental car 01:06:05.970 --> 01:06:07.890 align:middle line:84% prices or airline prices or whatever 01:06:07.890 --> 01:06:10.947 align:middle line:84% to ask how many of those things look like they have short run 01:06:10.947 --> 01:06:13.530 align:middle line:84% cycles versus how many of them don't look like they don't have 01:06:13.530 --> 01:06:14.420 align:middle line:90% short run cycles. 01:06:14.420 --> 01:06:22.160 align:middle line:90% 01:06:22.160 --> 01:06:23.530 align:middle line:90% But something that does go-- 01:06:23.530 --> 01:06:26.350 align:middle line:84% in mass control, it is, in some sense-- 01:06:26.350 --> 01:06:28.000 align:middle line:84% I think mass control does say you 01:06:28.000 --> 01:06:30.610 align:middle line:84% need to have pretty price sensitive things to create 01:06:30.610 --> 01:06:31.170 align:middle line:90% these cycles. 01:06:31.170 --> 01:06:33.690 align:middle line:90% 01:06:33.690 --> 01:06:36.810 align:middle line:84% It might be the internet price comparison sites make 01:06:36.810 --> 01:06:38.503 align:middle line:84% more things like gas stations that 01:06:38.503 --> 01:06:40.920 align:middle line:84% used to be like gas stations and might cause more of these 01:06:40.920 --> 01:06:41.900 align:middle line:90% than we used to see. 01:06:41.900 --> 01:06:46.070 align:middle line:90% 01:06:46.070 --> 01:06:46.570 align:middle line:90% OK. 01:06:46.570 --> 01:06:50.740 align:middle line:84% So Then the final thing I wanted to do is give a little bit of-- 01:06:50.740 --> 01:06:54.350 align:middle line:84% again, this is an empirical part of this paper. 01:06:54.350 --> 01:06:57.080 align:middle line:84% This paper, really, it's an empirical part motivating 01:06:57.080 --> 01:06:59.780 align:middle line:84% a theory model, but I figured I would just 01:06:59.780 --> 01:07:02.990 align:middle line:84% discuss the descriptive evidence as what evidence do 01:07:02.990 --> 01:07:06.020 align:middle line:84% we have on pricing algorithms and online sites 01:07:06.020 --> 01:07:09.650 align:middle line:84% and what are pricing algorithms doing to pricing. 01:07:09.650 --> 01:07:11.870 align:middle line:84% And actually, it's more suggesting 01:07:11.870 --> 01:07:14.550 align:middle line:84% that we need different models of pricing algorithms. 01:07:14.550 --> 01:07:17.420 align:middle line:84% But so what Brown and MacKay do, again, this 01:07:17.420 --> 01:07:23.360 align:middle line:84% is a totally graduate student capable paper where they've 01:07:23.360 --> 01:07:25.200 align:middle line:90% got no proprietary data at all. 01:07:25.200 --> 01:07:27.620 align:middle line:84% They're just web scraping prices. 01:07:27.620 --> 01:07:31.580 align:middle line:84% And what they did is they set up scraping programs 01:07:31.580 --> 01:07:37.100 align:middle line:84% to collect data on an hourly basis from April 2018 01:07:37.100 --> 01:07:38.940 align:middle line:90% through October 2019. 01:07:38.940 --> 01:07:41.090 align:middle line:84% So it's about a year and a half of data. 01:07:41.090 --> 01:07:44.310 align:middle line:84% And every hour they went to five different websites. 01:07:44.310 --> 01:07:45.770 align:middle line:84% They don't give the names, but they 01:07:45.770 --> 01:07:49.340 align:middle line:84% do say these are the five most common websites in which people 01:07:49.340 --> 01:07:51.570 align:middle line:84% buy allergy medicines in the United States. 01:07:51.570 --> 01:07:55.320 align:middle line:84% So I assume the five websites are Amazon, Walmart, Target, 01:07:55.320 --> 01:07:58.390 align:middle line:84% CVS, and Walgreens, but I don't know. 01:07:58.390 --> 01:08:01.470 align:middle line:84% Anyway, they picked seven allergy drugs, Allegra, 01:08:01.470 --> 01:08:03.430 align:middle line:90% Benadryl, Claritin, et cetera. 01:08:03.430 --> 01:08:07.470 align:middle line:84% And they just scraped the prices for every different package 01:08:07.470 --> 01:08:10.200 align:middle line:90% size of those items. 01:08:10.200 --> 01:08:17.580 align:middle line:84% Because you can buy Xyzal in 80 tablet jars and 55 tablet jars 01:08:17.580 --> 01:08:20.130 align:middle line:84% and 30 tablet jars and something or other, 01:08:20.130 --> 01:08:26.319 align:middle line:84% these seven drugs create, like, 150 different products. 01:08:26.319 --> 01:08:28.300 align:middle line:84% There are a lot of different packages. 01:08:28.300 --> 01:08:31.890 align:middle line:84% Many of the stores only sell the three most common packages 01:08:31.890 --> 01:08:33.990 align:middle line:90% for them. 01:08:33.990 --> 01:08:37.080 align:middle line:84% But anyway, so what they scraped is every price 01:08:37.080 --> 01:08:39.300 align:middle line:84% that they could find for any version 01:08:39.300 --> 01:08:42.399 align:middle line:84% of these drugs every hour for a year and a half. 01:08:42.399 --> 01:08:44.500 align:middle line:84% And then they're trying to record basically, 01:08:44.500 --> 01:08:47.880 align:middle line:84% what happens to learn about the pricing algorithms 01:08:47.880 --> 01:08:49.950 align:middle line:90% that these firms are doing. 01:08:49.950 --> 01:08:52.800 align:middle line:84% Thinking about the q-learning theory. 01:08:52.800 --> 01:08:55.120 align:middle line:84% Are these firms running q-learning algorithms? 01:08:55.120 --> 01:08:56.245 align:middle line:90% What are they doing? 01:08:56.245 --> 01:08:58.620 align:middle line:84% What can we pick up from just observing the price changes 01:08:58.620 --> 01:09:00.953 align:middle line:84% that they're making as a function of their rivals prices 01:09:00.953 --> 01:09:03.540 align:middle line:90% and so on? 01:09:03.540 --> 01:09:07.000 align:middle line:84% There's clear heterogeneity in practices across website. 01:09:07.000 --> 01:09:09.569 align:middle line:84% There's one retailer they call A. 01:09:09.569 --> 01:09:12.420 align:middle line:84% I'm going to guess that that's Amazon. 01:09:12.420 --> 01:09:14.593 align:middle line:84% Retailer A is very different from the others. 01:09:14.593 --> 01:09:16.260 align:middle line:84% One thing that retailer A is, is clearly 01:09:16.260 --> 01:09:17.859 align:middle line:90% is much bigger than the others. 01:09:17.859 --> 01:09:22.689 align:middle line:84% Every other website has roughly 50 of these drugs available. 01:09:22.689 --> 01:09:25.830 align:middle line:84% Retailer A has 150, so retailer A just 01:09:25.830 --> 01:09:28.452 align:middle line:90% has many, many more versions. 01:09:28.452 --> 01:09:29.910 align:middle line:84% Retailer A seems to have just about 01:09:29.910 --> 01:09:32.910 align:middle line:84% everything in the world that's sold is on retailer A, 01:09:32.910 --> 01:09:35.220 align:middle line:90% whereas it's not on the others. 01:09:35.220 --> 01:09:37.800 align:middle line:84% Retailer A is also changing its prices 01:09:37.800 --> 01:09:41.290 align:middle line:90% 1.89 times per day per product. 01:09:41.290 --> 01:09:47.520 align:middle line:84% So if you ask, do you want to buy 80 tablets of Xyzal, 01:09:47.520 --> 01:09:50.290 align:middle line:84% how long do you have to wait for the price of Xyzal to change? 01:09:50.290 --> 01:09:54.800 align:middle line:84% The answer is, like, 15 hours the price is going to change, 01:09:54.800 --> 01:09:56.640 align:middle line:84% and it just changes over and over again. 01:09:56.640 --> 01:09:59.180 align:middle line:84% So when they graph the raw price data, 01:09:59.180 --> 01:10:03.630 align:middle line:84% the black is the retailer A's prices. 01:10:03.630 --> 01:10:05.270 align:middle line:90% Yeah? 01:10:05.270 --> 01:10:07.430 align:middle line:84% AUDIENCE: Anecdotally, I know when you-- 01:10:07.430 --> 01:10:10.340 align:middle line:84% I think I've heard that sometimes when 01:10:10.340 --> 01:10:12.830 align:middle line:84% you look at Amazon a lot and keep putting, 01:10:12.830 --> 01:10:14.390 align:middle line:84% they just change the price for you. 01:10:14.390 --> 01:10:15.510 align:middle line:84% GLENN ELLISON: For you personally. 01:10:15.510 --> 01:10:17.302 align:middle line:84% AUDIENCE: So could it just be that Amazon's 01:10:17.302 --> 01:10:18.620 align:middle line:90% doing personalized pricing. 01:10:18.620 --> 01:10:20.300 align:middle line:84% GLENN ELLISON: So it could be they're 01:10:20.300 --> 01:10:25.970 align:middle line:84% getting something that Amazon is picking up their program. 01:10:25.970 --> 01:10:27.620 align:middle line:90% And I don't know-- 01:10:27.620 --> 01:10:29.660 align:middle line:84% I don't know the details of what kind of program 01:10:29.660 --> 01:10:31.862 align:middle line:84% they were using to access this data 01:10:31.862 --> 01:10:34.070 align:middle line:84% and whether Amazon could have recognized that program 01:10:34.070 --> 01:10:35.870 align:middle line:90% and done something to it. 01:10:35.870 --> 01:10:40.400 align:middle line:84% But anyway, this black is the graph 01:10:40.400 --> 01:10:43.220 align:middle line:84% of what Amazon was returning to them, 01:10:43.220 --> 01:10:46.430 align:middle line:84% which is these prices changing quite frequently. 01:10:46.430 --> 01:10:49.490 align:middle line:84% Other thing you actually notice is price levels differ. 01:10:49.490 --> 01:10:54.840 align:middle line:84% It looks like this A was changing prices all the time 01:10:54.840 --> 01:10:57.450 align:middle line:84% and substantially undercutting what 01:10:57.450 --> 01:11:01.590 align:middle line:84% the red and the green and the blue firm are doing. 01:11:01.590 --> 01:11:02.490 align:middle line:90% OK? 01:11:02.490 --> 01:11:04.920 align:middle line:84% The other one that changes price fairly frequently 01:11:04.920 --> 01:11:08.100 align:middle line:84% is the green firm, B. B changes prices, 01:11:08.100 --> 01:11:10.890 align:middle line:90% on average, 0.28 times per day. 01:11:10.890 --> 01:11:15.018 align:middle line:84% And so firm B is this green line here that's doing that there. 01:11:15.018 --> 01:11:17.310 align:middle line:84% At the end here it's harder to see because the green is 01:11:17.310 --> 01:11:19.080 align:middle line:90% more mixed in with the black. 01:11:19.080 --> 01:11:21.030 align:middle line:84% And then you have the three firms, C, D, 01:11:21.030 --> 01:11:23.670 align:middle line:84% and E, that are the ones up here. 01:11:23.670 --> 01:11:27.330 align:middle line:84% E is really like this blue line that's largely here. 01:11:27.330 --> 01:11:31.410 align:middle line:84% Red and the orange ones are going up and down occasionally. 01:11:31.410 --> 01:11:34.230 align:middle line:84% And there are gaps in this data because you write a program 01:11:34.230 --> 01:11:37.650 align:middle line:84% to scrape data every hour and then the websites change 01:11:37.650 --> 01:11:40.578 align:middle line:84% their interfaces and you don't notice it for a while 01:11:40.578 --> 01:11:42.120 align:middle line:84% that your program is broken, and then 01:11:42.120 --> 01:11:43.770 align:middle line:90% you have to fix your program. 01:11:43.770 --> 01:11:45.702 align:middle line:84% Many, many times during a year and a half, 01:11:45.702 --> 01:11:47.160 align:middle line:84% they changed their website in a way 01:11:47.160 --> 01:11:48.785 align:middle line:84% that your scraping program doesn't work 01:11:48.785 --> 01:11:52.540 align:middle line:84% and you have to fix it, and then you have missing data. 01:11:52.540 --> 01:11:54.430 align:middle line:90% So some findings. 01:11:54.430 --> 01:11:56.680 align:middle line:84% Firms seem to be updating on a schedule 01:11:56.680 --> 01:11:59.010 align:middle line:84% and there seem to be two technologies going on here. 01:11:59.010 --> 01:12:02.620 align:middle line:90% So remember, this is 2018, 2019. 01:12:02.620 --> 01:12:06.790 align:middle line:84% Firms A and B seem to be running some kind of price updating 01:12:06.790 --> 01:12:10.600 align:middle line:84% program that's roughly equally likely to change 01:12:10.600 --> 01:12:15.160 align:middle line:84% the price at any hour of the day or night on any day of the week. 01:12:15.160 --> 01:12:18.160 align:middle line:84% So these are the probability that a price change occurs 01:12:18.160 --> 01:12:19.840 align:middle line:84% in a particular hour of the week, 01:12:19.840 --> 01:12:26.850 align:middle line:84% running from Saturday midnight to Friday, 11:59 PM or whatever. 01:12:26.850 --> 01:12:27.490 align:middle line:90% OK? 01:12:27.490 --> 01:12:29.590 align:middle line:84% A and B seem to be having these programs that 01:12:29.590 --> 01:12:31.753 align:middle line:84% just run all the time that change their prices. 01:12:31.753 --> 01:12:34.420 align:middle line:90% 01:12:34.420 --> 01:12:38.860 align:middle line:84% Retailer C seems to run a program between 3:00 and 6:00 AM 01:12:38.860 --> 01:12:41.180 align:middle line:90% and sometimes change prices. 01:12:41.180 --> 01:12:44.680 align:middle line:84% Again, remember, retailer C is changing prices roughly-- 01:12:44.680 --> 01:12:47.920 align:middle line:84% each individual drug's price changes roughly once a month 01:12:47.920 --> 01:12:53.670 align:middle line:84% at retailer C. So they seem to change prices roughly 01:12:53.670 --> 01:12:54.400 align:middle line:90% once a month. 01:12:54.400 --> 01:12:58.140 align:middle line:84% When they do change prices, it's between 3:00 and 6:00 AM on some 01:12:58.140 --> 01:13:00.690 align:middle line:84% day of the week, and all days they're doing price changing 01:13:00.690 --> 01:13:02.390 align:middle line:90% activity. 01:13:02.390 --> 01:13:07.330 align:middle line:84% And retailers D and E are changing their prices 01:13:07.330 --> 01:13:12.940 align:middle line:84% at midnight Eastern time on Sunday, going into-- 01:13:12.940 --> 01:13:15.250 align:middle line:84% Saturday night, going into Sunday morning. 01:13:15.250 --> 01:13:17.500 align:middle line:84% So slightly after midnight Saturday 01:13:17.500 --> 01:13:20.800 align:middle line:84% night, they do change their prices. 01:13:20.800 --> 01:13:25.330 align:middle line:84% Retailer E, it's almost always right around midnight, 01:13:25.330 --> 01:13:28.810 align:middle line:84% and then retailer D, it's between midnight and 3:00 AM 01:13:28.810 --> 01:13:31.050 align:middle line:90% or something the prices change. 01:13:31.050 --> 01:13:34.890 align:middle line:84% So clearly these firms are using different technologies, 01:13:34.890 --> 01:13:36.720 align:middle line:84% and it seems like two of them are 01:13:36.720 --> 01:13:38.970 align:middle line:84% using a very old fashioned technology of having 01:13:38.970 --> 01:13:41.610 align:middle line:84% a meeting where they get together and discuss what prices 01:13:41.610 --> 01:13:46.580 align:middle line:84% should be, and then others are doing the change it more often. 01:13:46.580 --> 01:13:52.830 align:middle line:90% 01:13:52.830 --> 01:13:56.130 align:middle line:84% Are these pricing algorithms like the pricing algorithms 01:13:56.130 --> 01:13:58.920 align:middle line:84% I discussed in the q-learning theory 01:13:58.920 --> 01:14:03.690 align:middle line:84% where your price depends on the other firm's prices 01:14:03.690 --> 01:14:05.440 align:middle line:90% and you're reacting to them? 01:14:05.440 --> 01:14:08.520 align:middle line:84% And yes, firms A and B, their algorithms 01:14:08.520 --> 01:14:09.850 align:middle line:90% are reacting to prices. 01:14:09.850 --> 01:14:12.330 align:middle line:90% What they do here is that-- 01:14:12.330 --> 01:14:16.020 align:middle line:84% you remember D makes those changes very, very rarely. 01:14:16.020 --> 01:14:19.000 align:middle line:84% D is making those changes once every two or three months, 01:14:19.000 --> 01:14:22.110 align:middle line:84% and it's always making those changes just after midnight. 01:14:22.110 --> 01:14:26.190 align:middle line:84% And so they ask, do A and B do something 01:14:26.190 --> 01:14:32.700 align:middle line:84% different on those rare Sundays when D changes its price? 01:14:32.700 --> 01:14:37.690 align:middle line:84% And what you find here is that, yes, A and B are reacting to D. 01:14:37.690 --> 01:14:41.950 align:middle line:84% But A and B are not reacting all that quickly to D. 01:14:41.950 --> 01:14:44.770 align:middle line:84% So it's not like they're doing hourly reactions. 01:14:44.770 --> 01:14:48.530 align:middle line:84% So this is, is Amazon reacting to what goes on in cvs.com? 01:14:48.530 --> 01:14:52.095 align:middle line:84% And the answer is yes, but not for a while. 01:14:52.095 --> 01:14:53.470 align:middle line:84% So if you look at the probability 01:14:53.470 --> 01:14:58.630 align:middle line:84% that there will be a price change after D changes 01:14:58.630 --> 01:15:01.870 align:middle line:84% its price, and the answer is for 24 to 36 hours, 01:15:01.870 --> 01:15:04.630 align:middle line:84% there's no additional probability 01:15:04.630 --> 01:15:06.430 align:middle line:84% that A has changed its price based 01:15:06.430 --> 01:15:08.480 align:middle line:90% on D having changed its price. 01:15:08.480 --> 01:15:11.350 align:middle line:84% But then starting in about 36 hours, 01:15:11.350 --> 01:15:14.740 align:middle line:84% going up to 48 or 60 hours, yes, they 01:15:14.740 --> 01:15:17.080 align:middle line:84% are then reacting to what D does. 01:15:17.080 --> 01:15:18.850 align:middle line:84% And there's more probability that they 01:15:18.850 --> 01:15:24.430 align:middle line:84% will change their price in the 36 to 72 hour window 01:15:24.430 --> 01:15:27.930 align:middle line:84% after D has changed its price than there would be otherwise. 01:15:27.930 --> 01:15:32.720 align:middle line:84% And so again, 36 to 72 hours later means looking, 01:15:32.720 --> 01:15:34.360 align:middle line:90% it's Saturday midnight. 01:15:34.360 --> 01:15:36.110 align:middle line:90% D changes its price. 01:15:36.110 --> 01:15:40.550 align:middle line:84% And so then it's Monday at noon to Tuesday at noon. 01:15:40.550 --> 01:15:42.860 align:middle line:84% A is more likely to change its price if D has made 01:15:42.860 --> 01:15:44.580 align:middle line:90% a price change in that period. 01:15:44.580 --> 01:15:48.290 align:middle line:84% So it looks like A is running a price responsive algorithm, 01:15:48.290 --> 01:15:50.150 align:middle line:84% but it's not something that would 01:15:50.150 --> 01:15:52.910 align:middle line:84% cause the mass control gasoline price cycles 01:15:52.910 --> 01:15:55.160 align:middle line:84% thing that CVS changes its price, 01:15:55.160 --> 01:15:56.610 align:middle line:84% they're immediately undercutting. 01:15:56.610 --> 01:15:59.017 align:middle line:90% It seems like a slower undercut. 01:15:59.017 --> 01:16:01.100 align:middle line:84% But if you think about that, that might make sense 01:16:01.100 --> 01:16:02.683 align:middle line:84% because if you know that D isn't going 01:16:02.683 --> 01:16:05.900 align:middle line:84% to change its price for a month, you can just-- or two months, 01:16:05.900 --> 01:16:08.150 align:middle line:84% you can figure out, what do we want to do now 01:16:08.150 --> 01:16:09.840 align:middle line:90% that D is cheaper than us? 01:16:09.840 --> 01:16:12.770 align:middle line:84% We'll just pick a price and then we'll stick-- 01:16:12.770 --> 01:16:14.790 align:middle line:84% and then we'll be optimally responding to them. 01:16:14.790 --> 01:16:15.980 align:middle line:84% We don't need to respond to high frequency 01:16:15.980 --> 01:16:17.095 align:middle line:90% because d isn't changing. 01:16:17.095 --> 01:16:19.910 align:middle line:90% 01:16:19.910 --> 01:16:22.160 align:middle line:84% Also what they find in this paper is 01:16:22.160 --> 01:16:26.990 align:middle line:84% that the frequency of price changes 01:16:26.990 --> 01:16:29.990 align:middle line:84% is clearly related to the prices being charged. 01:16:29.990 --> 01:16:34.040 align:middle line:84% So if you make it price index where A's price for a product 01:16:34.040 --> 01:16:41.000 align:middle line:84% is normalized to one, B is charging 5% more than A, 01:16:41.000 --> 01:16:43.880 align:middle line:84% C is charging 10% more than A, and D and E 01:16:43.880 --> 01:16:47.570 align:middle line:84% are charging roughly 30% more than A for the same product. 01:16:47.570 --> 01:16:50.960 align:middle line:84% And those price levels seem to line up 01:16:50.960 --> 01:16:53.790 align:middle line:84% with the frequency of the price effects. 01:16:53.790 --> 01:16:56.330 align:middle line:84% But it's not just that D and E change their prices much less 01:16:56.330 --> 01:16:59.040 align:middle line:84% often and only do it on Sunday mornings occasionally. 01:16:59.040 --> 01:17:01.910 align:middle line:84% They're also just setting higher prices. 01:17:01.910 --> 01:17:04.970 align:middle line:84% And then B and C, which are changing prices 01:17:04.970 --> 01:17:08.330 align:middle line:84% more often than D and E but not as often as A, 01:17:08.330 --> 01:17:13.340 align:middle line:84% they're setting intermediate prices on average. 01:17:13.340 --> 01:17:16.670 align:middle line:84% And so motivated by these findings, what the paper then 01:17:16.670 --> 01:17:19.340 align:middle line:84% has is it's got this theory section looking 01:17:19.340 --> 01:17:22.460 align:middle line:84% at saying this question is what prices would 01:17:22.460 --> 01:17:25.190 align:middle line:84% we expect if firms had different pricing technologies 01:17:25.190 --> 01:17:27.680 align:middle line:84% and some were faster than the others? 01:17:27.680 --> 01:17:33.020 align:middle line:84% And both A is best reacting to D and D is also optimally pricing, 01:17:33.020 --> 01:17:35.360 align:middle line:84% knowing that we're going to set a price and then A 01:17:35.360 --> 01:17:37.160 align:middle line:84% is going to undercut us and then we're 01:17:37.160 --> 01:17:39.980 align:middle line:84% going to be stuck for a month being undercut by A. 01:17:39.980 --> 01:17:42.700 align:middle line:90% How should we price? 01:17:42.700 --> 01:17:44.890 align:middle line:84% And what they argue is that if you think about that, 01:17:44.890 --> 01:17:46.473 align:middle line:84% even just think about the static game. 01:17:46.473 --> 01:17:48.243 align:middle line:84% We're going to play this game just once. 01:17:48.243 --> 01:17:49.660 align:middle line:84% I'm going to set my price and then 01:17:49.660 --> 01:17:50.710 align:middle line:84% you're going to get to undercut me 01:17:50.710 --> 01:17:51.550 align:middle line:84% and I'm going to be stuck because I 01:17:51.550 --> 01:17:53.690 align:middle line:84% have an inferior updating technology. 01:17:53.690 --> 01:17:55.630 align:middle line:90% How should you price? 01:17:55.630 --> 01:17:57.100 align:middle line:84% And what they argue is that if you 01:17:57.100 --> 01:18:01.150 align:middle line:84% think about a differentiated product like Hotelling model 01:18:01.150 --> 01:18:04.060 align:middle line:84% or logit demand model, what you're going to do 01:18:04.060 --> 01:18:06.430 align:middle line:84% is the Stackelberg leader here, the person who moves 01:18:06.430 --> 01:18:09.790 align:middle line:84% first is disadvantaged, because you set your price first 01:18:09.790 --> 01:18:11.420 align:middle line:84% and then the other guy undercuts you. 01:18:11.420 --> 01:18:14.420 align:middle line:90% But what are you going to do? 01:18:14.420 --> 01:18:17.500 align:middle line:84% You're going to price higher than the Nash equilibrium price 01:18:17.500 --> 01:18:19.150 align:middle line:84% knowing you're going to be undercut, 01:18:19.150 --> 01:18:20.620 align:middle line:84% and the equilibrium is going to be 01:18:20.620 --> 01:18:25.510 align:middle line:84% where the firm that moves slow moves first, 01:18:25.510 --> 01:18:28.930 align:middle line:84% prices well above the Nash equilibrium, 01:18:28.930 --> 01:18:31.150 align:middle line:84% trying to induce the firm, undercutting 01:18:31.150 --> 01:18:34.250 align:middle line:84% it to also undercut to an above Nash price. 01:18:34.250 --> 01:18:36.970 align:middle line:84% And we end up with two above Nash prices 01:18:36.970 --> 01:18:40.330 align:middle line:84% where the initial early mover is disadvantaged, 01:18:40.330 --> 01:18:44.600 align:middle line:84% but it's disadvantaged, but it sets a high price 01:18:44.600 --> 01:18:48.090 align:middle line:84% to try to induce the other firm to go high underneath it. 01:18:48.090 --> 01:18:50.270 align:middle line:84% And knowing that you're going to be undercut 01:18:50.270 --> 01:18:52.010 align:middle line:84% causes you to price higher than you would 01:18:52.010 --> 01:18:54.060 align:middle line:84% in a simultaneous move pricing game. 01:18:54.060 --> 01:18:59.840 align:middle line:84% And they argue that this asymmetric speed of adjustment 01:18:59.840 --> 01:19:02.930 align:middle line:84% could be something that leads to super competitive prices 01:19:02.930 --> 01:19:06.050 align:middle line:84% where, in their model, even A is pricing above the static Nash 01:19:06.050 --> 01:19:06.980 align:middle line:90% point. 01:19:06.980 --> 01:19:10.220 align:middle line:84% But it's like D and E are pricing well above static Nash 01:19:10.220 --> 01:19:12.962 align:middle line:84% to try to pull A up knowing that they're going to be undercut, 01:19:12.962 --> 01:19:14.420 align:middle line:84% but at least they're being undercut 01:19:14.420 --> 01:19:18.090 align:middle line:84% and getting a moderate market share at a higher price. 01:19:18.090 --> 01:19:21.050 align:middle line:84% And again, like in pure Bertrand, 01:19:21.050 --> 01:19:22.230 align:middle line:90% this would not work well. 01:19:22.230 --> 01:19:26.060 align:middle line:84% But if you think about this as like CVS competing with Amazon, 01:19:26.060 --> 01:19:29.190 align:middle line:84% most people buying from CVS are just going to cvs.com. 01:19:29.190 --> 01:19:32.450 align:middle line:84% There's got to be some cross elasticity between cvs.com 01:19:32.450 --> 01:19:33.420 align:middle line:90% and Amazon. 01:19:33.420 --> 01:19:36.470 align:middle line:84% But a lot of people just buying Zyrtec at CVS 01:19:36.470 --> 01:19:39.020 align:middle line:84% just go to cvs.com, they buy their Zyrtec. 01:19:39.020 --> 01:19:42.740 align:middle line:84% And so if you have these finite price elasticities, 01:19:42.740 --> 01:19:44.690 align:middle line:84% they argue this pattern is something 01:19:44.690 --> 01:19:48.630 align:middle line:84% you might expect to see of slow price adjustment 01:19:48.630 --> 01:19:51.035 align:middle line:84% being a factor that can lead to higher price levels. 01:19:51.035 --> 01:19:56.770 align:middle line:90% 01:19:56.770 --> 01:19:58.030 align:middle line:90% Any questions? 01:19:58.030 --> 01:19:59.810 align:middle line:90% So that is it for today. 01:19:59.810 --> 01:20:01.430 align:middle line:90% So, yeah. 01:20:01.430 --> 01:20:02.940 align:middle line:90% Next week's topic is entry. 01:20:02.940 --> 01:20:06.640 align:middle line:84% Again, entry theory, Monday, empirics, Wednesday. 01:20:06.640 --> 01:20:10.550 align:middle line:84% And Monday's lecture, really, it's a lot of textbook material. 01:20:10.550 --> 01:20:17.710 align:middle line:84% I will be covering the Nikhil Vellodi paper towards the end, 01:20:17.710 --> 01:20:20.850 align:middle line:84% but it'll mostly be stuff from Tirole's book. 01:20:20.850 --> 01:20:44.000 align:middle line:90%