WEBVTT 00:00:00.000 --> 00:00:01.988 align:middle line:90% [SQUEAKING] 00:00:01.988 --> 00:00:03.479 align:middle line:90% [RUSTLING] 00:00:03.479 --> 00:00:04.970 align:middle line:90% [CLICKING] 00:00:04.970 --> 00:00:16.655 align:middle line:90% 00:00:16.655 --> 00:00:18.530 align:middle line:84% GLENN ELLISON: OK, anyway, so online markets. 00:00:18.530 --> 00:00:23.000 align:middle line:84% So what I've got several different papers 00:00:23.000 --> 00:00:24.820 align:middle line:90% I figured I'd talk about today. 00:00:24.820 --> 00:00:28.030 align:middle line:84% All of them are talking about online retail 00:00:28.030 --> 00:00:33.670 align:middle line:84% in one form or another and just different issues 00:00:33.670 --> 00:00:36.110 align:middle line:84% in online retail trying to understand the industry. 00:00:36.110 --> 00:00:39.910 align:middle line:84% So the first paper I was going to talk about, Einav, Kuchler, 00:00:39.910 --> 00:00:41.800 align:middle line:84% Levin, Sundaresan, Learning from Seller 00:00:41.800 --> 00:00:43.600 align:middle line:90% Experiments in Online Markets. 00:00:43.600 --> 00:00:47.980 align:middle line:90% This is a paper studying eBay. 00:00:47.980 --> 00:00:52.570 align:middle line:84% And while eBay has been in a long run decline, 00:00:52.570 --> 00:00:55.430 align:middle line:84% it's still an important way in which things get sold. 00:00:55.430 --> 00:00:56.860 align:middle line:84% And I also just wanted to cover it 00:00:56.860 --> 00:01:01.120 align:middle line:84% because I felt the methodology of the paper 00:01:01.120 --> 00:01:04.690 align:middle line:84% is interesting as a way to attack problems and these ideas 00:01:04.690 --> 00:01:08.470 align:middle line:84% that what they're talking about is an alternative to doing field 00:01:08.470 --> 00:01:12.340 align:middle line:84% experiments or setting up agreements to run experiments 00:01:12.340 --> 00:01:14.290 align:middle line:84% with online service providers is what they 00:01:14.290 --> 00:01:17.870 align:middle line:90% call opportunistic experiments. 00:01:17.870 --> 00:01:20.040 align:middle line:84% There are many, many businesses in this world. 00:01:20.040 --> 00:01:22.830 align:middle line:84% Those businesses are trying to operate optimally. 00:01:22.830 --> 00:01:25.170 align:middle line:84% They run lots of experiments themselves. 00:01:25.170 --> 00:01:28.310 align:middle line:84% And so their idea is to take advantage 00:01:28.310 --> 00:01:31.010 align:middle line:84% of all these experiments that sellers are running on eBay 00:01:31.010 --> 00:01:32.430 align:middle line:90% and analyze those. 00:01:32.430 --> 00:01:35.593 align:middle line:84% And obviously, you know, you get less controlled experiments 00:01:35.593 --> 00:01:37.760 align:middle line:84% than if you were setting up the experiment yourself. 00:01:37.760 --> 00:01:42.020 align:middle line:84% Each experiment is not ideal, but, you know, there's thousands 00:01:42.020 --> 00:01:43.910 align:middle line:84% and thousands of not ideal experiments 00:01:43.910 --> 00:01:48.320 align:middle line:84% may be more informative and cheaper 00:01:48.320 --> 00:01:51.140 align:middle line:84% to run than just one experiment you design yourself. 00:01:51.140 --> 00:01:52.970 align:middle line:84% So to start with, they note that, you know, 00:01:52.970 --> 00:01:56.690 align:middle line:84% very few things on eBay are actually unique. 00:01:56.690 --> 00:01:59.720 align:middle line:84% They said, you know, I guess interesting fact about eBay, 00:01:59.720 --> 00:02:02.450 align:middle line:84% for about a half of the items that are sold on eBay, 00:02:02.450 --> 00:02:04.730 align:middle line:84% you can find a nearly identical item sold 00:02:04.730 --> 00:02:07.259 align:middle line:90% by exactly the same seller. 00:02:07.259 --> 00:02:09.259 align:middle line:84% And so when they're talking about an experiment, 00:02:09.259 --> 00:02:14.270 align:middle line:84% an experiment meaning you have exactly the same seller 00:02:14.270 --> 00:02:17.210 align:middle line:84% selling a product with exactly the same listing 00:02:17.210 --> 00:02:20.190 align:middle line:84% title listed in the same category and subtitle. 00:02:20.190 --> 00:02:23.540 align:middle line:84% And they find 55 million seller experiments 00:02:23.540 --> 00:02:27.680 align:middle line:84% involving a total of 350 million listings. 00:02:27.680 --> 00:02:30.473 align:middle line:84% For much of their analysis, they're 00:02:30.473 --> 00:02:32.390 align:middle line:84% restricting attention to experiments that have 00:02:32.390 --> 00:02:34.020 align:middle line:90% several different features. 00:02:34.020 --> 00:02:35.970 align:middle line:84% There are at least two items auctioned. 00:02:35.970 --> 00:02:38.885 align:middle line:84% There's also at least one successful posted price sale. 00:02:38.885 --> 00:02:40.760 align:middle line:84% They want a posted price sale because they're 00:02:40.760 --> 00:02:42.590 align:middle line:84% going to use that as a reference for what's 00:02:42.590 --> 00:02:44.550 align:middle line:84% the market price of this if you just sell it. 00:02:44.550 --> 00:02:48.000 align:middle line:90% So not through an auction. 00:02:48.000 --> 00:02:49.580 align:middle line:84% And even with those restrictions, 00:02:49.580 --> 00:02:53.910 align:middle line:84% they come up with 244,000 experiments that have 7.6 00:02:53.910 --> 00:02:55.440 align:middle line:90% million listings. 00:02:55.440 --> 00:02:58.370 align:middle line:84% So the idea is just that eBay sellers are just 00:02:58.370 --> 00:03:01.110 align:middle line:84% experimenting all the time with all kinds of things. 00:03:01.110 --> 00:03:03.200 align:middle line:84% And this is something we can take advantage of. 00:03:03.200 --> 00:03:09.350 align:middle line:84% The paper itself has multiple different sub papers discussing 00:03:09.350 --> 00:03:10.320 align:middle line:90% what to do with this. 00:03:10.320 --> 00:03:14.570 align:middle line:84% So one of the sub papers is how do "Buy It Now" 00:03:14.570 --> 00:03:16.460 align:middle line:90% prices affect outcomes? 00:03:16.460 --> 00:03:18.790 align:middle line:84% On eBay, if you've never been on an eBay auction, 00:03:18.790 --> 00:03:22.030 align:middle line:84% it's like it's classic English ascending bid auction, 00:03:22.030 --> 00:03:24.190 align:middle line:84% but there's also this option of just click 00:03:24.190 --> 00:03:27.350 align:middle line:84% here to buy it for $125 and end the auction. 00:03:27.350 --> 00:03:28.900 align:middle line:84% And so one of the papers they do is 00:03:28.900 --> 00:03:33.310 align:middle line:84% how does having a Buy It Now price affect the outcome? 00:03:33.310 --> 00:03:36.230 align:middle line:84% Second is, how does shipping fees affect outcomes? 00:03:36.230 --> 00:03:40.540 align:middle line:84% There's this interesting earlier paper by Hossain and Morgan 00:03:40.540 --> 00:03:44.560 align:middle line:84% that argued that consumers are fairly naive in dealing 00:03:44.560 --> 00:03:47.590 align:middle line:84% with shipping fees and that you make more money by having 00:03:47.590 --> 00:03:52.430 align:middle line:84% good with $9.95 shipping than good with $5.95 shipping. 00:03:52.430 --> 00:03:56.495 align:middle line:84% Just somehow you offer the same good, different shipping prices, 00:03:56.495 --> 00:03:58.120 align:middle line:84% people somehow focus on the item price, 00:03:58.120 --> 00:03:59.210 align:middle line:90% not on the shipping price. 00:03:59.210 --> 00:04:00.918 align:middle line:84% And so the plus shipping and handling you 00:04:00.918 --> 00:04:04.480 align:middle line:84% make more money selling with a high shipping and handling fee. 00:04:04.480 --> 00:04:07.390 align:middle line:84% Anyway, they're going to do an experiment on that 00:04:07.390 --> 00:04:09.545 align:middle line:84% or many sellers have done experiments on that. 00:04:09.545 --> 00:04:10.920 align:middle line:84% And then they also have this sort 00:04:10.920 --> 00:04:14.410 align:middle line:84% of third paper about how and why the reserve prices affect 00:04:14.410 --> 00:04:15.500 align:middle line:90% revenues. 00:04:15.500 --> 00:04:17.600 align:middle line:84% I'm just going to talk about the third one 00:04:17.600 --> 00:04:20.269 align:middle line:84% because it feeds back to things I talked about in the auction 00:04:20.269 --> 00:04:24.780 align:middle line:84% literature about optimal reserve prices and auctions. 00:04:24.780 --> 00:04:27.330 align:middle line:84% OK, here's just their motivating figure 00:04:27.330 --> 00:04:30.270 align:middle line:90% to explain eBay experiments. 00:04:30.270 --> 00:04:34.710 align:middle line:84% Like, so they did a search for "TaylorMade driver." 00:04:34.710 --> 00:04:38.155 align:middle line:84% TaylorMade is a brand of golf club manufacturer. 00:04:38.155 --> 00:04:39.780 align:middle line:84% You do a search for "TaylorMade driver" 00:04:39.780 --> 00:04:43.410 align:middle line:84% on eBay and you just see a screen like this just full 00:04:43.410 --> 00:04:49.360 align:middle line:84% of people selling new and used golf clubs. 00:04:49.360 --> 00:04:51.180 align:middle line:84% And it looks like there's just a zillion 00:04:51.180 --> 00:04:53.260 align:middle line:84% different golf clubs out there on eBay. 00:04:53.260 --> 00:04:56.160 align:middle line:84% But if you do a more specific search. 00:04:56.160 --> 00:05:00.420 align:middle line:84% So for this specific search, they typed in "TaylorMade burner 00:05:00.420 --> 00:05:05.700 align:middle line:84% 09 driver 2009 golf club new quote 10RH." 00:05:05.700 --> 00:05:07.890 align:middle line:84% And so you do this very specific search 00:05:07.890 --> 00:05:09.240 align:middle line:90% and then what do you find? 00:05:09.240 --> 00:05:11.520 align:middle line:84% You find that there is one seller who 00:05:11.520 --> 00:05:15.180 align:middle line:84% uses that as his title when he sells TaylorMade drivers online. 00:05:15.180 --> 00:05:18.000 align:middle line:84% And so all the searches say "TaylorMade burner 09 driver 00:05:18.000 --> 00:05:21.990 align:middle line:84% 2009 golf club new 10.5 degrees right hand." 00:05:21.990 --> 00:05:24.460 align:middle line:84% And what you'll notice in this thing 00:05:24.460 --> 00:05:27.080 align:middle line:84% is that the seller has these many, many auctions. 00:05:27.080 --> 00:05:29.600 align:middle line:84% This one is closing in 3 hours and 21 minutes. 00:05:29.600 --> 00:05:31.850 align:middle line:84% This one closes in 3 hours and 22 minutes. 00:05:31.850 --> 00:05:34.070 align:middle line:84% This one closes in 6 hours and 9 minutes. 00:05:34.070 --> 00:05:35.325 align:middle line:90% 6 hours and 9 minutes. 00:05:35.325 --> 00:05:35.950 align:middle line:90% 8 hours and 59. 00:05:35.950 --> 00:05:38.380 align:middle line:90% 9, 41. 00:05:38.380 --> 00:05:41.020 align:middle line:84% So this seller is selling auctioning off eight 00:05:41.020 --> 00:05:43.010 align:middle line:90% of these in the next 12 hours. 00:05:43.010 --> 00:05:46.030 align:middle line:84% And you can see this seller is conducting an experiment. 00:05:46.030 --> 00:05:49.070 align:middle line:84% This golf club here-- maybe you can't see it, I can. 00:05:49.070 --> 00:05:51.020 align:middle line:90% This one has $7.99 shipping. 00:05:51.020 --> 00:05:53.410 align:middle line:90% This one has $9.99 shipping. 00:05:53.410 --> 00:05:55.190 align:middle line:90% This one has $9.99 shipping. 00:05:55.190 --> 00:05:56.750 align:middle line:90% This one has $7.99 shipping. 00:05:56.750 --> 00:06:00.230 align:middle line:84% So here it's clear these have Buy It Now prices. 00:06:00.230 --> 00:06:02.740 align:middle line:90% This is just $125 plus $9.99. 00:06:02.740 --> 00:06:05.240 align:middle line:90% This $125 plus $7.99. 00:06:05.240 --> 00:06:07.180 align:middle line:84% If anyone saw these two listings, 00:06:07.180 --> 00:06:09.380 align:middle line:84% obviously no one would buy this one. 00:06:09.380 --> 00:06:12.290 align:middle line:84% But they're just mixed in with that big list. 00:06:12.290 --> 00:06:14.240 align:middle line:84% People may only see one versus the other. 00:06:14.240 --> 00:06:18.160 align:middle line:84% It's unclear what's going to happen. 00:06:18.160 --> 00:06:21.170 align:middle line:84% Anyway, so this is one seller's experiment. 00:06:21.170 --> 00:06:23.650 align:middle line:84% And their idea is that there are many, many, 00:06:23.650 --> 00:06:25.160 align:middle line:84% many of these experiments out there. 00:06:25.160 --> 00:06:32.475 align:middle line:84% But this is one of the 244,000 they're analyzing. 00:06:32.475 --> 00:06:34.850 align:middle line:84% AUDIENCE: What's the logic from the seller's perspective? 00:06:34.850 --> 00:06:36.880 align:middle line:90% Like, where does it come from? 00:06:36.880 --> 00:06:40.510 align:middle line:84% GLENN ELLISON: [INAUDIBLE] well, so I guess the seller probably 00:06:40.510 --> 00:06:44.200 align:middle line:84% doesn't know whether a $7.99 shipping or a $9.99 shipping 00:06:44.200 --> 00:06:46.420 align:middle line:90% produces better profits. 00:06:46.420 --> 00:06:48.698 align:middle line:84% So they're probably just auctioning both of them off, 00:06:48.698 --> 00:06:50.740 align:middle line:84% and they may have been doing this for a few weeks 00:06:50.740 --> 00:06:54.850 align:middle line:84% and then stopped and settled on whichever one was better. 00:06:54.850 --> 00:06:57.860 align:middle line:84% Certainly this Buy It Now versus the auction, 00:06:57.860 --> 00:07:00.770 align:middle line:84% they're probably some people who prefer one to the other. 00:07:00.770 --> 00:07:02.520 align:middle line:84% And so you probably sell both ways. 00:07:02.520 --> 00:07:07.070 align:middle line:90% 00:07:07.070 --> 00:07:09.290 align:middle line:84% But, you know, something you see in their data 00:07:09.290 --> 00:07:12.365 align:middle line:84% is you see people conducting auctions 00:07:12.365 --> 00:07:14.990 align:middle line:84% where the starting price in the auction is above the Buy It Now 00:07:14.990 --> 00:07:16.400 align:middle line:90% price. 00:07:16.400 --> 00:07:19.400 align:middle line:84% So it does seem like there are people who just see an auction 00:07:19.400 --> 00:07:21.920 align:middle line:84% and decide to enter it and other people who just are looking 00:07:21.920 --> 00:07:24.470 align:middle line:84% for a fixed price and buy it and not really paying attention 00:07:24.470 --> 00:07:25.450 align:middle line:90% to the opposite ones. 00:07:25.450 --> 00:07:33.570 align:middle line:90% 00:07:33.570 --> 00:07:37.560 align:middle line:84% OK, anyway, so how do we think about effects of reserve prices 00:07:37.560 --> 00:07:39.630 align:middle line:90% in IPV auctions? 00:07:39.630 --> 00:07:44.568 align:middle line:84% If you remember basic theory of reserve 00:07:44.568 --> 00:07:46.360 align:middle line:84% prices in an English auction is the reserve 00:07:46.360 --> 00:07:49.430 align:middle line:84% price in English auction is the monopoly price to one bidder. 00:07:49.430 --> 00:07:52.410 align:middle line:84% And the logic was just, you're selling this good, 00:07:52.410 --> 00:07:54.160 align:middle line:84% once you found out what the second highest 00:07:54.160 --> 00:07:58.240 align:middle line:84% value is, the willingness to pay of the high value bidder 00:07:58.240 --> 00:08:00.070 align:middle line:84% is just the distribution truncated 00:08:00.070 --> 00:08:02.630 align:middle line:84% at the point at which you know their value is at least that. 00:08:02.630 --> 00:08:05.620 align:middle line:84% And so you're just monopoly pricing on the single bidder 00:08:05.620 --> 00:08:07.210 align:middle line:90% distribution. 00:08:07.210 --> 00:08:09.872 align:middle line:84% And so it's just the optimal reserve price 00:08:09.872 --> 00:08:11.080 align:middle line:90% is the monopoly price, right? 00:08:11.080 --> 00:08:16.460 align:middle line:84% You have this distribution of values. 00:08:16.460 --> 00:08:19.060 align:middle line:84% Once everyone else has dropped out here, 00:08:19.060 --> 00:08:21.460 align:middle line:84% all you know is that you're selling 00:08:21.460 --> 00:08:27.420 align:middle line:84% to someone whose draw is from this is the second highest bid. 00:08:27.420 --> 00:08:30.380 align:middle line:90% 00:08:30.380 --> 00:08:31.809 align:middle line:90% Nope. 00:08:31.809 --> 00:08:32.309 align:middle line:90% OK. 00:08:32.309 --> 00:08:34.190 align:middle line:84% Anyway, so you have this distribution 00:08:34.190 --> 00:08:35.539 align:middle line:90% of the second highest bid. 00:08:35.539 --> 00:08:36.559 align:middle line:90% You're pricing to that. 00:08:36.559 --> 00:08:39.860 align:middle line:84% You set the monopoly price on that demand curve. 00:08:39.860 --> 00:08:42.710 align:middle line:84% They also note that if we want to think about this empirically, 00:08:42.710 --> 00:08:46.010 align:middle line:84% how does a reserve price effect-- 00:08:46.010 --> 00:08:47.670 align:middle line:84% what's the optimal reserve price? 00:08:47.670 --> 00:08:49.760 align:middle line:84% Here's another good way to think about-- 00:08:49.760 --> 00:08:53.660 align:middle line:84% so this is density F of v. If you want to think about 00:08:53.660 --> 00:08:56.180 align:middle line:84% how does a reserve price affect things, what I can do 00:08:56.180 --> 00:08:57.890 align:middle line:84% is just imagine an experiment where 00:08:57.890 --> 00:09:01.040 align:middle line:84% the seller has used many, many different reserve prices. 00:09:01.040 --> 00:09:03.140 align:middle line:84% Like, they've started the bidding 00:09:03.140 --> 00:09:06.680 align:middle line:90% at $1, $5, $10, 20, 30, 40, 50. 00:09:06.680 --> 00:09:09.800 align:middle line:84% For every reserve price, you can just 00:09:09.800 --> 00:09:14.090 align:middle line:84% draw a graph, which is, if I use a reserve 00:09:14.090 --> 00:09:20.600 align:middle line:84% price of 5% of the expected sale price, 00:09:20.600 --> 00:09:25.010 align:middle line:84% what's the eventual price at which I sell the good, 00:09:25.010 --> 00:09:29.310 align:middle line:90% and what's the-- 00:09:29.310 --> 00:09:31.500 align:middle line:84% if I use a very low reserve price, 00:09:31.500 --> 00:09:35.760 align:middle line:84% I sell the good with probability 0.9 or 0.95, 00:09:35.760 --> 00:09:38.440 align:middle line:84% but I get a fairly low average price. 00:09:38.440 --> 00:09:41.550 align:middle line:84% And then if I sell it with a $20 reserve, 00:09:41.550 --> 00:09:44.550 align:middle line:84% maybe I sell the good with probability 0.8 00:09:44.550 --> 00:09:46.600 align:middle line:84% and I get a higher average price. 00:09:46.600 --> 00:09:48.720 align:middle line:84% And if I sell it with a reserve that's 00:09:48.720 --> 00:09:51.820 align:middle line:84% 120% of the average price for the good, 00:09:51.820 --> 00:09:54.450 align:middle line:84% maybe I sell it with just some very small probability, 00:09:54.450 --> 00:09:56.623 align:middle line:84% but I do get a very high price of selling. 00:09:56.623 --> 00:09:59.040 align:middle line:84% So what you can basically do is just look at every reserve 00:09:59.040 --> 00:10:01.560 align:middle line:84% price in the data and figure out what's 00:10:01.560 --> 00:10:03.540 align:middle line:84% the probability that the good sells 00:10:03.540 --> 00:10:06.820 align:middle line:84% and what's the expected price conditional on its selling, 00:10:06.820 --> 00:10:09.972 align:middle line:84% and that's basically just your demand curve. 00:10:09.972 --> 00:10:11.430 align:middle line:84% You can view that as a demand curve 00:10:11.430 --> 00:10:13.930 align:middle line:84% and you're just trying to monopoly price on that curve. 00:10:13.930 --> 00:10:18.660 align:middle line:84% So that's a second way to think about optimal reserve pricing is 00:10:18.660 --> 00:10:21.360 align:middle line:84% just draw this demand curve like object 00:10:21.360 --> 00:10:24.417 align:middle line:84% and think about what would be the monopoly price on that. 00:10:24.417 --> 00:10:26.250 align:middle line:84% And that would be the way to figure out what 00:10:26.250 --> 00:10:28.420 align:middle line:90% the optimal reserve price is. 00:10:28.420 --> 00:10:30.000 align:middle line:84% So that would be an experimental way 00:10:30.000 --> 00:10:32.910 align:middle line:84% to figure out my optimal reserve price without, you know-- 00:10:32.910 --> 00:10:34.420 align:middle line:84% theoretically you want to just say, 00:10:34.420 --> 00:10:35.970 align:middle line:84% what's the distribution from which the values are 00:10:35.970 --> 00:10:36.630 align:middle line:90% being drawn. 00:10:36.630 --> 00:10:38.490 align:middle line:84% But without thinking about what the distribution from which 00:10:38.490 --> 00:10:41.130 align:middle line:84% values are being drawn, you could just make this curve 00:10:41.130 --> 00:10:44.880 align:middle line:90% monopoly price relative to it. 00:10:44.880 --> 00:10:46.161 align:middle line:90% OK. 00:10:46.161 --> 00:10:49.050 align:middle line:84% You know, several reasons why this simple theory may not 00:10:49.050 --> 00:10:50.760 align:middle line:90% work in practice. 00:10:50.760 --> 00:10:52.540 align:middle line:90% One is entry costs. 00:10:52.540 --> 00:10:54.540 align:middle line:84% If you imagine there's some cost to a bidder 00:10:54.540 --> 00:10:58.740 align:middle line:84% of starting to bid in an auction and examining 00:10:58.740 --> 00:10:59.940 align:middle line:90% the good in detail. 00:10:59.940 --> 00:11:01.710 align:middle line:84% If there's a low price, they may be 00:11:01.710 --> 00:11:03.190 align:middle line:90% willing to pay the entry cost. 00:11:03.190 --> 00:11:05.340 align:middle line:84% If they see that the reserve price is $110, 00:11:05.340 --> 00:11:07.320 align:middle line:84% they may not bother to read all the fine print 00:11:07.320 --> 00:11:10.320 align:middle line:90% and enter that auction. 00:11:10.320 --> 00:11:12.700 align:middle line:84% There could be some kind of common value component. 00:11:12.700 --> 00:11:16.200 align:middle line:84% This theory was, once I know everybody else's bid, 00:11:16.200 --> 00:11:17.760 align:middle line:90% my bid is unchanged. 00:11:17.760 --> 00:11:19.720 align:middle line:84% There's common values, the theory changes. 00:11:19.720 --> 00:11:22.470 align:middle line:84% And then they also note that there could be-- 00:11:22.470 --> 00:11:26.130 align:middle line:84% speaking of this theory and why this theory may not work. 00:11:26.130 --> 00:11:29.590 align:middle line:84% You could also have behavioral preferences, like loss aversion. 00:11:29.590 --> 00:11:33.270 align:middle line:84% If you imagine the loss aversion is always like, 00:11:33.270 --> 00:11:35.208 align:middle line:84% I feel bad losing something that I have. 00:11:35.208 --> 00:11:37.500 align:middle line:84% If you start the bidding at a very low price and people 00:11:37.500 --> 00:11:39.480 align:middle line:84% get used to the idea of owning this fancy golf 00:11:39.480 --> 00:11:41.585 align:middle line:84% club because they see it being bid for $20, 00:11:41.585 --> 00:11:42.960 align:middle line:84% once they've got used to the fact 00:11:42.960 --> 00:11:45.670 align:middle line:84% that they have this fancy golf club, 00:11:45.670 --> 00:11:49.293 align:middle line:84% then they may feel loss aversion if they don't win the golf club. 00:11:49.293 --> 00:11:51.460 align:middle line:84% You might, by getting people tied up in the bidding, 00:11:51.460 --> 00:11:53.252 align:middle line:84% they may get used to the thought of winning 00:11:53.252 --> 00:11:56.440 align:middle line:84% and then they may bid more in the end. 00:11:56.440 --> 00:12:01.960 align:middle line:84% Anyway, to examine this, what they do is they find a data set 00:12:01.960 --> 00:12:04.450 align:middle line:84% where they found 19,000 experiments, 00:12:04.450 --> 00:12:07.390 align:middle line:84% where there are several features. 00:12:07.390 --> 00:12:10.930 align:middle line:84% One is there's variation in the start price, 00:12:10.930 --> 00:12:13.580 align:middle line:84% there's free shipping, which just simplifies things. 00:12:13.580 --> 00:12:16.340 align:middle line:84% There's no secret reserve price, no Buy It Now auction. 00:12:16.340 --> 00:12:20.770 align:middle line:84% So these are just straight, plain vanilla auctions 00:12:20.770 --> 00:12:23.050 align:middle line:90% of a good, many different times. 00:12:23.050 --> 00:12:30.570 align:middle line:84% And so about 25 listings per experiment. 00:12:30.570 --> 00:12:35.280 align:middle line:84% So again, these are just, you know, 00:12:35.280 --> 00:12:39.910 align:middle line:84% sellers experimenting with, if I'm selling $125 golf club, 00:12:39.910 --> 00:12:43.135 align:middle line:84% should I start the bidding at $5, or $100, or 50, 00:12:43.135 --> 00:12:44.135 align:middle line:90% or something in between? 00:12:44.135 --> 00:12:46.640 align:middle line:90% 00:12:46.640 --> 00:12:48.500 align:middle line:84% Figure 5 of the paper shows that, yes, 00:12:48.500 --> 00:12:50.640 align:middle line:90% sellers are doing this a lot. 00:12:50.640 --> 00:12:53.000 align:middle line:84% So let's look at a typical category like goods 00:12:53.000 --> 00:12:55.610 align:middle line:90% that cost between $30 and $100. 00:12:55.610 --> 00:12:58.572 align:middle line:84% And you can see goods where that cost, 00:12:58.572 --> 00:13:00.530 align:middle line:84% this is like the average price at which they're 00:13:00.530 --> 00:13:04.160 align:middle line:84% sold when they sell them through just a fixed price sale. 00:13:04.160 --> 00:13:07.190 align:middle line:84% And you can see goods that sell for $30 to $100, 00:13:07.190 --> 00:13:11.990 align:middle line:84% 20% of the sellers start them out at 0 in these experiments 00:13:11.990 --> 00:13:17.390 align:middle line:84% and then 37% start them out at between 85 to 100% 00:13:17.390 --> 00:13:19.520 align:middle line:90% of the average fixed price sale. 00:13:19.520 --> 00:13:21.870 align:middle line:90% 17% choose 60 to 80%. 00:13:21.870 --> 00:13:25.490 align:middle line:84% 13% are actually starting them off above the average 00:13:25.490 --> 00:13:28.080 align:middle line:84% Buy It Now average fixed price sales. 00:13:28.080 --> 00:13:30.780 align:middle line:84% But we get this sort of bimodal distribution. 00:13:30.780 --> 00:13:33.513 align:middle line:84% Notably, there are very few sellers starting out in here. 00:13:33.513 --> 00:13:35.180 align:middle line:84% It seems like if you're going to go low, 00:13:35.180 --> 00:13:37.100 align:middle line:84% you just go all the way down to the bottom 00:13:37.100 --> 00:13:40.620 align:middle line:90% and start at almost 0. 00:13:40.620 --> 00:13:44.380 align:middle line:84% And then there's with-- a lot of within experiment variation. 00:13:44.380 --> 00:13:48.270 align:middle line:84% So if you look at auctions in which their sellers are 00:13:48.270 --> 00:13:50.950 align:middle line:84% sometimes selling it at less than 5%, 00:13:50.950 --> 00:13:52.440 align:middle line:84% those same sellers were sometimes 00:13:52.440 --> 00:13:56.280 align:middle line:84% selling it at less than 5% are also sometimes selling 5 to 45%, 00:13:56.280 --> 00:14:01.800 align:middle line:84% 45 to 85%, 85 to 100%, 100 to 120%, greater than 120%. 00:14:01.800 --> 00:14:03.630 align:middle line:84% So there are just many experiments 00:14:03.630 --> 00:14:07.450 align:middle line:84% where sellers are trying out the whole range themselves. 00:14:07.450 --> 00:14:10.412 align:middle line:84% This is our experimental paper that we 00:14:10.412 --> 00:14:12.120 align:middle line:84% get to run without acting to actually pay 00:14:12.120 --> 00:14:13.026 align:middle line:90% for the experiments. 00:14:13.026 --> 00:14:17.040 align:middle line:90% 00:14:17.040 --> 00:14:17.540 align:middle line:90% OK. 00:14:17.540 --> 00:14:23.980 align:middle line:84% So does the standard auction demand theory work? 00:14:23.980 --> 00:14:26.870 align:middle line:90% And the answer is no. 00:14:26.870 --> 00:14:28.880 align:middle line:84% There's one interesting feature, which 00:14:28.880 --> 00:14:32.660 align:middle line:84% is the demand curve turns up at very low prices. 00:14:32.660 --> 00:14:35.060 align:middle line:84% So remember, I drew this curve for you, 00:14:35.060 --> 00:14:38.975 align:middle line:84% saying, if I start the good at only $5, 00:14:38.975 --> 00:14:40.100 align:middle line:90% if I start the good at $1-- 00:14:40.100 --> 00:14:43.280 align:middle line:90% 00:14:43.280 --> 00:14:45.140 align:middle line:84% if I start the good at $1, I sell it 00:14:45.140 --> 00:14:46.585 align:middle line:84% almost entirely for sure, but I'm 00:14:46.585 --> 00:14:48.710 align:middle line:84% going to have a lower average price than if I start 00:14:48.710 --> 00:14:50.600 align:middle line:90% the good at a higher price. 00:14:50.600 --> 00:14:54.350 align:middle line:84% What they find is that the curve doesn't look like this. 00:14:54.350 --> 00:14:56.200 align:middle line:90% The curve looks like this. 00:14:56.200 --> 00:15:01.350 align:middle line:90% 00:15:01.350 --> 00:15:05.130 align:middle line:84% So as I start at a lower and lower price, not 00:15:05.130 --> 00:15:08.400 align:middle line:84% only do I sell with a higher probability, 00:15:08.400 --> 00:15:11.970 align:middle line:84% I sell at a higher price, conditional on selling. 00:15:11.970 --> 00:15:15.018 align:middle line:84% So I'm up here, I'm selling at 90% probability 00:15:15.018 --> 00:15:16.560 align:middle line:84% and I'm getting a higher price than I 00:15:16.560 --> 00:15:18.268 align:middle line:84% would get if I chose a reserve price that 00:15:18.268 --> 00:15:20.670 align:middle line:90% only sold with 70% probability. 00:15:20.670 --> 00:15:23.700 align:middle line:84% So the demand curve turns up, suggesting 00:15:23.700 --> 00:15:28.510 align:middle line:84% that I put the $100 item out there, start the bidding at $1, 00:15:28.510 --> 00:15:30.360 align:middle line:90% lots of people see it at $1. 00:15:30.360 --> 00:15:32.040 align:middle line:84% 10 hours before the auction is over, 00:15:32.040 --> 00:15:33.970 align:middle line:84% they all start following the auction. 00:15:33.970 --> 00:15:35.640 align:middle line:90% They all start bidding. 00:15:35.640 --> 00:15:40.840 align:middle line:84% I get more entry than I would if I started at it at $20 instead. 00:15:40.840 --> 00:15:43.650 align:middle line:84% So the curve, instead of looking like this, 00:15:43.650 --> 00:15:44.930 align:middle line:90% looks like this instead. 00:15:44.930 --> 00:15:50.940 align:middle line:90% 00:15:50.940 --> 00:15:52.860 align:middle line:84% Obviously, how do we do monopoly pricing? 00:15:52.860 --> 00:15:54.930 align:middle line:84% Like, the way to think about monopoly pricing 00:15:54.930 --> 00:15:57.220 align:middle line:84% is you do marginal revenue equals marginal cost. 00:15:57.220 --> 00:15:59.580 align:middle line:84% So they take this demand curve, they 00:15:59.580 --> 00:16:01.830 align:middle line:84% construct the marginal revenue curve, 00:16:01.830 --> 00:16:07.260 align:middle line:84% and the marginal revenue curve also turns up. 00:16:07.260 --> 00:16:10.530 align:middle line:84% Like, we're used to marginal revenue is downward sloping. 00:16:10.530 --> 00:16:12.480 align:middle line:90% You draw marginal cost across. 00:16:12.480 --> 00:16:14.530 align:middle line:84% Where marginal revenue hits marginal cost, 00:16:14.530 --> 00:16:17.620 align:middle line:90% that's the monopoly price. 00:16:17.620 --> 00:16:21.090 align:middle line:84% What do you do when marginal revenue turns up? 00:16:21.090 --> 00:16:26.760 align:middle line:84% So if this is my marginal revenue curve, 00:16:26.760 --> 00:16:28.290 align:middle line:90% my marginal cost is flat. 00:16:28.290 --> 00:16:30.870 align:middle line:84% That's my normal monopoly pricing. 00:16:30.870 --> 00:16:34.650 align:middle line:84% What do I do if marginal revenue turns up? 00:16:34.650 --> 00:16:39.060 align:middle line:84% Then I can draw the line across where marginal revenue equals 00:16:39.060 --> 00:16:40.620 align:middle line:90% marginal cost. 00:16:40.620 --> 00:16:44.800 align:middle line:84% This is a locally optimal monopoly price. 00:16:44.800 --> 00:16:49.640 align:middle line:84% But the solution is either going to be this point or this point. 00:16:49.640 --> 00:16:51.600 align:middle line:84% You know, it doesn't have to intersect again. 00:16:51.600 --> 00:16:56.010 align:middle line:84% It could be that you get a, here your revenue's increasing, 00:16:56.010 --> 00:16:58.640 align:middle line:84% so it may be that you want to price there. 00:16:58.640 --> 00:17:00.560 align:middle line:90% You may want to price here. 00:17:00.560 --> 00:17:02.300 align:middle line:84% That may be the global optimum or this 00:17:02.300 --> 00:17:03.410 align:middle line:84% could be the global optimum, but you're 00:17:03.410 --> 00:17:04.618 align:middle line:90% going to do one or the other. 00:17:04.618 --> 00:17:08.089 align:middle line:90% 00:17:08.089 --> 00:17:08.630 align:middle line:90% Anyway. 00:17:08.630 --> 00:17:10.669 align:middle line:84% And so that's, I guess, in some sense the-- 00:17:10.669 --> 00:17:13.900 align:middle line:90% 00:17:13.900 --> 00:17:17.150 align:middle line:84% this is how they explain what's going on in the data. 00:17:17.150 --> 00:17:19.930 align:middle line:84% If you remember in the data, there was a lot of firms-- 00:17:19.930 --> 00:17:20.920 align:middle line:90% figure 5. 00:17:20.920 --> 00:17:25.630 align:middle line:84% Like, there were a lot of firms charging fairly high reserve 00:17:25.630 --> 00:17:28.030 align:middle line:84% prices and then some firms charging very low reserve 00:17:28.030 --> 00:17:29.090 align:middle line:90% prices. 00:17:29.090 --> 00:17:31.690 align:middle line:84% And so their thought is that what's going on in the data 00:17:31.690 --> 00:17:34.840 align:middle line:84% is probably something like this, where the marginal cost is 00:17:34.840 --> 00:17:35.960 align:middle line:90% pretty high. 00:17:35.960 --> 00:17:40.210 align:middle line:84% So you either set a quite high reserve price and this 00:17:40.210 --> 00:17:43.100 align:middle line:84% is the optimum or you set this right end point. 00:17:43.100 --> 00:17:45.360 align:middle line:84% And for different sellers, different ones are optimal. 00:17:45.360 --> 00:17:50.480 align:middle line:90% 00:17:50.480 --> 00:17:52.400 align:middle line:84% And in general, if your marginal cost is high, 00:17:52.400 --> 00:17:54.400 align:middle line:84% you're going to just set the high reserve price. 00:17:54.400 --> 00:17:56.093 align:middle line:84% Your marginal cost is low, you're 00:17:56.093 --> 00:17:57.760 align:middle line:84% going to be setting the right end point. 00:17:57.760 --> 00:18:00.152 align:middle line:90% 00:18:00.152 --> 00:18:02.610 align:middle line:84% You know, and then the paper tries to do other predictions. 00:18:02.610 --> 00:18:07.430 align:middle line:84% It tries to have a test of upper tail hypothesis of, 00:18:07.430 --> 00:18:11.150 align:middle line:84% does the upper tail of the price distribution 00:18:11.150 --> 00:18:14.850 align:middle line:84% look the same regardless of what reserve price you use? 00:18:14.850 --> 00:18:17.220 align:middle line:84% And that test only works so well. 00:18:17.220 --> 00:18:21.020 align:middle line:84% So it does seem like the classic theory, 00:18:21.020 --> 00:18:24.140 align:middle line:84% this sort of upper tail distribution 00:18:24.140 --> 00:18:25.610 align:middle line:84% is just what you get by truncating 00:18:25.610 --> 00:18:27.470 align:middle line:90% isn't quite right here. 00:18:27.470 --> 00:18:29.480 align:middle line:90% And it does seem to enter-- 00:18:29.480 --> 00:18:32.480 align:middle line:84% it's entry costs are important for online auctions, 00:18:32.480 --> 00:18:36.730 align:middle line:84% I think is the main message there. 00:18:36.730 --> 00:18:38.090 align:middle line:90% OK, questions? 00:18:38.090 --> 00:18:43.010 align:middle line:90% 00:18:43.010 --> 00:18:43.510 align:middle line:90% Yeah. 00:18:43.510 --> 00:18:48.150 align:middle line:84% And obviously, I think that's something, you know-- 00:18:48.150 --> 00:18:50.850 align:middle line:90% that work was done inside eBay. 00:18:50.850 --> 00:18:52.620 align:middle line:84% For a long time, eBay actually had 00:18:52.620 --> 00:18:56.598 align:middle line:84% this sort of very, very open policy to academics. 00:18:56.598 --> 00:18:58.140 align:middle line:84% And we actually had several students, 00:18:58.140 --> 00:18:59.940 align:middle line:84% like, after they finished their PhDs here, 00:18:59.940 --> 00:19:02.760 align:middle line:84% go to a postdoc for a year and stay at eBay because eBay would 00:19:02.760 --> 00:19:06.120 align:middle line:84% let you sit there and operate inside the company 00:19:06.120 --> 00:19:08.670 align:middle line:84% and do all kinds of stuff with their full data set. 00:19:08.670 --> 00:19:12.820 align:middle line:84% That seems to have faded away three or four years ago. 00:19:12.820 --> 00:19:14.820 align:middle line:84% But obviously one could still scrape eBay and do 00:19:14.820 --> 00:19:17.430 align:middle line:90% the same kinds of things. 00:19:17.430 --> 00:19:21.732 align:middle line:84% OK, so then the second topic that I guess 00:19:21.732 --> 00:19:23.190 align:middle line:84% is common to the rest of the papers 00:19:23.190 --> 00:19:26.370 align:middle line:84% I'm going to talk about is just understanding 00:19:26.370 --> 00:19:30.990 align:middle line:84% how online retail works and why it's successful. 00:19:30.990 --> 00:19:34.980 align:middle line:84% You know, if you read people writing about E-commerce 00:19:34.980 --> 00:19:39.060 align:middle line:84% early in the internet era, there was this question of, 00:19:39.060 --> 00:19:42.750 align:middle line:84% is online retail ever going to really take off and displace 00:19:42.750 --> 00:19:44.410 align:middle line:90% traditional retail? 00:19:44.410 --> 00:19:47.520 align:middle line:84% And the biggest argument against online retail taking off 00:19:47.520 --> 00:19:51.030 align:middle line:84% was that it's just much more expensive. 00:19:51.030 --> 00:19:54.660 align:middle line:84% You know, like, if you look at the long history of retail 00:19:54.660 --> 00:19:59.460 align:middle line:84% in America, you know, in the 1800s, 00:19:59.460 --> 00:20:02.010 align:middle line:84% people shopped in department stores. 00:20:02.010 --> 00:20:06.270 align:middle line:84% And then there was this golden era of catalog retail 00:20:06.270 --> 00:20:09.690 align:middle line:84% where, like the Sears catalog in some sense 00:20:09.690 --> 00:20:12.780 align:middle line:84% brought products to-- you know, they were only formally 00:20:12.780 --> 00:20:14.770 align:middle line:84% available in cities to all of America. 00:20:14.770 --> 00:20:16.890 align:middle line:84% And they had like this phone book, you know, 00:20:16.890 --> 00:20:20.260 align:middle line:84% 1,000-page thick catalog that they would mail you. 00:20:20.260 --> 00:20:22.830 align:middle line:84% You could buy anything from the Sears catalog. 00:20:22.830 --> 00:20:26.550 align:middle line:84% You know, the Sears catalog just took off in the, I guess 00:20:26.550 --> 00:20:29.790 align:middle line:84% it's late 1800s after there were trains, or whatever, 00:20:29.790 --> 00:20:32.770 align:middle line:84% that would bring packages to people. 00:20:32.770 --> 00:20:35.890 align:middle line:84% When the car was invented, the Sears catalog just 00:20:35.890 --> 00:20:37.420 align:middle line:90% went into a prolonged decline. 00:20:37.420 --> 00:20:41.290 align:middle line:84% And Sears opened, you know, massive, you know, not the size 00:20:41.290 --> 00:20:44.350 align:middle line:84% of a Walmart, but, you know, massive department stores 00:20:44.350 --> 00:20:45.650 align:middle line:90% in many US cities. 00:20:45.650 --> 00:20:47.840 align:middle line:84% And again, the catalog business declined. 00:20:47.840 --> 00:20:51.940 align:middle line:84% The physical retail business took off. 00:20:51.940 --> 00:20:54.190 align:middle line:90% Thoughts on why that was. 00:20:54.190 --> 00:20:57.310 align:middle line:84% Catalog retail was just always much more expensive 00:20:57.310 --> 00:20:58.770 align:middle line:90% than traditional retail. 00:20:58.770 --> 00:21:04.820 align:middle line:84% Like, if you look at Walmart, Walmart's cost of selling goods 00:21:04.820 --> 00:21:11.030 align:middle line:84% is roughly 20% of the merchandise that they sell. 00:21:11.030 --> 00:21:15.650 align:middle line:84% Catalog retail had costs more like 40 or 50%. 00:21:15.650 --> 00:21:19.190 align:middle line:84% And why does it cost Walmart only 20% 00:21:19.190 --> 00:21:22.260 align:middle line:84% when it costs catalog retail 50%? 00:21:22.260 --> 00:21:25.900 align:middle line:84% You know, it's because catalog retail, 00:21:25.900 --> 00:21:27.760 align:middle line:84% you need to have stuff in a warehouse. 00:21:27.760 --> 00:21:29.580 align:middle line:84% Every single item that gets sold, 00:21:29.580 --> 00:21:32.580 align:middle line:84% someone needs to put it in a box, put the box in the mail, 00:21:32.580 --> 00:21:33.890 align:middle line:90% pay for the mail. 00:21:33.890 --> 00:21:37.440 align:middle line:84% Whereas Walmart's cost of putting-- 00:21:37.440 --> 00:21:39.220 align:middle line:84% like, the simplest example of things 00:21:39.220 --> 00:21:40.387 align:middle line:90% that should never be online. 00:21:40.387 --> 00:21:41.880 align:middle line:84% Like, a kiddie swimming pool, one 00:21:41.880 --> 00:21:44.520 align:middle line:84% of those plastic four-foot diameter swimming pools. 00:21:44.520 --> 00:21:47.580 align:middle line:84% You can put a whole bunch of four-foot diameter plastic 00:21:47.580 --> 00:21:50.780 align:middle line:84% swimming pools, stack them together, put them in a truck, 00:21:50.780 --> 00:21:53.280 align:middle line:84% take them to the Walmart, and the cost of getting them there 00:21:53.280 --> 00:21:55.780 align:middle line:84% is, you know, $0.50 or something like that. 00:21:55.780 --> 00:21:58.170 align:middle line:84% You try to put a kiddie swimming pool in the mail 00:21:58.170 --> 00:22:00.120 align:middle line:84% and send it to somebody, you've spent, 00:22:00.120 --> 00:22:03.210 align:middle line:84% you know, $30 mailing something that's only worth 00:22:03.210 --> 00:22:04.660 align:middle line:90% $5 or something. 00:22:04.660 --> 00:22:06.360 align:middle line:84% And, you know, putting things in a truck 00:22:06.360 --> 00:22:10.320 align:middle line:84% is just much, much cheaper than handling every package 00:22:10.320 --> 00:22:10.930 align:middle line:90% individually. 00:22:10.930 --> 00:22:13.020 align:middle line:84% And you're getting some sense the free labor of people 00:22:13.020 --> 00:22:15.353 align:middle line:84% to just walk through the store and take it off the shelf 00:22:15.353 --> 00:22:18.130 align:middle line:90% and carry it home. 00:22:18.130 --> 00:22:23.267 align:middle line:84% Anyway, early on, Amazon was undercutting Barnes & Noble 00:22:23.267 --> 00:22:25.600 align:middle line:84% and selling books for lower than you would buy the books 00:22:25.600 --> 00:22:27.060 align:middle line:90% in a physical bookstore. 00:22:27.060 --> 00:22:29.500 align:middle line:84% And this question of, could Amazon ever 00:22:29.500 --> 00:22:32.960 align:middle line:84% actually make a profit selling things not at a loss? 00:22:32.960 --> 00:22:36.430 align:middle line:84% And if Amazon were pricing things at cost or above cost, 00:22:36.430 --> 00:22:39.880 align:middle line:84% could it be competitive with traditional retail? 00:22:39.880 --> 00:22:43.450 align:middle line:84% One of the big arguments that was made early on 00:22:43.450 --> 00:22:46.030 align:middle line:84% was that, yes, Amazon doesn't need 00:22:46.030 --> 00:22:48.770 align:middle line:84% to have a cost advantage to sell things online. 00:22:48.770 --> 00:22:51.310 align:middle line:84% It could be that you have a cost disadvantage that people 00:22:51.310 --> 00:22:54.100 align:middle line:84% buy online because of the product variety. 00:22:54.100 --> 00:22:57.010 align:middle line:84% The product variety advantages are much bigger 00:22:57.010 --> 00:22:59.030 align:middle line:90% than the cost disadvantage. 00:22:59.030 --> 00:23:02.057 align:middle line:84% So therefore people buy online to get the book they want 00:23:02.057 --> 00:23:03.640 align:middle line:84% or whatever product they want and have 00:23:03.640 --> 00:23:06.070 align:middle line:90% a much bigger selection. 00:23:06.070 --> 00:23:08.770 align:middle line:84% And they note that when Amazon launched, 00:23:08.770 --> 00:23:10.530 align:middle line:90% Amazon had one million titles. 00:23:10.530 --> 00:23:13.270 align:middle line:84% Obviously, it later expanded to several million titles. 00:23:13.270 --> 00:23:17.060 align:middle line:84% And a very large physical bookstore at the time would have 00:23:17.060 --> 00:23:18.470 align:middle line:90% 100,000 titles. 00:23:18.470 --> 00:23:22.820 align:middle line:84% You're typical bookstore might have 50,000 titles in it. 00:23:22.820 --> 00:23:26.490 align:middle line:84% Perhaps there's some superstores that would have 200,000. 00:23:26.490 --> 00:23:30.350 align:middle line:84% But Amazon had many, many more books, 00:23:30.350 --> 00:23:35.395 align:middle line:84% even from the very beginning, than physical bookstores did. 00:23:35.395 --> 00:23:37.520 align:middle line:84% And so what Brynjolfsson, Hu, and Smith tried to do 00:23:37.520 --> 00:23:41.270 align:middle line:84% is just say, could we estimate how big the product variety 00:23:41.270 --> 00:23:45.480 align:middle line:84% benefit is by just, in some sense, looking at this demand 00:23:45.480 --> 00:23:45.980 align:middle line:90% curve? 00:23:45.980 --> 00:23:53.630 align:middle line:84% Like, finding out what fraction of the sales of Amazon are 00:23:53.630 --> 00:23:58.178 align:middle line:84% to books that are beyond the 100,000th book? 00:23:58.178 --> 00:24:00.470 align:middle line:84% So let's say there are 100,000 books that are available 00:24:00.470 --> 00:24:01.980 align:middle line:90% in your physical bookstore. 00:24:01.980 --> 00:24:03.410 align:middle line:84% There are millions of other books 00:24:03.410 --> 00:24:05.500 align:middle line:84% that are not available in a physical bookstore. 00:24:05.500 --> 00:24:09.350 align:middle line:84% How much utility do people get from buying all those books from 00:24:09.350 --> 00:24:12.320 align:middle line:90% 100,000 out to 2 million? 00:24:12.320 --> 00:24:18.270 align:middle line:84% And the way they do that is to just basically convert Amazon 00:24:18.270 --> 00:24:21.870 align:middle line:84% sales ranks to quantities, figure out how many copies 00:24:21.870 --> 00:24:25.260 align:middle line:84% of every book are being sold, and then just count up how many 00:24:25.260 --> 00:24:28.150 align:middle line:84% books-- what fraction of Amazon sales are books outside the top 00:24:28.150 --> 00:24:31.800 align:middle line:84% 100,000 and then multiply them by just a consumer surplus 00:24:31.800 --> 00:24:33.090 align:middle line:90% estimate per title. 00:24:33.090 --> 00:24:36.270 align:middle line:84% If you have a general sense of the demand elasticity, 00:24:36.270 --> 00:24:38.490 align:middle line:84% you know the average price, you know the consumer, 00:24:38.490 --> 00:24:39.990 align:middle line:84% you can just guess what the consumer 00:24:39.990 --> 00:24:43.850 align:middle line:90% surplus is and estimate that. 00:24:43.850 --> 00:24:47.360 align:middle line:84% And what they show in this figure is that if you look 00:24:47.360 --> 00:24:50.150 align:middle line:84% at books that are outside the top 100,000, 00:24:50.150 --> 00:24:52.910 align:middle line:84% there seems to be this very, very thick or very, 00:24:52.910 --> 00:24:55.030 align:middle line:90% very slow decline in sales. 00:24:55.030 --> 00:24:59.630 align:middle line:84% So book number 100,000 is selling 1 and 1/2 copies a week. 00:24:59.630 --> 00:25:02.100 align:middle line:84% Book number 200,000 is selling 0.8. 00:25:02.100 --> 00:25:05.400 align:middle line:84% Book number 400 is selling 0.6 copies a week. 00:25:05.400 --> 00:25:08.460 align:middle line:84% Book number 6 is selling like 0.4. 00:25:08.460 --> 00:25:10.650 align:middle line:84% But then when you get to book 2.2 million, 00:25:10.650 --> 00:25:13.580 align:middle line:84% book 2.2 million is not selling that many fewer copies than book 00:25:13.580 --> 00:25:15.000 align:middle line:90% 600,000. 00:25:15.000 --> 00:25:20.390 align:middle line:84% And so because this tail doesn't just go to 0 very quickly, 00:25:20.390 --> 00:25:22.430 align:middle line:84% there are lots of sales on Amazon 00:25:22.430 --> 00:25:24.620 align:middle line:90% that are these extra books. 00:25:24.620 --> 00:25:26.990 align:middle line:84% And the thought is that all these extra books 00:25:26.990 --> 00:25:31.190 align:middle line:84% are contributing a lot to consumer surplus. 00:25:31.190 --> 00:25:32.960 align:middle line:84% What they estimate in their paper 00:25:32.960 --> 00:25:36.590 align:middle line:84% is that the product variety back in 2000, when Amazon was still 00:25:36.590 --> 00:25:40.760 align:middle line:84% young, was adding $700 million to a billion 00:25:40.760 --> 00:25:44.480 align:middle line:84% to consumer surplus, and that was about-- 00:25:44.480 --> 00:25:46.820 align:middle line:84% Amazon was lowering prices, I think it was by-- 00:25:46.820 --> 00:25:50.730 align:middle line:84% Amazon was 5 to 10% cheaper than physical bookstores at the time. 00:25:50.730 --> 00:25:53.480 align:middle line:84% And what they find is that look at the gain 00:25:53.480 --> 00:25:56.400 align:middle line:84% from the lower prices, the 5 to 10% lower prices. 00:25:56.400 --> 00:26:00.290 align:middle line:84% Look at the gains from the consumer surplus. 00:26:00.290 --> 00:26:02.060 align:middle line:84% And the gains from the lower prices 00:26:02.060 --> 00:26:03.790 align:middle line:84% were only about $100 million a year. 00:26:03.790 --> 00:26:06.140 align:middle line:84% The gains from the consumer surplus is like $1 billion 00:26:06.140 --> 00:26:06.910 align:middle line:90% a year. 00:26:06.910 --> 00:26:11.310 align:middle line:84% So consumer surplus is 10 times as important as the price cut. 00:26:11.310 --> 00:26:14.330 align:middle line:84% And so if you believe that, if consumer surplus is 10 times as 00:26:14.330 --> 00:26:17.780 align:middle line:84% important as a price cut, you could take away the $100 million 00:26:17.780 --> 00:26:19.640 align:middle line:84% or even have a negative $100 million 00:26:19.640 --> 00:26:22.850 align:middle line:84% or negative $200 million, and they say that Amazon really 00:26:22.850 --> 00:26:26.767 align:middle line:84% could raise prices and still be a positive value for consumers 00:26:26.767 --> 00:26:28.225 align:middle line:84% because the product variety effect. 00:26:28.225 --> 00:26:30.760 align:middle line:90% 00:26:30.760 --> 00:26:32.980 align:middle line:90% It's a very nice paper. 00:26:32.980 --> 00:26:35.250 align:middle line:84% The calculation itself is a little speculative 00:26:35.250 --> 00:26:37.667 align:middle line:90% in that it is a-- 00:26:37.667 --> 00:26:39.000 align:middle line:90% we don't have a lot of evidence. 00:26:39.000 --> 00:26:40.380 align:middle line:84% Like, if you think about it, like, 00:26:40.380 --> 00:26:42.060 align:middle line:84% any of these books that are sold very rarely, 00:26:42.060 --> 00:26:44.010 align:middle line:84% it's very hard to get much evidence of what the demand 00:26:44.010 --> 00:26:46.010 align:middle line:84% curves look like for them because they're hardly 00:26:46.010 --> 00:26:47.160 align:middle line:90% selling any copies. 00:26:47.160 --> 00:26:51.600 align:middle line:84% You are extrapolating on some things there. 00:26:51.600 --> 00:26:54.240 align:middle line:84% So then the paper I want to talk about more 00:26:54.240 --> 00:26:59.830 align:middle line:84% is Quan and Williams' Product Variety, Market Heterogeneity. 00:26:59.830 --> 00:27:02.670 align:middle line:84% And this is an argument that the Brynjolfsson, Hu, 00:27:02.670 --> 00:27:08.040 align:middle line:84% and Smith paper may really be overestimating product variety 00:27:08.040 --> 00:27:09.220 align:middle line:90% advantages. 00:27:09.220 --> 00:27:10.950 align:middle line:84% And the reason they would argue this 00:27:10.950 --> 00:27:13.440 align:middle line:84% could be really overestimating profit variety advantages 00:27:13.440 --> 00:27:15.750 align:middle line:84% is because different physical bookstores 00:27:15.750 --> 00:27:17.410 align:middle line:90% carry different books. 00:27:17.410 --> 00:27:18.990 align:middle line:84% So, for instance, you know, people 00:27:18.990 --> 00:27:21.930 align:middle line:84% at the time needed to get textbooks 00:27:21.930 --> 00:27:24.400 align:middle line:84% for all of their graduate economics classes. 00:27:24.400 --> 00:27:27.550 align:middle line:84% Where would you get textbooks for graduate economics classes? 00:27:27.550 --> 00:27:30.700 align:middle line:84% None of them are in the top 100,000 most popular books 00:27:30.700 --> 00:27:32.060 align:middle line:90% in the United States. 00:27:32.060 --> 00:27:34.750 align:middle line:84% However, you go to the MIT Coop in the week 00:27:34.750 --> 00:27:36.250 align:middle line:84% before classes start, and they would 00:27:36.250 --> 00:27:38.380 align:middle line:84% have all of the graduate economics textbooks 00:27:38.380 --> 00:27:41.140 align:middle line:84% and all of the other textbooks for all of your other classes 00:27:41.140 --> 00:27:42.680 align:middle line:90% would be in the MIT Coop. 00:27:42.680 --> 00:27:45.820 align:middle line:84% And, you know, not just the graduate economics textbooks 00:27:45.820 --> 00:27:48.940 align:middle line:84% from MIT Coop, but the MIT Coop also 00:27:48.940 --> 00:27:52.760 align:middle line:84% has a big selection of, you know, technical books. 00:27:52.760 --> 00:27:55.600 align:middle line:84% And if you just want books about computer programming 00:27:55.600 --> 00:27:57.550 align:middle line:84% or whatever, not for your classes, 00:27:57.550 --> 00:27:59.875 align:middle line:84% they have lots of those books in that bookstores. 00:27:59.875 --> 00:28:01.750 align:middle line:84% And then there are other bookstores in Boston 00:28:01.750 --> 00:28:03.220 align:middle line:84% where if you're interested in art 00:28:03.220 --> 00:28:05.920 align:middle line:84% that have lots of books about art and lots 00:28:05.920 --> 00:28:07.642 align:middle line:84% of big coffee table books or whatever. 00:28:07.642 --> 00:28:09.850 align:middle line:84% And so their thought was that-- the Quan and Williams 00:28:09.850 --> 00:28:12.582 align:middle line:84% argument is that you really could be overestimating what 00:28:12.582 --> 00:28:15.040 align:middle line:84% Amazon is doing because there are many different bookstores 00:28:15.040 --> 00:28:17.373 align:middle line:84% and many different bookstores have many different books. 00:28:17.373 --> 00:28:20.813 align:middle line:84% So the gain from the online over going to the bookstore 00:28:20.813 --> 00:28:22.230 align:middle line:84% that you know would have the books 00:28:22.230 --> 00:28:24.840 align:middle line:90% you want could be much smaller. 00:28:24.840 --> 00:28:27.820 align:middle line:84% They study this not in the case of books. 00:28:27.820 --> 00:28:29.230 align:middle line:90% They pick a different industry. 00:28:29.230 --> 00:28:31.350 align:middle line:90% Their industry is shoes. 00:28:31.350 --> 00:28:34.000 align:middle line:84% They don't say it, one suspects that they have-- 00:28:34.000 --> 00:28:36.690 align:middle line:84% they had a friend who worked at Zappos, because they are working 00:28:36.690 --> 00:28:41.850 align:middle line:84% with what they describe as a very large online shoe 00:28:41.850 --> 00:28:45.600 align:middle line:90% store in the year 2012, 2013. 00:28:45.600 --> 00:28:48.480 align:middle line:84% And just here's one example of why 00:28:48.480 --> 00:28:52.350 align:middle line:84% we would think that physical shoe stores would 00:28:52.350 --> 00:28:53.500 align:middle line:90% carry different things. 00:28:53.500 --> 00:28:59.790 align:middle line:84% This graph is entirely done with sales on the online site. 00:28:59.790 --> 00:29:03.180 align:middle line:84% But what they graph is, this is a state level analysis 00:29:03.180 --> 00:29:05.790 align:middle line:90% for, I think, 49 states. 00:29:05.790 --> 00:29:07.260 align:middle line:84% And what they put on this is just 00:29:07.260 --> 00:29:11.610 align:middle line:84% a dot for every state of what fraction of the shoes 00:29:11.610 --> 00:29:14.670 align:middle line:84% that are sold to that state are boots 00:29:14.670 --> 00:29:17.070 align:middle line:90% and what fraction are sandals. 00:29:17.070 --> 00:29:20.830 align:middle line:84% And what do you notice on the right side of this graph? 00:29:20.830 --> 00:29:23.050 align:middle line:84% We take the two warmest US states, which 00:29:23.050 --> 00:29:25.190 align:middle line:90% would be Hawaii and Florida. 00:29:25.190 --> 00:29:27.250 align:middle line:84% And in Hawaii and Florida, people 00:29:27.250 --> 00:29:31.090 align:middle line:84% don't buy a whole lot of boots and they buy a lot of sandals. 00:29:31.090 --> 00:29:34.180 align:middle line:84% I guess especially in Florida, people buy a lot of sandals. 00:29:34.180 --> 00:29:36.820 align:middle line:84% And then you go down here and you 00:29:36.820 --> 00:29:40.500 align:middle line:84% look at what people are buying in the coldest states. 00:29:40.500 --> 00:29:44.410 align:middle line:84% So it's, I don't know if that's North Dakota, Wyoming, 00:29:44.410 --> 00:29:47.170 align:middle line:84% Minnesota, places like that, people buy 00:29:47.170 --> 00:29:49.030 align:middle line:90% a lot of boots in North Dakota. 00:29:49.030 --> 00:29:53.230 align:middle line:84% People buy fewer sandals in North Dakota. 00:29:53.230 --> 00:29:55.090 align:middle line:84% Although I guess sort of surprising things 00:29:55.090 --> 00:29:56.040 align:middle line:90% you learned from this. 00:29:56.040 --> 00:29:57.498 align:middle line:84% It turns out people in North Dakota 00:29:57.498 --> 00:29:59.050 align:middle line:84% and Wyoming actually buy more sandals 00:29:59.050 --> 00:30:01.990 align:middle line:84% than people in middle temperature states. 00:30:01.990 --> 00:30:03.318 align:middle line:90% I guess it may be-- 00:30:03.318 --> 00:30:03.860 align:middle line:90% I don't know. 00:30:03.860 --> 00:30:05.277 align:middle line:84% Maybe if you live in a cold place, 00:30:05.277 --> 00:30:08.335 align:middle line:84% you go on vacation to Florida and you need to buy sandals. 00:30:08.335 --> 00:30:10.210 align:middle line:84% Or if you live in a medium temperature state, 00:30:10.210 --> 00:30:11.920 align:middle line:84% you just don't go on as many vacations 00:30:11.920 --> 00:30:13.370 align:middle line:90% to warm weather states. 00:30:13.370 --> 00:30:14.230 align:middle line:90% I don't know. 00:30:14.230 --> 00:30:16.990 align:middle line:84% But anyway, but their thought from this graph 00:30:16.990 --> 00:30:18.550 align:middle line:84% is if you look at if this is what 00:30:18.550 --> 00:30:20.590 align:middle line:90% happens for online retailers. 00:30:20.590 --> 00:30:22.690 align:middle line:90% You imagine that you walk into-- 00:30:22.690 --> 00:30:24.760 align:middle line:84% even take a national chain store, 00:30:24.760 --> 00:30:28.940 align:middle line:84% you imagine that the national chains in their North Dakota 00:30:28.940 --> 00:30:31.580 align:middle line:84% stores, they have lots of boots on the shelves, 00:30:31.580 --> 00:30:34.760 align:middle line:84% and in their Florida stores, they have very few winter 00:30:34.760 --> 00:30:36.530 align:middle line:90% boots on the shelves. 00:30:36.530 --> 00:30:38.370 align:middle line:84% And so this would just be an example. 00:30:38.370 --> 00:30:42.440 align:middle line:84% You know, this doesn't prove that local stores 00:30:42.440 --> 00:30:47.650 align:middle line:84% tailor their selections to what customers want. 00:30:47.650 --> 00:30:50.330 align:middle line:84% But this at least suggests that local stores ought 00:30:50.330 --> 00:30:52.490 align:middle line:84% to tailor their selections to their customers, 00:30:52.490 --> 00:30:55.580 align:middle line:84% and that ought to give you back some of the gains 00:30:55.580 --> 00:30:58.130 align:middle line:84% that we think is Amazon's gains from product variety should 00:30:58.130 --> 00:31:00.670 align:middle line:84% be gains from product variety at these stores. 00:31:00.670 --> 00:31:04.210 align:middle line:90% 00:31:04.210 --> 00:31:10.430 align:middle line:84% You know, the paper has a second data set where they also 00:31:10.430 --> 00:31:14.390 align:middle line:84% got access to the inventories of local stores 00:31:14.390 --> 00:31:19.400 align:middle line:84% at Macy's and Payless Shoe stores, which are just two chain 00:31:19.400 --> 00:31:21.270 align:middle line:90% retail stores. 00:31:21.270 --> 00:31:24.410 align:middle line:84% And they find that there is substantial differences 00:31:24.410 --> 00:31:27.883 align:middle line:84% in the SKUs carried at the different stores. 00:31:27.883 --> 00:31:29.300 align:middle line:84% They don't really have enough data 00:31:29.300 --> 00:31:31.050 align:middle line:84% to say whether it's just completely random 00:31:31.050 --> 00:31:36.333 align:middle line:84% or how much of it is this store carrying boots and this one 00:31:36.333 --> 00:31:38.750 align:middle line:84% carrying sandals, or this one having just different brands 00:31:38.750 --> 00:31:39.417 align:middle line:90% from the others. 00:31:39.417 --> 00:31:42.860 align:middle line:84% But what they find is that especially at the Macy's, there 00:31:42.860 --> 00:31:45.380 align:middle line:84% are almost no shoes that all the Macy's have. 00:31:45.380 --> 00:31:47.855 align:middle line:84% Different Macy's just have different shoes. 00:31:47.855 --> 00:31:49.730 align:middle line:84% There are a lot of them, shoes, that are only 00:31:49.730 --> 00:31:51.510 align:middle line:90% found in one or two stores. 00:31:51.510 --> 00:31:53.780 align:middle line:84% Very few that are found in 100% of stores. 00:31:53.780 --> 00:31:56.800 align:middle line:84% At the Payless, there are a lot of store shoes that 00:31:56.800 --> 00:31:59.300 align:middle line:84% are found in every Payless and a lot of shoes that are found 00:31:59.300 --> 00:32:00.776 align:middle line:90% in like one or two Paylesses. 00:32:00.776 --> 00:32:04.280 align:middle line:90% 00:32:04.280 --> 00:32:08.210 align:middle line:84% OK, anyway, so how do they estimate? 00:32:08.210 --> 00:32:12.230 align:middle line:84% So what they want to do is estimate how important is 00:32:12.230 --> 00:32:14.780 align:middle line:84% this variation in consumer preferences 00:32:14.780 --> 00:32:18.050 align:middle line:84% and to estimate how important if stores just 00:32:18.050 --> 00:32:20.900 align:middle line:84% match their local preferences, would stores 00:32:20.900 --> 00:32:22.910 align:middle line:84% be able to get most of the benefits 00:32:22.910 --> 00:32:24.113 align:middle line:90% that you get from Amazon? 00:32:24.113 --> 00:32:26.990 align:middle line:90% 00:32:26.990 --> 00:32:29.970 align:middle line:84% So what they do is demand estimation, 00:32:29.970 --> 00:32:33.720 align:middle line:84% it's going to be a standard kind of nested logit demand. 00:32:33.720 --> 00:32:40.130 align:middle line:84% So this is utility, customer I, location L, buying product J. 00:32:40.130 --> 00:32:45.590 align:middle line:84% It's going to be a mean utility of product J in location L, 00:32:45.590 --> 00:32:50.160 align:middle line:84% plus sort of nested logit models, there's a shock. 00:32:50.160 --> 00:32:52.100 align:middle line:84% You have this idiosyncratic preference 00:32:52.100 --> 00:32:57.170 align:middle line:84% of a customer for all products in category J. 00:32:57.170 --> 00:32:58.700 align:middle line:84% So if I'm looking for snow boots, 00:32:58.700 --> 00:33:04.370 align:middle line:84% I have a snow boot specific shock, and then a weight. 00:33:04.370 --> 00:33:05.640 align:middle line:90% I have two shocks. 00:33:05.640 --> 00:33:07.280 align:middle line:84% One is the category-specific shock, 00:33:07.280 --> 00:33:09.990 align:middle line:84% and then one is the brand-specific shock. 00:33:09.990 --> 00:33:12.330 align:middle line:84% And they're going to estimate this lambda parameter 00:33:12.330 --> 00:33:14.163 align:middle line:84% so they can tell whether people's shocks are 00:33:14.163 --> 00:33:19.610 align:middle line:84% mostly category-specific or mostly product-specific. 00:33:19.610 --> 00:33:21.470 align:middle line:84% And then what they're going to assume 00:33:21.470 --> 00:33:24.650 align:middle line:84% is that the mean utility provided 00:33:24.650 --> 00:33:28.310 align:middle line:84% by product J in location L is a function 00:33:28.310 --> 00:33:33.060 align:middle line:84% of the product characteristics, a function of the product price. 00:33:33.060 --> 00:33:37.520 align:middle line:84% And then this other xi sub LJ term, 00:33:37.520 --> 00:33:42.440 align:middle line:84% which that's the idiosyncratic preference 00:33:42.440 --> 00:33:46.580 align:middle line:90% for this product in location L. 00:33:46.580 --> 00:33:48.150 align:middle line:84% They don't have the temperature here. 00:33:48.150 --> 00:33:51.330 align:middle line:84% So the people in Florida don't like boots would be here 00:33:51.330 --> 00:33:54.510 align:middle line:84% or people in Hawaii like sandals would be here. 00:33:54.510 --> 00:33:57.410 align:middle line:84% But then you could imagine there's also just 00:33:57.410 --> 00:33:59.660 align:middle line:84% like these mayonnaise-like preferences. 00:33:59.660 --> 00:34:04.230 align:middle line:84% Like, there's just UGGs boots are popular in New England. 00:34:04.230 --> 00:34:06.320 align:middle line:84% They're unfashionable in California. 00:34:06.320 --> 00:34:09.420 align:middle line:84% Or this brand of shoes is popular here, 00:34:09.420 --> 00:34:11.010 align:middle line:90% it's not popular here. 00:34:11.010 --> 00:34:14.014 align:middle line:84% All of that would go into the xi LJs. 00:34:14.014 --> 00:34:18.989 align:middle line:90% 00:34:18.989 --> 00:34:22.080 align:middle line:84% A real challenge to estimating demand 00:34:22.080 --> 00:34:26.969 align:middle line:84% is that most products, you know, like there's 00:34:26.969 --> 00:34:30.900 align:middle line:84% a very, very large number of shoes in the world. 00:34:30.900 --> 00:34:35.790 align:middle line:84% And it turns out that most shoes in most cities in most months 00:34:35.790 --> 00:34:38.179 align:middle line:90% have zero sales. 00:34:38.179 --> 00:34:42.170 align:middle line:84% And so dealing with quantity data when you have zeros 00:34:42.170 --> 00:34:43.730 align:middle line:90% is very difficult. 00:34:43.730 --> 00:34:47.000 align:middle line:84% If you remember, one of the things we-- 00:34:47.000 --> 00:34:50.224 align:middle line:84% standard approaches people do in the load 00:34:50.224 --> 00:34:52.500 align:middle line:84% to demand estimation is you have, 00:34:52.500 --> 00:34:54.920 align:middle line:84% you know, log of the share of product 00:34:54.920 --> 00:34:58.730 align:middle line:84% J minus log of the share of the outside good 00:34:58.730 --> 00:35:03.655 align:middle line:84% was X, J, beta minus Alpha, P, J, something like that. 00:35:03.655 --> 00:35:07.100 align:middle line:90% 00:35:07.100 --> 00:35:10.040 align:middle line:84% You convert to the market shares to get back to the product 00:35:10.040 --> 00:35:12.530 align:middle line:84% characteristics, and then you can run IV regressions 00:35:12.530 --> 00:35:14.010 align:middle line:90% to estimate demand. 00:35:14.010 --> 00:35:17.090 align:middle line:84% You know, you can't do these log transformations 00:35:17.090 --> 00:35:20.450 align:middle line:84% when you have a whole bunch of zeros in your data. 00:35:20.450 --> 00:35:23.030 align:middle line:84% And so what do you do when you have a whole bunch of zeros 00:35:23.030 --> 00:35:24.020 align:middle line:90% in your data? 00:35:24.020 --> 00:35:29.960 align:middle line:84% Anyway, what they do in this paper is discuss a GMM 00:35:29.960 --> 00:35:34.850 align:middle line:84% approach to try to deal with all the zeros 00:35:34.850 --> 00:35:36.600 align:middle line:84% You also don't want to throw out-- 00:35:36.600 --> 00:35:38.308 align:middle line:84% another thing people do is sometimes just 00:35:38.308 --> 00:35:39.430 align:middle line:90% throw out all the zeros. 00:35:39.430 --> 00:35:41.790 align:middle line:84% You clearly don't want to throw out all the zeros 00:35:41.790 --> 00:35:45.060 align:middle line:84% because the goods that have zero sales in one month 00:35:45.060 --> 00:35:47.140 align:middle line:84% will have positive sales in the next month. 00:35:47.140 --> 00:35:48.820 align:middle line:84% They all have positive mean utilities. 00:35:48.820 --> 00:35:50.550 align:middle line:84% You don't want to remove them and say, 00:35:50.550 --> 00:35:54.570 align:middle line:84% no one wants to buy from those goods. 00:35:54.570 --> 00:35:59.280 align:middle line:84% In particular, what they note is that the number 00:35:59.280 --> 00:36:03.420 align:middle line:84% of shoes that are sold by this online website 00:36:03.420 --> 00:36:06.540 align:middle line:84% in a particular city, in a particular month 00:36:06.540 --> 00:36:11.240 align:middle line:84% is smaller than the inventory of a large shoe store. 00:36:11.240 --> 00:36:15.470 align:middle line:84% So if you actually were to take the realized shoe sales and say, 00:36:15.470 --> 00:36:17.120 align:middle line:84% those people's deltas are only based 00:36:17.120 --> 00:36:19.850 align:middle line:84% on the shoes that actually got bought, what you'd conclude 00:36:19.850 --> 00:36:21.800 align:middle line:84% is that a single physical store could actually 00:36:21.800 --> 00:36:23.678 align:middle line:90% supply the entire market. 00:36:23.678 --> 00:36:25.220 align:middle line:84% And so it's just, it's very important 00:36:25.220 --> 00:36:27.710 align:middle line:84% to correctly deal with all of these shoes that 00:36:27.710 --> 00:36:31.350 align:middle line:84% are in the store and have zero sales in a month. 00:36:31.350 --> 00:36:33.445 align:middle line:84% But, you know, people are buying. 00:36:33.445 --> 00:36:34.820 align:middle line:84% You just have these many products 00:36:34.820 --> 00:36:37.710 align:middle line:84% all with a fraction of expected sales. 00:36:37.710 --> 00:36:39.213 align:middle line:84% But those are what people buy among. 00:36:39.213 --> 00:36:40.880 align:middle line:84% And so they have to deal with the zeros. 00:36:40.880 --> 00:36:44.470 align:middle line:84% Anyway, one thing you'll notice in the paper, 00:36:44.470 --> 00:36:47.170 align:middle line:84% like, they spend a while talking about how to deal with demand 00:36:47.170 --> 00:36:49.288 align:middle line:84% estimation when you have lots of zeros. 00:36:49.288 --> 00:36:50.830 align:middle line:84% If you ever have a paper where you're 00:36:50.830 --> 00:36:54.700 align:middle line:84% trying to do demand estimation with lots of zeros, 00:36:54.700 --> 00:36:56.802 align:middle line:84% they discuss their method and other methods 00:36:56.802 --> 00:36:58.510 align:middle line:84% that people have used to deal with zeros. 00:36:58.510 --> 00:37:01.210 align:middle line:84% I think theirs is thought to be a fairly good method for doing 00:37:01.210 --> 00:37:05.610 align:middle line:84% demand estimation when you have a whole bunch of zeros. 00:37:05.610 --> 00:37:11.460 align:middle line:84% But so anyway, the basic idea is they aggregate sales 00:37:11.460 --> 00:37:13.150 align:middle line:90% to the location month level. 00:37:13.150 --> 00:37:16.020 align:middle line:90% They have 213 locations. 00:37:16.020 --> 00:37:22.150 align:middle line:84% Most of the 213 locations are metropolitan areas. 00:37:22.150 --> 00:37:27.120 align:middle line:84% But then in addition, they have each metropolitan area 00:37:27.120 --> 00:37:28.950 align:middle line:84% as a location observation, and then they 00:37:28.950 --> 00:37:31.680 align:middle line:84% have the rest of the state not in metropolitan areas 00:37:31.680 --> 00:37:33.390 align:middle line:90% as an observation. 00:37:33.390 --> 00:37:36.780 align:middle line:84% Anyway, so they have 213 different locations. 00:37:36.780 --> 00:37:41.040 align:middle line:84% They do something that's kind of like a BLP estimation 00:37:41.040 --> 00:37:43.000 align:middle line:90% except dealing with the zeros. 00:37:43.000 --> 00:37:45.870 align:middle line:90% So they aggregate sales. 00:37:45.870 --> 00:37:48.353 align:middle line:84% They do something very much like, I 00:37:48.353 --> 00:37:50.520 align:middle line:84% don't know if you remember back to Tobias's lecture, 00:37:50.520 --> 00:37:53.520 align:middle line:84% you're BLP instruments for price are 00:37:53.520 --> 00:37:55.500 align:middle line:84% some function of the characteristics of all 00:37:55.500 --> 00:37:57.130 align:middle line:90% the competing brands. 00:37:57.130 --> 00:38:00.330 align:middle line:84% So they have not a very convincing 00:38:00.330 --> 00:38:03.070 align:middle line:84% instrumentation strategy, but that's what's available. 00:38:03.070 --> 00:38:06.520 align:middle line:84% So they have a BLP style instrumentation strategy 00:38:06.520 --> 00:38:08.030 align:middle line:90% dealing with the zeros. 00:38:08.030 --> 00:38:10.480 align:middle line:84% And it's not a pure cross section 00:38:10.480 --> 00:38:14.710 align:middle line:84% because there's price variation, because the prices vary 00:38:14.710 --> 00:38:15.560 align:middle line:90% across months. 00:38:15.560 --> 00:38:18.550 align:middle line:84% And you can see when this shoe is on sale in some month, 00:38:18.550 --> 00:38:20.620 align:middle line:90% do sales go up. 00:38:20.620 --> 00:38:23.210 align:middle line:84% Product reviews are also time varying. 00:38:23.210 --> 00:38:25.330 align:middle line:84% Shoes come out and they have no reviews over time. 00:38:25.330 --> 00:38:27.160 align:middle line:90% They build up reviews. 00:38:27.160 --> 00:38:30.290 align:middle line:84% Obviously, higher reviews are good for sales. 00:38:30.290 --> 00:38:34.090 align:middle line:84% It turns out, though, also zero reviews are good for sales 00:38:34.090 --> 00:38:36.820 align:middle line:84% or positively associated with sales as well. 00:38:36.820 --> 00:38:40.240 align:middle line:84% And this may be because zero reviews means 00:38:40.240 --> 00:38:43.390 align:middle line:84% it's a brand new product that's just being introduced 00:38:43.390 --> 00:38:45.760 align:middle line:84% with the latest fashion or that's being advertised 00:38:45.760 --> 00:38:48.130 align:middle line:90% or something like that. 00:38:48.130 --> 00:38:52.210 align:middle line:84% Anyway, and then the products, that is also time varying. 00:38:52.210 --> 00:38:54.250 align:middle line:84% So anyway, with all of that, they're 00:38:54.250 --> 00:38:58.360 align:middle line:84% able to estimate all these coefficients. 00:38:58.360 --> 00:39:02.470 align:middle line:84% And they can estimate a distribution of the delta LJ's. 00:39:02.470 --> 00:39:06.280 align:middle line:84% And the delta LJ's indicate that there 00:39:06.280 --> 00:39:09.970 align:middle line:84% is a lot of cross-market heterogeneity in preferences. 00:39:09.970 --> 00:39:11.558 align:middle line:84% And, you know, you have to have faith 00:39:11.558 --> 00:39:13.600 align:middle line:84% that they're dealing with the zeros appropriately 00:39:13.600 --> 00:39:14.267 align:middle line:90% to deal with it. 00:39:14.267 --> 00:39:17.410 align:middle line:84% It's not just in some markets these shoes actually 00:39:17.410 --> 00:39:20.133 align:middle line:84% have sales, in some markets these have actually have sales. 00:39:20.133 --> 00:39:22.300 align:middle line:84% But I think they've done a fairly good job of trying 00:39:22.300 --> 00:39:25.030 align:middle line:84% to argue that, no, there's underlying 00:39:25.030 --> 00:39:27.910 align:middle line:84% cross sectional variation in the propensities 00:39:27.910 --> 00:39:30.350 align:middle line:84% to purchase different shoes in different places. 00:39:30.350 --> 00:39:32.080 align:middle line:84% And again, I think some of that is 00:39:32.080 --> 00:39:34.700 align:middle line:84% like the boots versus sandals approach. 00:39:34.700 --> 00:39:39.110 align:middle line:84% And some of that is just, you know, shoes are a fashion good. 00:39:39.110 --> 00:39:40.870 align:middle line:84% And there are brands of sneakers, 00:39:40.870 --> 00:39:43.360 align:middle line:84% brands of shoes, brands of whatever that 00:39:43.360 --> 00:39:45.980 align:middle line:84% are more or less fashionable in different cities. 00:39:45.980 --> 00:39:48.940 align:middle line:84% For whatever reason, this has become something people wear 00:39:48.940 --> 00:39:53.180 align:middle line:84% and people are willing to pay more for those. 00:39:53.180 --> 00:39:58.520 align:middle line:84% OK, so how do they estimate the gains from product variety? 00:39:58.520 --> 00:40:00.350 align:middle line:84% Estimated gains from product variety, this 00:40:00.350 --> 00:40:02.420 align:middle line:90% is done with a counterfactual. 00:40:02.420 --> 00:40:06.470 align:middle line:84% So they know all the sales that Zappos, 00:40:06.470 --> 00:40:08.750 align:middle line:84% or whatever the retailer is, made. 00:40:08.750 --> 00:40:10.790 align:middle line:84% And then they think about a counterfactual world 00:40:10.790 --> 00:40:13.850 align:middle line:84% where instead of choosing from all the shoes at Zappos, 00:40:13.850 --> 00:40:16.910 align:middle line:84% people could only choose from a limited supply. 00:40:16.910 --> 00:40:18.720 align:middle line:84% What do they do for the limited supply? 00:40:18.720 --> 00:40:21.260 align:middle line:84% Well, the first thing they do is they 00:40:21.260 --> 00:40:24.980 align:middle line:84% use the data on the Macy's and Payless Shoe Stores 00:40:24.980 --> 00:40:28.820 align:middle line:84% to estimate the total number of products 00:40:28.820 --> 00:40:32.130 align:middle line:84% available in each market as a function of the population. 00:40:32.130 --> 00:40:35.780 align:middle line:84% So they would be doing, how many different brands of shoes 00:40:35.780 --> 00:40:37.460 align:middle line:90% are available in Boston? 00:40:37.460 --> 00:40:39.860 align:middle line:84% Which you get by looking at the overlap of all the 00:40:39.860 --> 00:40:42.140 align:middle line:84% Macy's in Boston and the overlap of the Macy's 00:40:42.140 --> 00:40:44.790 align:middle line:84% in Boston with the Payless Shoe Stores in Boston. 00:40:44.790 --> 00:40:46.700 align:middle line:84% And then they have census data knowing 00:40:46.700 --> 00:40:49.820 align:middle line:84% how many shoe stores there are in Boston 00:40:49.820 --> 00:40:51.690 align:middle line:84% and how large they are in Boston. 00:40:51.690 --> 00:40:55.350 align:middle line:84% They assume that they have the same overlap as they're seeing 00:40:55.350 --> 00:40:57.070 align:middle line:90% in the Macy's and Paylesses. 00:40:57.070 --> 00:41:00.630 align:middle line:84% And they're trying to estimate how many total shoes are there 00:41:00.630 --> 00:41:05.423 align:middle line:84% in the Boston area that you can buy at physical retail stores. 00:41:05.423 --> 00:41:07.590 align:middle line:84% Obviously, that's a bit of an aggressive assumption, 00:41:07.590 --> 00:41:09.690 align:middle line:84% and that is assuming that people could actually 00:41:09.690 --> 00:41:11.610 align:middle line:84% know which local Boston shoe store 00:41:11.610 --> 00:41:15.750 align:middle line:84% to go to buy whatever particular brand of boots they want. 00:41:15.750 --> 00:41:19.830 align:middle line:84% And then what they assume is that offline consumers 00:41:19.830 --> 00:41:22.440 align:middle line:84% have the choice of the NL products 00:41:22.440 --> 00:41:26.910 align:middle line:84% that have the highest delta LJ's to choose from so 00:41:26.910 --> 00:41:32.400 align:middle line:84% that the physical stores in each state or each city 00:41:32.400 --> 00:41:34.780 align:middle line:90% carry the most popular brands. 00:41:34.780 --> 00:41:38.580 align:middle line:84% So again, it's kind of an aggressive assumption in that, 00:41:38.580 --> 00:41:41.370 align:middle line:84% you know, local stores are tailoring their demand to what 00:41:41.370 --> 00:41:44.020 align:middle line:84% people want and they're tailoring it perfectly. 00:41:44.020 --> 00:41:45.790 align:middle line:84% They're picking the most popular brands. 00:41:45.790 --> 00:41:50.760 align:middle line:84% And so, you know, it's both we observe. 00:41:50.760 --> 00:41:55.050 align:middle line:84% It's both we know where to buy shoes physically, 00:41:55.050 --> 00:41:58.290 align:middle line:84% depending on what we want, which store to go to, 00:41:58.290 --> 00:42:00.960 align:middle line:84% and the shoe stores are doing perfect tailoring and carrying 00:42:00.960 --> 00:42:02.893 align:middle line:84% the most popular shoes, not making mistakes. 00:42:02.893 --> 00:42:05.490 align:middle line:90% 00:42:05.490 --> 00:42:09.900 align:middle line:84% But anyway, so subject to that caveat, as the paper suggests, 00:42:09.900 --> 00:42:12.110 align:middle line:84% the online product variety benefits 00:42:12.110 --> 00:42:16.080 align:middle line:84% are much smaller than what Brynjolfsson, Hu, and Smith 00:42:16.080 --> 00:42:16.870 align:middle line:90% said. 00:42:16.870 --> 00:42:20.070 align:middle line:84% And they're much smaller in two different senses. 00:42:20.070 --> 00:42:24.690 align:middle line:84% The first is that if you look at the gain from product variety, 00:42:24.690 --> 00:42:29.010 align:middle line:84% they argue that about half of the gain from product variety 00:42:29.010 --> 00:42:33.270 align:middle line:84% would be provided anyway by local stores 00:42:33.270 --> 00:42:36.510 align:middle line:84% by just having heterogeneity in local stores. 00:42:36.510 --> 00:42:39.146 align:middle line:84% So if you think about this sort of continuum. 00:42:39.146 --> 00:42:42.220 align:middle line:90% 00:42:42.220 --> 00:42:44.920 align:middle line:84% Let me do the Brynjolfsson, Hu, and Smith numbers. 00:42:44.920 --> 00:42:50.050 align:middle line:90% 00:42:50.050 --> 00:42:52.832 align:middle line:84% So Brynjolfsson, Hu, and Smith are all looking at sort 00:42:52.832 --> 00:42:57.670 align:middle line:84% of, what's the consumer surplus from there being one bookstore 00:42:57.670 --> 00:43:01.800 align:middle line:90% that has 100,000 titles? 00:43:01.800 --> 00:43:08.750 align:middle line:84% And then we have this sort of Amazon with 2 million titles 00:43:08.750 --> 00:43:11.450 align:middle line:84% and they're arguing that this is 10. 00:43:11.450 --> 00:43:14.340 align:middle line:84% This is like, if this is 1, this is 10. 00:43:14.340 --> 00:43:15.733 align:middle line:90% You get 10 times-- 00:43:15.733 --> 00:43:17.650 align:middle line:84% I'm putting the consumer surplus in this axis. 00:43:17.650 --> 00:43:21.420 align:middle line:90% 00:43:21.420 --> 00:43:26.880 align:middle line:84% And then what Quan and Williams are saying is that we have-- 00:43:26.880 --> 00:43:28.845 align:middle line:90% so this is-- 00:43:28.845 --> 00:43:30.720 align:middle line:84% Quan and Williams saying, well, our benchmark 00:43:30.720 --> 00:43:37.080 align:middle line:84% is going to be not one bookstore, but all Boston 00:43:37.080 --> 00:43:39.150 align:middle line:90% bookstores. 00:43:39.150 --> 00:43:41.220 align:middle line:84% They don't know what all Boston bookstores have, 00:43:41.220 --> 00:43:43.590 align:middle line:84% but they just assume that the Boston bookstores have 00:43:43.590 --> 00:43:46.290 align:middle line:84% these non-overlapping sets of books 00:43:46.290 --> 00:43:49.740 align:middle line:84% and the non-overlapping sets of books tailored to what the MIT 00:43:49.740 --> 00:43:53.070 align:middle line:84% students want, tailored to what everybody else in Boston wants. 00:43:53.070 --> 00:43:57.780 align:middle line:84% And what they find is that, you know-- 00:43:57.780 --> 00:44:00.600 align:middle line:84% well, they do all shoe stores, but all shoe stores 00:44:00.600 --> 00:44:02.800 align:middle line:84% thing compared to the one shoe store, 00:44:02.800 --> 00:44:06.690 align:middle line:84% they're arguing that all shoe stores get you about half 00:44:06.690 --> 00:44:08.800 align:middle line:84% of the way to the Amazon rankings. 00:44:08.800 --> 00:44:10.890 align:middle line:84% So in some sense that Quan and Williams 00:44:10.890 --> 00:44:18.980 align:middle line:84% may be overstating the gains by 100% or something. 00:44:18.980 --> 00:44:20.757 align:middle line:90% And so that's one point. 00:44:20.757 --> 00:44:22.340 align:middle line:84% And then the second point of the paper 00:44:22.340 --> 00:44:25.165 align:middle line:84% that is not emphasized as much in just 00:44:25.165 --> 00:44:26.540 align:middle line:84% in the title and the introduction 00:44:26.540 --> 00:44:34.310 align:middle line:84% is that they also find that the gains from product variety 00:44:34.310 --> 00:44:39.080 align:middle line:84% are just much smaller than are reported elsewhere. 00:44:39.080 --> 00:44:49.950 align:middle line:84% So that whereas Brynjolfsson, Hu, and Smith 00:44:49.950 --> 00:44:52.500 align:middle line:84% are saying that these gains are 10 times the gain 00:44:52.500 --> 00:44:55.020 align:middle line:90% from a 5% price cut. 00:44:55.020 --> 00:44:58.770 align:middle line:84% What they find is that the gains in the shoe store data 00:44:58.770 --> 00:45:06.060 align:middle line:84% look like they're, you know, 2 times the benefit of a-- 00:45:06.060 --> 00:45:07.200 align:middle line:90% so anyway, sorry. 00:45:07.200 --> 00:45:10.530 align:middle line:90% 00:45:10.530 --> 00:45:12.450 align:middle line:90% If I make this 1, this is 2. 00:45:12.450 --> 00:45:15.270 align:middle line:84% So the gain is just equivalent to a 5. 00:45:15.270 --> 00:45:19.710 align:middle line:84% The gain is like equivalent to a 5% price cut. 00:45:19.710 --> 00:45:23.440 align:middle line:84% So it's like this number is here, this number is here. 00:45:23.440 --> 00:45:25.460 align:middle line:84% So what they're arguing is that the-- 00:45:25.460 --> 00:45:27.900 align:middle line:90% actually, no, I take that back. 00:45:27.900 --> 00:45:29.970 align:middle line:84% It's this one compared to this one 00:45:29.970 --> 00:45:33.480 align:middle line:84% I think is roughly equivalent to a 5% price cut. 00:45:33.480 --> 00:45:36.930 align:middle line:84% So what they're finding is that the gains from product variety 00:45:36.930 --> 00:45:39.360 align:middle line:84% are just much, much smaller in shoes 00:45:39.360 --> 00:45:42.400 align:middle line:84% than Brynjolfsson, Hu, and Smith said they were in books. 00:45:42.400 --> 00:45:45.840 align:middle line:84% And because the gains are so much smaller, 00:45:45.840 --> 00:45:49.770 align:middle line:84% and you take out this thing is halfway there already, what 00:45:49.770 --> 00:45:52.160 align:middle line:84% you're getting is the product variety gains 00:45:52.160 --> 00:45:56.120 align:middle line:84% are really only like equivalent to a 5% price cut. 00:45:56.120 --> 00:45:59.360 align:middle line:90% 00:45:59.360 --> 00:46:03.590 align:middle line:84% So it's saying, at least in the shoe sense, 00:46:03.590 --> 00:46:06.620 align:middle line:84% product variety gains are, you know, 00:46:06.620 --> 00:46:08.555 align:middle line:84% 10 times smaller than you would have 00:46:08.555 --> 00:46:10.430 align:middle line:84% might have thought from the Brynjolfsson, Hu, 00:46:10.430 --> 00:46:11.630 align:middle line:90% and Smith paper. 00:46:11.630 --> 00:46:14.510 align:middle line:84% And obviously, you know, is that because shoes 00:46:14.510 --> 00:46:16.360 align:middle line:90% are very different from books? 00:46:16.360 --> 00:46:17.730 align:middle line:90% That's one possibility. 00:46:17.730 --> 00:46:20.253 align:middle line:84% You know, like, every book is very, very different 00:46:20.253 --> 00:46:21.170 align:middle line:90% from every other book. 00:46:21.170 --> 00:46:23.990 align:middle line:84% You could imagine, once you've got a million shoes, 00:46:23.990 --> 00:46:26.150 align:middle line:84% the ability to make a million first shoe 00:46:26.150 --> 00:46:29.070 align:middle line:84% that's very different from all existing shoes is very hard. 00:46:29.070 --> 00:46:31.430 align:middle line:84% Whereas every new book is just completely different 00:46:31.430 --> 00:46:32.940 align:middle line:90% text from every other book. 00:46:32.940 --> 00:46:34.820 align:middle line:84% There may be more inherent variety in books 00:46:34.820 --> 00:46:37.160 align:middle line:84% than there are in shoes, but I think 00:46:37.160 --> 00:46:41.300 align:middle line:84% it may also be that they may have done a better 00:46:41.300 --> 00:46:44.810 align:middle line:84% job of dealing with estimating what's 00:46:44.810 --> 00:46:47.790 align:middle line:84% true product variety versus spurious product variety. 00:46:47.790 --> 00:46:50.313 align:middle line:90% You know, like the-- 00:46:50.313 --> 00:46:51.230 align:middle line:90% I don't if I remember. 00:46:51.230 --> 00:46:52.880 align:middle line:84% Like, I've discussed, like, when you're 00:46:52.880 --> 00:46:55.220 align:middle line:84% trying to estimate consumer variety in demand, 00:46:55.220 --> 00:46:57.020 align:middle line:84% like, consumer welfare from demand 00:46:57.020 --> 00:46:59.010 align:middle line:84% and you have heterogeneous products, 00:46:59.010 --> 00:47:04.580 align:middle line:84% you know, everyone's always doing this X, J, Beta 00:47:04.580 --> 00:47:09.142 align:middle line:84% minus alpha, P, J plus Epsilon, IJ. 00:47:09.142 --> 00:47:10.850 align:middle line:84% And they're always assuming these Epsilon 00:47:10.850 --> 00:47:14.360 align:middle line:84% IJ's are logit distributed, which has a fairly thick tail, 00:47:14.360 --> 00:47:16.280 align:middle line:84% and that every Epsilon IJ is independent 00:47:16.280 --> 00:47:19.400 align:middle line:90% of every other Epsilon IJ. 00:47:19.400 --> 00:47:22.370 align:middle line:84% Again, once you get to the millionth different shoe, 00:47:22.370 --> 00:47:25.010 align:middle line:84% is it really an Epsilon IJ that's completely independent 00:47:25.010 --> 00:47:27.420 align:middle line:84% of your taste for every other shoe out there? 00:47:27.420 --> 00:47:29.180 align:middle line:84% It's highly likely those Epsilons are 00:47:29.180 --> 00:47:31.340 align:middle line:90% highly correlated across shoes. 00:47:31.340 --> 00:47:32.900 align:middle line:90% At least by having the-- 00:47:32.900 --> 00:47:37.310 align:middle line:84% you know, they've got those category-specific shocks 00:47:37.310 --> 00:47:39.200 align:middle line:84% by putting in the category-specific shocks 00:47:39.200 --> 00:47:42.680 align:middle line:84% and assuming the magnitude and doing a careful job of thinking 00:47:42.680 --> 00:47:45.320 align:middle line:84% about the zeros, they may be avoiding 00:47:45.320 --> 00:47:48.170 align:middle line:84% overstating how big all these Epsilon IJ's are 00:47:48.170 --> 00:47:50.490 align:middle line:84% and all this surplus we're getting from all 00:47:50.490 --> 00:47:53.840 align:middle line:90% these new different products. 00:47:53.840 --> 00:47:55.490 align:middle line:84% And I guess one other paper I thought 00:47:55.490 --> 00:47:57.115 align:middle line:84% I'd mention-- one other thing I thought 00:47:57.115 --> 00:48:00.110 align:middle line:84% I'd mentioned in the paper is that they 00:48:00.110 --> 00:48:02.240 align:middle line:84% know that Macy's and Payless carry different shoes 00:48:02.240 --> 00:48:03.140 align:middle line:90% in different stores. 00:48:03.140 --> 00:48:06.320 align:middle line:84% They ask, how important is it for stores 00:48:06.320 --> 00:48:08.160 align:middle line:90% to tailor their selections? 00:48:08.160 --> 00:48:09.680 align:middle line:84% And what they find is that there's 00:48:09.680 --> 00:48:11.010 align:middle line:90% a gain that firms should do it. 00:48:11.010 --> 00:48:12.180 align:middle line:90% It's not that large. 00:48:12.180 --> 00:48:16.550 align:middle line:84% The gain from customizing stores is estimated to be about 6% 00:48:16.550 --> 00:48:17.240 align:middle line:90% in revenue. 00:48:17.240 --> 00:48:19.520 align:middle line:84% So you could just carry all the standard shoes 00:48:19.520 --> 00:48:21.830 align:middle line:84% or you could tailor the demand in your city. 00:48:21.830 --> 00:48:25.370 align:middle line:84% You tailor the demand in your city, you get 6% more profit. 00:48:25.370 --> 00:48:27.600 align:middle line:84% So it seems like it's worth doing. 00:48:27.600 --> 00:48:29.600 align:middle line:84% Firms should want to do this, especially if they 00:48:29.600 --> 00:48:30.990 align:middle line:90% have thin profit margins. 00:48:30.990 --> 00:48:37.250 align:middle line:84% But it's boundedly-- tailoring is boundedly profitable? 00:48:37.250 --> 00:48:43.920 align:middle line:90% 00:48:43.920 --> 00:48:44.780 align:middle line:90% Yeah? 00:48:44.780 --> 00:48:46.560 align:middle line:84% AUDIENCE: Was this paper actually 00:48:46.560 --> 00:48:47.800 align:middle line:90% response to names on paper? 00:48:47.800 --> 00:48:49.543 align:middle line:84% Or was it just kind of disjointed? 00:48:49.543 --> 00:48:51.460 align:middle line:84% Because couldn't you just do the same exercise 00:48:51.460 --> 00:48:55.075 align:middle line:84% they did in the Amazon paper but with shoes [INAUDIBLE] exchange? 00:48:55.075 --> 00:48:57.200 align:middle line:84% GLENN ELLISON: Yeah, so certainly they could have-- 00:48:57.200 --> 00:49:00.580 align:middle line:90% 00:49:00.580 --> 00:49:02.530 align:middle line:84% you know, so they actually have-- 00:49:02.530 --> 00:49:03.730 align:middle line:90% I don't know. 00:49:03.730 --> 00:49:06.080 align:middle line:84% They do several different calculations in the paper. 00:49:06.080 --> 00:49:07.910 align:middle line:84% Like, suppose we estimate demand this way. 00:49:07.910 --> 00:49:09.410 align:middle line:84% Suppose we estimate demand this way. 00:49:09.410 --> 00:49:11.230 align:middle line:84% Suppose we estimate demand this way. 00:49:11.230 --> 00:49:13.810 align:middle line:84% You know, in some sense doing demand worse and worse 00:49:13.810 --> 00:49:16.460 align:middle line:84% and worse, doing worse and worse demand estimation. 00:49:16.460 --> 00:49:19.430 align:middle line:84% And they do present several things. 00:49:19.430 --> 00:49:21.250 align:middle line:84% If you've done this estimate, instead you'd 00:49:21.250 --> 00:49:23.050 align:middle line:84% get numbers that are two times as big. 00:49:23.050 --> 00:49:24.467 align:middle line:84% If you do these estimates instead, 00:49:24.467 --> 00:49:26.240 align:middle line:84% you get things that are five times as big. 00:49:26.240 --> 00:49:29.020 align:middle line:90% So they do get-- 00:49:29.020 --> 00:49:33.873 align:middle line:84% you know, it's not exactly the same calculation as the Amazon 00:49:33.873 --> 00:49:35.290 align:middle line:84% data, but they do something that's 00:49:35.290 --> 00:49:39.370 align:middle line:84% more like the naive calculations where you treat everything 00:49:39.370 --> 00:49:42.100 align:middle line:84% as national market shares and use logit models 00:49:42.100 --> 00:49:44.080 align:middle line:84% and show that you can get things many times 00:49:44.080 --> 00:49:45.788 align:middle line:84% bigger than the estimates they've gotten. 00:49:45.788 --> 00:49:52.490 align:middle line:90% 00:49:52.490 --> 00:49:55.313 align:middle line:84% Yeah, I think these are the things that obviously you 00:49:55.313 --> 00:49:56.980 align:middle line:84% couldn't have done with the Amazon data. 00:49:56.980 --> 00:49:59.630 align:middle line:84% Like, the Brynjolfsson, Hu, and Smith 00:49:59.630 --> 00:50:02.420 align:middle line:84% didn't have anything about what is in local bookstores 00:50:02.420 --> 00:50:05.120 align:middle line:84% and how much overlap there is across local bookstores 00:50:05.120 --> 00:50:06.115 align:middle line:90% and stuff like that. 00:50:06.115 --> 00:50:13.270 align:middle line:90% 00:50:13.270 --> 00:50:16.660 align:middle line:84% But certainly, the paper is very much written in response 00:50:16.660 --> 00:50:19.600 align:middle line:84% to literature saying there are these really big private-- 00:50:19.600 --> 00:50:21.980 align:middle line:84% really big product variety benefits. 00:50:21.980 --> 00:50:29.260 align:middle line:84% And I think the paper is more about the benefits are smaller 00:50:29.260 --> 00:50:31.780 align:middle line:84% because of the variety in retail than the benefits are 00:50:31.780 --> 00:50:34.040 align:middle line:84% smaller if you just do the econometrics correctly. 00:50:34.040 --> 00:50:35.295 align:middle line:90% But I think it is about both. 00:50:35.295 --> 00:50:38.690 align:middle line:90% 00:50:38.690 --> 00:50:39.514 align:middle line:90% OK. 00:50:39.514 --> 00:50:42.500 align:middle line:84% So anyway, final paper I want to cover. 00:50:42.500 --> 00:50:46.430 align:middle line:84% This is a working paper that Sara and I have, 00:50:46.430 --> 00:50:49.700 align:middle line:84% which I guess has two different motivations. 00:50:49.700 --> 00:50:53.570 align:middle line:84% You know, one is also this is a paper about trying to estimate 00:50:53.570 --> 00:50:56.240 align:middle line:84% how big product variety benefits are 00:50:56.240 --> 00:50:58.940 align:middle line:84% and how big product variety benefits are in an area where 00:50:58.940 --> 00:51:01.460 align:middle line:84% we think the product variety benefits would be bigger, 00:51:01.460 --> 00:51:03.000 align:middle line:90% which is in used books. 00:51:03.000 --> 00:51:06.800 align:middle line:84% And then the other is also sometimes you can say it is a, 00:51:06.800 --> 00:51:12.260 align:middle line:84% you know, Chicago school critique of our search 00:51:12.260 --> 00:51:13.610 align:middle line:90% and obfuscation paper. 00:51:13.610 --> 00:51:15.690 align:middle line:84% Where in the search and obfuscation paper 00:51:15.690 --> 00:51:20.030 align:middle line:84% we argued that you see these not super low 00:51:20.030 --> 00:51:22.370 align:middle line:84% and dispersed prices for computer components 00:51:22.370 --> 00:51:27.830 align:middle line:84% because the firms have an incentive 00:51:27.830 --> 00:51:29.900 align:middle line:84% to obfuscate what they're selling the price 00:51:29.900 --> 00:51:32.930 align:middle line:84% is at and invent these add-ons or invent complicated things 00:51:32.930 --> 00:51:37.370 align:middle line:84% that raise search costs and keep prices high and disperses 00:51:37.370 --> 00:51:39.110 align:middle line:84% as they are in a search cost model 00:51:39.110 --> 00:51:42.470 align:middle line:84% rather than letting perfect competition come out here. 00:51:42.470 --> 00:51:47.090 align:middle line:84% This paper is going to provide an alternate explanation for why 00:51:47.090 --> 00:51:48.810 align:middle line:90% prices can be high in dispersed. 00:51:48.810 --> 00:51:51.110 align:middle line:84% It's going to be that in some sense, 00:51:51.110 --> 00:51:53.490 align:middle line:84% prices are high and dispersed for two different reasons. 00:51:53.490 --> 00:51:56.150 align:middle line:84% One is prices are high on the internet 00:51:56.150 --> 00:51:59.700 align:middle line:84% because the internet is so good at matching products to people. 00:51:59.700 --> 00:52:06.770 align:middle line:84% And if you think about a used book that you had as a child. 00:52:06.770 --> 00:52:10.700 align:middle line:84% You know, and then you have a used book as a child, 00:52:10.700 --> 00:52:12.090 align:middle line:90% you now have a niece or nephew. 00:52:12.090 --> 00:52:16.310 align:middle line:84% You want to buy them the same picture book that you remember 00:52:16.310 --> 00:52:17.640 align:middle line:90% from when you were a kid. 00:52:17.640 --> 00:52:20.170 align:middle line:84% You're willing to pay many dollars for that book. 00:52:20.170 --> 00:52:21.920 align:middle line:84% There are a million other people out there 00:52:21.920 --> 00:52:23.820 align:middle line:84% who didn't remember that book when they were kids. 00:52:23.820 --> 00:52:26.445 align:middle line:84% There are probably hundreds of copies of it sitting in physical 00:52:26.445 --> 00:52:29.210 align:middle line:84% used bookstores with $0.10 marked inside the cover that no 00:52:29.210 --> 00:52:30.510 align:middle line:90% one is ever buying. 00:52:30.510 --> 00:52:32.930 align:middle line:84% If we get those books in the hands of the right people, 00:52:32.930 --> 00:52:34.950 align:middle line:84% the willingness to pay is way, way higher. 00:52:34.950 --> 00:52:39.170 align:middle line:84% In fact, actually, we say the paper started when Sara was 00:52:39.170 --> 00:52:41.300 align:middle line:84% writing a paper on pharmaceuticals and needed some 00:52:41.300 --> 00:52:46.400 align:middle line:84% obscure pharmaceutical book that someone had written in the 1970s 00:52:46.400 --> 00:52:48.770 align:middle line:84% that the MIT libraries didn't carry. 00:52:48.770 --> 00:52:53.280 align:middle line:84% You know, websites selling used books online were new. 00:52:53.280 --> 00:52:56.690 align:middle line:84% She actually went on to websites selling used books, 00:52:56.690 --> 00:52:59.750 align:middle line:84% found the book, and it's like, wow, I can find this book. 00:52:59.750 --> 00:53:01.320 align:middle line:90% I've been trying to find this. 00:53:01.320 --> 00:53:02.233 align:middle line:90% It's only $15. 00:53:02.233 --> 00:53:03.900 align:middle line:84% I can pay it out of my research account. 00:53:03.900 --> 00:53:06.140 align:middle line:84% I'm thrilled to get it for $15 instead of $100. 00:53:06.140 --> 00:53:09.000 align:middle line:84% Gets it and then got the book in like $0.25 00:53:09.000 --> 00:53:11.140 align:middle line:84% was penciled inside the front cover of it. 00:53:11.140 --> 00:53:14.250 align:middle line:84% And she realized it had been sitting in a bookstore in Iowa 00:53:14.250 --> 00:53:16.500 align:middle line:84% City since some economics professor retired 00:53:16.500 --> 00:53:19.570 align:middle line:84% for the last decade, and no one wanted to pay $0.25 for this. 00:53:19.570 --> 00:53:21.750 align:middle line:84% And she was thrilled to get it for $15 00:53:21.750 --> 00:53:24.480 align:middle line:84% and realizing that this is potential huge increase 00:53:24.480 --> 00:53:27.162 align:middle line:84% in social welfare of match quality. 00:53:27.162 --> 00:53:28.620 align:middle line:84% Anyway, so part of what we're doing 00:53:28.620 --> 00:53:31.423 align:middle line:84% is underestimate that match quality. 00:53:31.423 --> 00:53:33.090 align:middle line:84% And then the second thought in the paper 00:53:33.090 --> 00:53:36.840 align:middle line:84% is that, it's nice to explain two things with one story. 00:53:36.840 --> 00:53:38.520 align:middle line:84% Like, prices are high and dispersed 00:53:38.520 --> 00:53:39.520 align:middle line:90% because of search costs. 00:53:39.520 --> 00:53:41.320 align:middle line:84% But we think there are two stories here. 00:53:41.320 --> 00:53:43.530 align:middle line:84% Prices are high because of match quality. 00:53:43.530 --> 00:53:46.912 align:middle line:84% And prices are dispersed because some websites are better 00:53:46.912 --> 00:53:48.870 align:middle line:84% at selling things than others, and the websites 00:53:48.870 --> 00:53:49.860 align:middle line:84% that are better at selling things 00:53:49.860 --> 00:53:52.470 align:middle line:84% are able to command much higher prices than the ones that 00:53:52.470 --> 00:53:54.660 align:middle line:90% have limited attention. 00:53:54.660 --> 00:53:56.830 align:middle line:84% Anyway, so what does the paper do? 00:53:56.830 --> 00:53:58.940 align:middle line:90% It's a three-part paper. 00:53:58.940 --> 00:54:02.910 align:middle line:84% First, it's developing simple models to basically illustrate 00:54:02.910 --> 00:54:04.623 align:middle line:90% the ideas. 00:54:04.623 --> 00:54:06.290 align:middle line:84% You know, one and two are really, again, 00:54:06.290 --> 00:54:09.920 align:middle line:84% just me always aspiring to be like Hendrix and Porter 00:54:09.920 --> 00:54:12.295 align:middle line:84% and having this theory, having predictions of the theory, 00:54:12.295 --> 00:54:14.503 align:middle line:84% testing the predictions of the theory to convince you 00:54:14.503 --> 00:54:15.360 align:middle line:90% that it's right. 00:54:15.360 --> 00:54:16.880 align:middle line:84% And then the back part of the paper 00:54:16.880 --> 00:54:20.300 align:middle line:84% is structural estimates to try to say how big are the welfare 00:54:20.300 --> 00:54:24.860 align:middle line:84% gains from improved accessibility of used books sold 00:54:24.860 --> 00:54:28.050 align:middle line:84% online versus in physical stores. 00:54:28.050 --> 00:54:30.490 align:middle line:90% OK. 00:54:30.490 --> 00:54:34.080 align:middle line:84% So first, let me start with the basic theory. 00:54:34.080 --> 00:54:38.900 align:middle line:84% And this is a simple model of dynamic monopoly pricing 00:54:38.900 --> 00:54:40.490 align:middle line:90% of a variety that we don't-- 00:54:40.490 --> 00:54:43.580 align:middle line:84% I guess I never covered in the IO class. 00:54:43.580 --> 00:54:45.800 align:middle line:84% But so imagine that you have a seller 00:54:45.800 --> 00:54:48.090 align:middle line:84% and the seller just has one unit to sell. 00:54:48.090 --> 00:54:49.350 align:middle line:90% There's no production here. 00:54:49.350 --> 00:54:51.180 align:middle line:90% So this is a unique item. 00:54:51.180 --> 00:54:52.760 align:middle line:90% You're just trying to sell it. 00:54:52.760 --> 00:54:54.620 align:middle line:84% And then you have customers who have 00:54:54.620 --> 00:54:58.040 align:middle line:84% values VJ drawn from some distribution F 00:54:58.040 --> 00:55:00.980 align:middle line:84% and they just arrive at some Poisson rate gamma. 00:55:00.980 --> 00:55:02.780 align:middle line:90% They walk into your store. 00:55:02.780 --> 00:55:05.000 align:middle line:84% When they walk into your store and are 00:55:05.000 --> 00:55:06.950 align:middle line:84% interested in this pharmaceutical textbook, 00:55:06.950 --> 00:55:10.260 align:middle line:84% you have to either sell it to them immediately or not at all. 00:55:10.260 --> 00:55:13.007 align:middle line:84% Because if they walk into your store, they see the book, 00:55:13.007 --> 00:55:15.090 align:middle line:84% they look at the price, they don't want to buy it, 00:55:15.090 --> 00:55:15.780 align:middle line:90% they walk out. 00:55:15.780 --> 00:55:16.947 align:middle line:90% You'll never see them again. 00:55:16.947 --> 00:55:19.310 align:middle line:90% 00:55:19.310 --> 00:55:21.610 align:middle line:90% How should you set your prices? 00:55:21.610 --> 00:55:24.800 align:middle line:84% Well, you know, the demand from these 00:55:24.800 --> 00:55:26.610 align:middle line:84% are one-time consumers who are myopic. 00:55:26.610 --> 00:55:28.460 align:middle line:84% So the demand at price P is just 1 00:55:28.460 --> 00:55:30.510 align:middle line:84% minus capital F of P, the probability 00:55:30.510 --> 00:55:33.420 align:middle line:84% that the consumer's value is above P. 00:55:33.420 --> 00:55:37.317 align:middle line:84% So the expected profit is the expectation 00:55:37.317 --> 00:55:39.150 align:middle line:84% of setting a price of P and leaving it there 00:55:39.150 --> 00:55:42.600 align:middle line:84% is the expectation of pe to the minus rt, where 00:55:42.600 --> 00:55:48.000 align:middle line:84% t is the random time at which you will sell the book. 00:55:48.000 --> 00:55:49.530 align:middle line:84% In this model, as long as you set 00:55:49.530 --> 00:55:51.865 align:middle line:84% any p that's less than the upper bound of F, 00:55:51.865 --> 00:55:53.490 align:middle line:84% you'll always sell the book eventually, 00:55:53.490 --> 00:55:55.800 align:middle line:84% you just may sell it a million years in the future 00:55:55.800 --> 00:55:58.980 align:middle line:84% and the [INAUDIBLE] rt will eat up all your profits. 00:55:58.980 --> 00:56:03.150 align:middle line:84% OK, so what we're going to do is integrate pe 00:56:03.150 --> 00:56:06.960 align:middle line:84% to the minus rt times the density 00:56:06.960 --> 00:56:10.440 align:middle line:84% of the probability of selling the book 00:56:10.440 --> 00:56:12.760 align:middle line:90% in the interval tt plus dt. 00:56:12.760 --> 00:56:16.100 align:middle line:90% 00:56:16.100 --> 00:56:18.720 align:middle line:84% Gamma is the Poisson arrival rate. 00:56:18.720 --> 00:56:22.430 align:middle line:84% Gamma D of P is the Poisson arrival rate 00:56:22.430 --> 00:56:26.240 align:middle line:84% of people willing to pay at least P. And so 00:56:26.240 --> 00:56:28.280 align:middle line:84% the density of the time you sell it 00:56:28.280 --> 00:56:34.033 align:middle line:84% is gamma D of pe to the minus gamma D of pt. 00:56:34.033 --> 00:56:36.200 align:middle line:84% You may have seen like if you have a Poisson arrival 00:56:36.200 --> 00:56:40.460 align:middle line:84% rate, gamma, you know, the density of arrivals 00:56:40.460 --> 00:56:45.950 align:middle line:90% is gamma e to the minus gamma t. 00:56:45.950 --> 00:56:49.860 align:middle line:84% And so the arrival rate of people willing to pay p is this. 00:56:49.860 --> 00:56:54.390 align:middle line:84% So this is the density of the arrival distribution. 00:56:54.390 --> 00:56:57.330 align:middle line:84% And it turns out this is a nice, well-behaved integral. 00:56:57.330 --> 00:57:03.330 align:middle line:84% And so the price is a function of p as this expression. 00:57:03.330 --> 00:57:08.310 align:middle line:84% It's p times gamma D of p over r plus gamma D of p. 00:57:08.310 --> 00:57:13.120 align:middle line:84% And so as gamma D of p goes to 1, 00:57:13.120 --> 00:57:17.410 align:middle line:84% you're selling it almost instantly and your profit is p. 00:57:17.410 --> 00:57:21.595 align:middle line:84% As gamma D of p goes to zero, the r over 00:57:21.595 --> 00:57:23.470 align:middle line:84% outweighs the gamma D of p, and you basically 00:57:23.470 --> 00:57:26.500 align:middle line:84% are getting something that's going towards zero because this 00:57:26.500 --> 00:57:31.210 align:middle line:84% is the discounting and this is the price at which you sell it. 00:57:31.210 --> 00:57:35.030 align:middle line:84% And so we can think of monopoly pricing in two different ways. 00:57:35.030 --> 00:57:37.540 align:middle line:84% One way to think about it is just the monopoly price 00:57:37.540 --> 00:57:41.140 align:middle line:84% is the maximizer of this expression. 00:57:41.140 --> 00:57:43.750 align:middle line:84% Second way to think about it is like a dynamic programming 00:57:43.750 --> 00:57:44.980 align:middle line:90% approach. 00:57:44.980 --> 00:57:47.085 align:middle line:84% This good doesn't have any production costs, 00:57:47.085 --> 00:57:49.210 align:middle line:84% but it has an opportunity cost because you can only 00:57:49.210 --> 00:57:50.500 align:middle line:90% sell it once. 00:57:50.500 --> 00:57:55.570 align:middle line:84% And so if pi star is your equilibrium expected profit 00:57:55.570 --> 00:57:58.060 align:middle line:84% from holding one unit of this good, 00:57:58.060 --> 00:58:00.940 align:middle line:84% you're thinking about someone's in my store now, what price 00:58:00.940 --> 00:58:03.250 align:middle line:90% should I quote them? 00:58:03.250 --> 00:58:07.060 align:middle line:84% What I want to do is maximize p minus pi star D of p. 00:58:07.060 --> 00:58:09.820 align:middle line:84% I'm maximizing price minus opportunity cost 00:58:09.820 --> 00:58:10.910 align:middle line:90% times the demand. 00:58:10.910 --> 00:58:12.410 align:middle line:84% And so another way to think about it 00:58:12.410 --> 00:58:19.610 align:middle line:84% is that pm has to solve this where pi star is pi of pm. 00:58:19.610 --> 00:58:21.350 align:middle line:84% So it's two different ways to think 00:58:21.350 --> 00:58:23.843 align:middle line:84% about monopoly pricing for a unique item 00:58:23.843 --> 00:58:25.260 align:middle line:84% when you have a customer arrivals. 00:58:25.260 --> 00:58:29.030 align:middle line:90% 00:58:29.030 --> 00:58:32.360 align:middle line:84% Observations from this monopoly pricing model. 00:58:32.360 --> 00:58:36.795 align:middle line:84% Monopoly prices increasing in the customer arrival rate. 00:58:36.795 --> 00:58:38.670 align:middle line:84% Why is the increase in customer arrival rate? 00:58:38.670 --> 00:58:40.670 align:middle line:84% Because the higher is the customer arrival rate, 00:58:40.670 --> 00:58:42.620 align:middle line:84% the higher is your opportunity cost. 00:58:42.620 --> 00:58:47.040 align:middle line:84% This is, again, a 14.121 standard exam question. 00:58:47.040 --> 00:58:49.040 align:middle line:84% Your monopoly price is increasing 00:58:49.040 --> 00:58:52.850 align:middle line:84% in your cost, full stop because of this function 00:58:52.850 --> 00:58:54.420 align:middle line:90% as increasing differences. 00:58:54.420 --> 00:58:59.930 align:middle line:84% So monopoly price is increasing opportunity cost, therefore 00:58:59.930 --> 00:59:02.240 align:middle line:84% dynamic monopoly price increases in arrival rate 00:59:02.240 --> 00:59:04.760 align:middle line:84% because arrival rate increases your opportunity set. 00:59:04.760 --> 00:59:07.370 align:middle line:90% Opportunity cost goes up. 00:59:07.370 --> 00:59:10.640 align:middle line:84% So anyway, if consumers start to arrive at higher rates 00:59:10.640 --> 00:59:12.560 align:middle line:84% online than they did in physical stores, 00:59:12.560 --> 00:59:14.340 align:middle line:84% you're going to get higher prices. 00:59:14.340 --> 00:59:16.430 align:middle line:84% And if there's a rival rate heterogeneity 00:59:16.430 --> 00:59:18.710 align:middle line:84% across online stores, that's going 00:59:18.710 --> 00:59:20.010 align:middle line:90% to lead to price dispersion. 00:59:20.010 --> 00:59:22.545 align:middle line:84% And stores that are frequently visited by online consumers 00:59:22.545 --> 00:59:24.920 align:middle line:84% are going to be able to set higher prices because they're 00:59:24.920 --> 00:59:26.040 align:middle line:90% more patient. 00:59:26.040 --> 00:59:30.780 align:middle line:84% Those that have fewer visitors are going to set lower prices. 00:59:30.780 --> 00:59:33.840 align:middle line:84% How sensitive prices are to arrival rates 00:59:33.840 --> 00:59:38.450 align:middle line:84% depends on the thickness of the upper tail. 00:59:38.450 --> 00:59:41.540 align:middle line:84% If you have an extremely thick tailed distribution, 00:59:41.540 --> 00:59:45.830 align:middle line:84% like constant times p to the minus 1, then 00:59:45.830 --> 00:59:47.930 align:middle line:84% the monopoly price increases almost 00:59:47.930 --> 00:59:50.030 align:middle line:90% linearly in arrival rates. 00:59:50.030 --> 00:59:52.220 align:middle line:84% If you have a very thin tailed distribution 00:59:52.220 --> 00:59:55.970 align:middle line:84% with a greater than 1, you know, things increase more slowly 00:59:55.970 --> 00:59:58.850 align:middle line:90% with arrival rates. 00:59:58.850 --> 01:00:03.290 align:middle line:84% And then with if demand has this constant elasticity form, 01:00:03.290 --> 01:00:08.060 align:middle line:84% like this, then you get this very unusual theorem. 01:00:08.060 --> 01:00:10.820 align:middle line:84% So the theorem is, suppose the distribution of consumer 01:00:10.820 --> 01:00:12.680 align:middle line:84% valuations is such that demand has 01:00:12.680 --> 01:00:15.320 align:middle line:84% a truncated, constant elasticity form. 01:00:15.320 --> 01:00:17.990 align:middle line:84% Suppose the monopoly price is not 01:00:17.990 --> 01:00:19.850 align:middle line:84% at the kink of this demand curve at 1, 01:00:19.850 --> 01:00:22.970 align:middle line:84% but it's on the downward sloping part. 01:00:22.970 --> 01:00:27.530 align:middle line:84% Then the monopoly price maximizes social welfare, 01:00:27.530 --> 01:00:32.190 align:middle line:84% and social welfare is the price numerically, just numerically. 01:00:32.190 --> 01:00:35.420 align:middle line:84% Like, I want to say, what's the expected producer 01:00:35.420 --> 01:00:37.950 align:middle line:84% surplus plus consumer surplus from selling this book? 01:00:37.950 --> 01:00:40.520 align:middle line:84% The answer is, if that book is being sold for $12, 01:00:40.520 --> 01:00:42.620 align:middle line:84% the answer expected welfare is $12. 01:00:42.620 --> 01:00:45.080 align:middle line:84% So in some sense, like, it's the most powerful 01:00:45.080 --> 01:00:47.840 align:middle line:84% structural estimation theorem of all time, 01:00:47.840 --> 01:00:50.720 align:middle line:84% saying, if you want to estimate the social welfare 01:00:50.720 --> 01:00:53.250 align:middle line:84% from dynamic sales, you just write down the price. 01:00:53.250 --> 01:00:55.458 align:middle line:84% You don't need to compute the demand elasticities. 01:00:55.458 --> 01:00:57.750 align:middle line:84% You don't need to figure out consumer surplus, nothing. 01:00:57.750 --> 01:01:01.040 align:middle line:84% You just write down the price and that is the welfare. 01:01:01.040 --> 01:01:03.050 align:middle line:84% And then there's also this optimality theorem 01:01:03.050 --> 01:01:06.140 align:middle line:84% that there's no monopoly distortion. 01:01:06.140 --> 01:01:10.200 align:middle line:84% And so, like, why does this work out? 01:01:10.200 --> 01:01:15.590 align:middle line:84% You know, obviously the monopoly price 01:01:15.590 --> 01:01:19.720 align:middle line:90% maximizes monopolist profits. 01:01:19.720 --> 01:01:22.690 align:middle line:90% But what is consumer surplus? 01:01:22.690 --> 01:01:27.670 align:middle line:84% So if I have this distribution of consumer valuations. 01:01:27.670 --> 01:01:32.860 align:middle line:84% So I have this distribution of consumer valuations, F of V, 01:01:32.860 --> 01:01:35.380 align:middle line:84% and imagine it looks something like that. 01:01:35.380 --> 01:01:37.360 align:middle line:84% And suppose I set some price, pm. 01:01:37.360 --> 01:01:40.110 align:middle line:90% 01:01:40.110 --> 01:01:41.400 align:middle line:90% OK. 01:01:41.400 --> 01:01:47.490 align:middle line:84% When I sell the good, there's some consumer surplus. 01:01:47.490 --> 01:01:49.440 align:middle line:90% What's the consumer surplus? 01:01:49.440 --> 01:01:54.143 align:middle line:84% That's the expectation of V minus pm conditional on V 01:01:54.143 --> 01:01:54.810 align:middle line:90% greater than pm. 01:01:54.810 --> 01:01:58.150 align:middle line:90% 01:01:58.150 --> 01:02:01.840 align:middle line:84% How does that vary with the monopoly price? 01:02:01.840 --> 01:02:11.110 align:middle line:84% Like, in a linear demand curve, as I set the price really high, 01:02:11.110 --> 01:02:13.600 align:middle line:84% you know, the expected value conditional on being bigger 01:02:13.600 --> 01:02:15.910 align:middle line:90% than this is almost nothing. 01:02:15.910 --> 01:02:19.000 align:middle line:84% And so consumer surplus drops substantially 01:02:19.000 --> 01:02:20.580 align:middle line:90% with the monopoly price. 01:02:20.580 --> 01:02:23.380 align:middle line:84% But in a constant elasticity model, 01:02:23.380 --> 01:02:28.120 align:middle line:84% the expected value of V minus pm conditional on V 01:02:28.120 --> 01:02:32.560 align:middle line:84% being bigger than pm is proportional to pm. 01:02:32.560 --> 01:02:35.680 align:middle line:84% So in a constant elasticity demand curve, 01:02:35.680 --> 01:02:40.780 align:middle line:84% consumer surplus is proportional to the price. 01:02:40.780 --> 01:02:43.750 align:middle line:84% So if you set a really, really high price, 01:02:43.750 --> 01:02:46.010 align:middle line:84% you don't sell for a very long time, 01:02:46.010 --> 01:02:49.250 align:middle line:84% and so the pe to the minus rt part gets really small. 01:02:49.250 --> 01:02:51.790 align:middle line:84% But when you do sell it, the person because they're 01:02:51.790 --> 01:02:53.950 align:middle line:84% willing to pay $1 million for this book, 01:02:53.950 --> 01:02:58.640 align:middle line:84% their expected surplus is also on the order of $1 million. 01:02:58.640 --> 01:03:01.910 align:middle line:84% And so with constant elasticity demand, 01:03:01.910 --> 01:03:04.130 align:middle line:84% the consumer surplus is just always 01:03:04.130 --> 01:03:07.700 align:middle line:84% proportional to the price being paid, expected proportional 01:03:07.700 --> 01:03:08.550 align:middle line:90% to price. 01:03:08.550 --> 01:03:12.055 align:middle line:84% Therefore, by maximizing expected price paid 01:03:12.055 --> 01:03:13.430 align:middle line:84% discounted expected price, you're 01:03:13.430 --> 01:03:16.850 align:middle line:84% also maximizing discounted expected social welfare. 01:03:16.850 --> 01:03:18.710 align:middle line:90% So I view this as more a-- 01:03:18.710 --> 01:03:21.440 align:middle line:84% I mean, it's a neat estimation trick or something, 01:03:21.440 --> 01:03:23.540 align:middle line:84% but I view it more as a cautionary tale. 01:03:23.540 --> 01:03:26.870 align:middle line:84% The cautionary tale being, if in this demand estimation, 01:03:26.870 --> 01:03:31.400 align:middle line:84% you can't tell the demand apart from-- 01:03:31.400 --> 01:03:34.670 align:middle line:84% if you can't tell the valuation curve apart 01:03:34.670 --> 01:03:37.100 align:middle line:84% from a constant elasticity demand curve, 01:03:37.100 --> 01:03:39.650 align:middle line:84% then you're going to end up finding 01:03:39.650 --> 01:03:41.278 align:middle line:84% that the expected welfare is the price, 01:03:41.278 --> 01:03:43.070 align:middle line:84% regardless of what the demand elasticity is 01:03:43.070 --> 01:03:45.610 align:middle line:90% and everything else are. 01:03:45.610 --> 01:03:48.437 align:middle line:84% So it can be a first approximation. 01:03:48.437 --> 01:03:50.520 align:middle line:84% And it's also, if you want a better approximation, 01:03:50.520 --> 01:03:53.250 align:middle line:84% you need to be able to tell the shape of this curve 01:03:53.250 --> 01:03:56.085 align:middle line:84% apart from a constant elasticity one, which 01:03:56.085 --> 01:03:57.960 align:middle line:84% is going to be hard to do when the upper tail 01:03:57.960 --> 01:04:01.610 align:middle line:84% because you don't see things in the upper tail very often. 01:04:01.610 --> 01:04:04.010 align:middle line:90% OK. 01:04:04.010 --> 01:04:07.340 align:middle line:84% so now let me think about that's a monopoly model. 01:04:07.340 --> 01:04:09.600 align:middle line:84% Again, think about an oligopoly model. 01:04:09.600 --> 01:04:11.880 align:middle line:84% So you have two types of consumers. 01:04:11.880 --> 01:04:14.670 align:middle line:84% You have some people who just go straight. 01:04:14.670 --> 01:04:17.360 align:middle line:84% Like, they know some online bookstore, 01:04:17.360 --> 01:04:20.420 align:middle line:84% like Better World Books or Powell's Books. 01:04:20.420 --> 01:04:23.810 align:middle line:84% And they just go directly to PowellsBooks.com, 01:04:23.810 --> 01:04:25.120 align:middle line:90% look for a book there. 01:04:25.120 --> 01:04:27.800 align:middle line:90% 01:04:27.800 --> 01:04:30.830 align:middle line:84% We're going to have some consumers who have this, 01:04:30.830 --> 01:04:32.700 align:middle line:90% or you do a Google search. 01:04:32.700 --> 01:04:35.480 align:middle line:84% Google search reveals the name of one bookstore 01:04:35.480 --> 01:04:36.518 align:middle line:90% for the title you want. 01:04:36.518 --> 01:04:38.060 align:middle line:84% You just go to that bookstore and you 01:04:38.060 --> 01:04:41.050 align:middle line:84% buy from the first, the one that's highest ranked on Google. 01:04:41.050 --> 01:04:42.830 align:middle line:84% And then we have other people who 01:04:42.830 --> 01:04:45.920 align:middle line:84% are going to be shoppers who are going to arrive at a Poisson 01:04:45.920 --> 01:04:49.910 align:middle line:84% rate and they observe the prices at all the stores. 01:04:49.910 --> 01:04:52.562 align:middle line:90% Do I have a picture of AbeBooks? 01:04:52.562 --> 01:04:54.020 align:middle line:84% I don't have a picture of AbeBooks. 01:04:54.020 --> 01:04:54.520 align:middle line:90% OK. 01:04:54.520 --> 01:05:01.370 align:middle line:84% Anyway, we use data in the paper from a website 01:05:01.370 --> 01:05:02.790 align:middle line:90% called AbeBooks.com. 01:05:02.790 --> 01:05:08.450 align:middle line:90% 01:05:08.450 --> 01:05:12.810 align:middle line:84% AbeBooks, it's a used book aggregator site. 01:05:12.810 --> 01:05:14.040 align:middle line:90% You can go on there. 01:05:14.040 --> 01:05:17.400 align:middle line:84% You could type in any book you want that's out of print. 01:05:17.400 --> 01:05:20.240 align:middle line:84% It will find 100 online bookstores that 01:05:20.240 --> 01:05:22.660 align:middle line:90% sell that particular used book. 01:05:22.660 --> 01:05:24.950 align:middle line:90% OK. 01:05:24.950 --> 01:05:26.612 align:middle line:84% Actually, you know, AbeBooks used 01:05:26.612 --> 01:05:27.820 align:middle line:90% to be an independent company. 01:05:27.820 --> 01:05:29.080 align:middle line:90% It was purchased by Amazon. 01:05:29.080 --> 01:05:31.890 align:middle line:84% So if you go to the Amazon Buy Used button, 01:05:31.890 --> 01:05:36.200 align:middle line:84% Amazon Buy Used button is the AbeBooks listings. 01:05:36.200 --> 01:05:38.575 align:middle line:84% In some sense, Amazon invented the Buy Used button first. 01:05:38.575 --> 01:05:40.200 align:middle line:84% When they invented the Buy Used button, 01:05:40.200 --> 01:05:41.940 align:middle line:84% there was nothing there under Buy Used. 01:05:41.940 --> 01:05:44.590 align:middle line:84% You'd click Buy Used and there were no copies available. 01:05:44.590 --> 01:05:46.340 align:middle line:84% And then they realized, well, why don't we 01:05:46.340 --> 01:05:49.140 align:middle line:84% just buy AbeBooks that has 10 million used books listed 01:05:49.140 --> 01:05:51.430 align:middle line:84% and put all their listings under the Buy Used button? 01:05:51.430 --> 01:05:53.097 align:middle line:84% So now if you go to the Buy Used button, 01:05:53.097 --> 01:05:55.530 align:middle line:84% you're seeing the same AbeBooks listings. 01:05:55.530 --> 01:05:56.640 align:middle line:90% OK. 01:05:56.640 --> 01:05:58.087 align:middle line:84% Anyway, so our model is that there 01:05:58.087 --> 01:05:59.795 align:middle line:84% are sort of these two types of customers, 01:05:59.795 --> 01:06:02.280 align:middle line:84% the ones going to Amazon and clicking Buy Used 01:06:02.280 --> 01:06:05.560 align:middle line:84% and getting this curated list of all the things sorted by price, 01:06:05.560 --> 01:06:07.560 align:middle line:84% and then the people who are searching for a book 01:06:07.560 --> 01:06:10.200 align:middle line:84% by just typing it into Google or going directly 01:06:10.200 --> 01:06:12.030 align:middle line:84% to a used bookstore that they know online 01:06:12.030 --> 01:06:14.440 align:middle line:90% and finding it that way. 01:06:14.440 --> 01:06:16.860 align:middle line:84% You know, we know that this model 01:06:16.860 --> 01:06:19.953 align:middle line:84% can have pure strategy dispersed price equilibrium, not so 01:06:19.953 --> 01:06:21.870 align:middle line:84% different from the dispersed price equilibrium 01:06:21.870 --> 01:06:25.670 align:middle line:84% I did when I did differentiated product search models. 01:06:25.670 --> 01:06:27.450 align:middle line:84% You know, the firms here, the dispersion 01:06:27.450 --> 01:06:30.000 align:middle line:84% is going to be the firms that have the higher arrival rates 01:06:30.000 --> 01:06:33.440 align:middle line:84% are going to set high prices and mostly serve the non-shoppers. 01:06:33.440 --> 01:06:35.340 align:middle line:84% If you don't have a lot of non-shoppers, 01:06:35.340 --> 01:06:38.670 align:middle line:84% you're going to set a low price and compete for the shoppers. 01:06:38.670 --> 01:06:42.450 align:middle line:84% And when I'm thinking about, like, how does the online prices 01:06:42.450 --> 01:06:43.890 align:middle line:90% differ from the offline prices? 01:06:43.890 --> 01:06:46.030 align:middle line:84% It helps think that there are two reasons. 01:06:46.030 --> 01:06:47.762 align:middle line:90% One is the match quality reason. 01:06:47.762 --> 01:06:50.220 align:middle line:84% The higher arrival rates shift the distribution of monopoly 01:06:50.220 --> 01:06:52.710 align:middle line:90% prices to the right. 01:06:52.710 --> 01:06:54.930 align:middle line:84% The second is the competition effect, 01:06:54.930 --> 01:06:57.970 align:middle line:84% which is there's a desire to attract the shoppers. 01:06:57.970 --> 01:07:01.020 align:middle line:84% So the firms that have low arrival rates of non-shoppers 01:07:01.020 --> 01:07:03.150 align:middle line:84% are going to be competing against each other 01:07:03.150 --> 01:07:05.592 align:middle line:84% to try to sell in a more Bertrand-like environment 01:07:05.592 --> 01:07:07.550 align:middle line:84% to the people who are using the search engines. 01:07:07.550 --> 01:07:10.610 align:middle line:90% 01:07:10.610 --> 01:07:13.200 align:middle line:84% So I think on the next page I have this illustration. 01:07:13.200 --> 01:07:16.590 align:middle line:84% So here's a numerical simulation of the model. 01:07:16.590 --> 01:07:20.960 align:middle line:84% So this dashed gray line would be the distribution of monopoly 01:07:20.960 --> 01:07:22.850 align:middle line:84% prices if there are exponentially 01:07:22.850 --> 01:07:24.240 align:middle line:90% distributed arrival rates. 01:07:24.240 --> 01:07:26.390 align:middle line:84% And so the firms with high arrival rates 01:07:26.390 --> 01:07:30.320 align:middle line:84% at high prices, those with low arrival rates at low prices. 01:07:30.320 --> 01:07:34.700 align:middle line:84% If I increase the value of the good 01:07:34.700 --> 01:07:36.830 align:middle line:84% because people are getting a better matched good, 01:07:36.830 --> 01:07:38.390 align:middle line:84% the monopoly price distribution would 01:07:38.390 --> 01:07:42.250 align:middle line:84% shift to the right from the gray line to the blue line. 01:07:42.250 --> 01:07:45.735 align:middle line:84% But then if I add in this atom of shoppers 01:07:45.735 --> 01:07:47.860 align:middle line:84% who are just going to buy for the lowest price book 01:07:47.860 --> 01:07:51.040 align:middle line:84% and are fairly price insensitive, what you get 01:07:51.040 --> 01:07:52.930 align:middle line:90% is the distribution of prices. 01:07:52.930 --> 01:07:55.540 align:middle line:84% All the firms that are going to be setting the lowest prices 01:07:55.540 --> 01:07:58.180 align:middle line:84% are now Bertrand competing against the other firms 01:07:58.180 --> 01:08:00.220 align:middle line:84% who are also going to set very low prices. 01:08:00.220 --> 01:08:03.620 align:middle line:84% And so the low end of the price distribution gets pulled down. 01:08:03.620 --> 01:08:06.570 align:middle line:84% The upper tail really doesn't get affected at all. 01:08:06.570 --> 01:08:08.320 align:middle line:84% The upper tail doesn't get affected at all 01:08:08.320 --> 01:08:10.630 align:middle line:84% because these firms were just monopoly pricing 01:08:10.630 --> 01:08:13.870 align:middle line:84% and serving the high value customers who happen 01:08:13.870 --> 01:08:15.520 align:middle line:90% on their website occasionally. 01:08:15.520 --> 01:08:17.821 align:middle line:84% They don't bother to compete for the shoppers. 01:08:17.821 --> 01:08:19.779 align:middle line:84% They just leave the high price and just exploit 01:08:19.779 --> 01:08:21.340 align:middle line:90% the non-shoppers. 01:08:21.340 --> 01:08:23.710 align:middle line:84% And so what you get is this increase 01:08:23.710 --> 01:08:28.060 align:middle line:84% in dispersion, which comes from the upper tail really 01:08:28.060 --> 01:08:30.700 align:middle line:84% having moved up because people have higher values, 01:08:30.700 --> 01:08:35.600 align:middle line:84% and then the lower tail getting pulled down by the competition. 01:08:35.600 --> 01:08:38.170 align:middle line:84% And so we think those are the two things you should expect 01:08:38.170 --> 01:08:40.630 align:middle line:84% is that the online price distribution has this thicker 01:08:40.630 --> 01:08:43.850 align:middle line:84% upper tail relative to the offline price distribution, 01:08:43.850 --> 01:08:47.050 align:middle line:84% but then it also has more mass in the lower tail, perhaps. 01:08:47.050 --> 01:08:52.870 align:middle line:90% 01:08:52.870 --> 01:08:55.600 align:middle line:84% OK, so how do we try to test this in the data? 01:08:55.600 --> 01:09:00.250 align:middle line:84% So we gathered a data set that had 335 books 01:09:00.250 --> 01:09:03.580 align:middle line:84% representative of traditional used bookstores' inventories. 01:09:03.580 --> 01:09:07.000 align:middle line:90% The way this worked is we just-- 01:09:07.000 --> 01:09:08.819 align:middle line:84% actually, Sara did some of this herself. 01:09:08.819 --> 01:09:12.760 align:middle line:84% We also just sent an RA to some physical bookstores with 01:09:12.760 --> 01:09:17.979 align:middle line:84% the instructions, I want you to go to the physical bookstores 01:09:17.979 --> 01:09:22.060 align:middle line:84% and just go to the first shelf, take every 17th book off 01:09:22.060 --> 01:09:25.210 align:middle line:84% the shelf, take that book, write down the title, what it is, 01:09:25.210 --> 01:09:28.479 align:middle line:84% write down the price, and then just go shelf to shelf to shelf 01:09:28.479 --> 01:09:31.300 align:middle line:90% picking out books. 01:09:31.300 --> 01:09:33.460 align:middle line:84% So we have hundreds, what we call 01:09:33.460 --> 01:09:35.390 align:middle line:90% standard books in the data set. 01:09:35.390 --> 01:09:38.960 align:middle line:84% We also oversampled what we call local interest books. 01:09:38.960 --> 01:09:40.630 align:middle line:84% Like, if you go into a Boston bookstore, 01:09:40.630 --> 01:09:42.880 align:middle line:84% you'll notice there's a shelf of books 01:09:42.880 --> 01:09:46.270 align:middle line:90% about Boston and Boston stuff. 01:09:46.270 --> 01:09:50.319 align:middle line:84% And so what we said is, if there's a local interest 01:09:50.319 --> 01:09:53.359 align:middle line:84% section, so there's a Boston shelf in a Boston bookstore, 01:09:53.359 --> 01:09:55.790 align:middle line:84% one of the bookstores was in Atlanta, there 01:09:55.790 --> 01:09:58.185 align:middle line:84% if there's a book about Christmas in Georgia, 01:09:58.185 --> 01:10:00.560 align:middle line:84% you know, like, find the shelf that has the Georgia books 01:10:00.560 --> 01:10:03.440 align:middle line:90% and oversample those. 01:10:03.440 --> 01:10:06.580 align:middle line:84% And then we also separate out popular books. 01:10:06.580 --> 01:10:08.450 align:middle line:84% Popular books are the things like-- 01:10:08.450 --> 01:10:10.040 align:middle line:84% I mean, not necessarily Harry Potter, 01:10:10.040 --> 01:10:11.720 align:middle line:84% but things that are popular enough 01:10:11.720 --> 01:10:14.728 align:middle line:84% that every used bookstore is probably going to have a copy. 01:10:14.728 --> 01:10:17.270 align:middle line:84% There are some things that these are former mega bestsellers. 01:10:17.270 --> 01:10:19.490 align:middle line:84% So it's not so rare to find these, 01:10:19.490 --> 01:10:22.130 align:middle line:84% whereas the standard books are all going to be completely 01:10:22.130 --> 01:10:24.680 align:middle line:84% non-overlapping across stores because just the average used 01:10:24.680 --> 01:10:26.013 align:middle line:90% book-- 01:10:26.013 --> 01:10:27.180 align:middle line:90% there are millions of books. 01:10:27.180 --> 01:10:28.850 align:middle line:84% The average used book is just no one's going 01:10:28.850 --> 01:10:30.090 align:middle line:90% to have another copy of that. 01:10:30.090 --> 01:10:31.965 align:middle line:84% You could search the bookstores all your life 01:10:31.965 --> 01:10:34.720 align:middle line:84% and probably not find another copy of a standard book. 01:10:34.720 --> 01:10:35.600 align:middle line:90% OK. 01:10:35.600 --> 01:10:38.700 align:middle line:84% Anyway, so we have four data collections. 01:10:38.700 --> 01:10:41.420 align:middle line:84% What we did is we first just randomly selected these books 01:10:41.420 --> 01:10:43.310 align:middle line:84% from physical bookstores, and then we 01:10:43.310 --> 01:10:47.450 align:middle line:84% went online and searched for those books at AbeBooks.com. 01:10:47.450 --> 01:10:52.160 align:middle line:84% And then we noticed that in between the time 01:10:52.160 --> 01:10:54.470 align:middle line:84% we collected this and we wrote the paper, 01:10:54.470 --> 01:10:57.050 align:middle line:90% Amazon bought AbeBooks. 01:10:57.050 --> 01:11:00.320 align:middle line:84% And we thought, wow, that's a great second accidental 01:11:00.320 --> 01:11:03.800 align:middle line:84% opportunity here is that Amazon buying AbeBooks and putting 01:11:03.800 --> 01:11:06.000 align:middle line:84% the listings under Buy Used probably 01:11:06.000 --> 01:11:08.000 align:middle line:84% just dramatically increased the number of people 01:11:08.000 --> 01:11:12.960 align:middle line:84% who are looking at the curated search engine listings for them. 01:11:12.960 --> 01:11:16.550 align:middle line:84% And so then after Amazon bought and incorporated the AbeBooks 01:11:16.550 --> 01:11:19.140 align:middle line:84% listings, we went out and collected the data again 01:11:19.140 --> 01:11:20.720 align:middle line:90% a second time. 01:11:20.720 --> 01:11:22.780 align:middle line:84% And we collected the data again a second time so 01:11:22.780 --> 01:11:25.700 align:middle line:84% that we would have a before and after on prices, before 01:11:25.700 --> 01:11:27.980 align:middle line:90% and after the AbeBooks listings. 01:11:27.980 --> 01:11:30.270 align:middle line:84% And then we went and collected the data fourth time 01:11:30.270 --> 01:11:31.520 align:middle line:90% a few months later. 01:11:31.520 --> 01:11:34.340 align:middle line:84% We realized it would be useful to have quantity data, 01:11:34.340 --> 01:11:36.470 align:middle line:84% and a way to get quantity data is 01:11:36.470 --> 01:11:38.690 align:middle line:84% to search a bookstore for the book, 01:11:38.690 --> 01:11:40.838 align:middle line:84% search that same bookstore two months later, 01:11:40.838 --> 01:11:42.380 align:middle line:84% see if they're still listing the book 01:11:42.380 --> 01:11:44.270 align:middle line:84% and infer if they're not still listing 01:11:44.270 --> 01:11:47.362 align:middle line:84% the book two months later, they may have sold the book. 01:11:47.362 --> 01:11:49.320 align:middle line:84% Now they may have just also just taken it down, 01:11:49.320 --> 01:11:52.070 align:middle line:84% but it's a pain to take things down from your website. 01:11:52.070 --> 01:11:55.320 align:middle line:84% You probably take them down when you sell them. 01:11:55.320 --> 01:11:58.700 align:middle line:84% So anyway, that's so we have these four data collections. 01:11:58.700 --> 01:12:01.450 align:middle line:84% And last two is how we got quantity data. 01:12:01.450 --> 01:12:04.610 align:middle line:90% 01:12:04.610 --> 01:12:07.230 align:middle line:84% So what are the predictions we're trying to test? 01:12:07.230 --> 01:12:10.610 align:middle line:84% Prediction 1 is that online prices 01:12:10.610 --> 01:12:12.770 align:middle line:84% should be more dispersed than offline prices 01:12:12.770 --> 01:12:16.000 align:middle line:90% and have a thicker upper tail. 01:12:16.000 --> 01:12:19.870 align:middle line:84% Prediction 2 is, local interest books may already 01:12:19.870 --> 01:12:22.360 align:middle line:84% be well matched by the offline stores, 01:12:22.360 --> 01:12:25.330 align:middle line:84% and so the online-offline gap should be smaller 01:12:25.330 --> 01:12:27.110 align:middle line:90% for the local interest books. 01:12:27.110 --> 01:12:30.163 align:middle line:84% In particular, I put this quote up here. 01:12:30.163 --> 01:12:32.080 align:middle line:84% I found this online review of one of the books 01:12:32.080 --> 01:12:34.240 align:middle line:84% in our data set, the Mount Vernon Street Warrens-- 01:12:34.240 --> 01:12:35.030 align:middle line:90% A Boston Story. 01:12:35.030 --> 01:12:37.360 align:middle line:84% I don't know if you guys have been to Beacon Hill. 01:12:37.360 --> 01:12:41.650 align:middle line:84% Mount Vernon is one of the streets on Beacon Hill. 01:12:41.650 --> 01:12:43.690 align:middle line:84% It's a cute street or whatever, with the gas 01:12:43.690 --> 01:12:46.420 align:middle line:84% lamps and the brick sidewalks and stuff. 01:12:46.420 --> 01:12:50.170 align:middle line:84% Anyway, if you go into a local Boston bookstore, one of them, 01:12:50.170 --> 01:12:52.960 align:middle line:84% some RA found a copy of this Mount Vernon Street Warrens. 01:12:52.960 --> 01:12:54.700 align:middle line:84% Anyway, if you read that book, like, 01:12:54.700 --> 01:12:56.020 align:middle line:90% what's your overwhelming thing? 01:12:56.020 --> 01:12:57.970 align:middle line:84% So this is, you know, "telling the stories 01:12:57.970 --> 01:13:00.178 align:middle line:84% of these children who became notable for eccentricity 01:13:00.178 --> 01:13:00.890 align:middle line:90% and philanthropy. 01:13:00.890 --> 01:13:03.940 align:middle line:84% Green focuses on Ned Warren, a homosexual and mover 01:13:03.940 --> 01:13:06.010 align:middle line:84% in the international movement of aestheticism, 01:13:06.010 --> 01:13:08.677 align:middle line:84% who was determined to lead a 'grand but blighted life.' 01:13:08.677 --> 01:13:11.260 align:middle line:84% This somewhat jumbled tale of a family sundering through greed 01:13:11.260 --> 01:13:12.350 align:middle line:90% and suicide." 01:13:12.350 --> 01:13:14.650 align:middle line:84% You know, basically I read this and I'm like, 01:13:14.650 --> 01:13:16.140 align:middle line:90% I don't want to read that book. 01:13:16.140 --> 01:13:18.560 align:middle line:84% I mean, I don't know so much about what aestheticism was, 01:13:18.560 --> 01:13:21.670 align:middle line:84% but how much do I want to really read about greed and suicide 01:13:21.670 --> 01:13:22.920 align:middle line:90% in Boston in the 19th century? 01:13:22.920 --> 01:13:25.860 align:middle line:84% And the whole point is, in fact, nobody wants to read that book. 01:13:25.860 --> 01:13:28.260 align:middle line:84% So who is the high value purchaser of that book? 01:13:28.260 --> 01:13:31.040 align:middle line:84% A tourist who's just been to Beacon Hill and seen Mount 01:13:31.040 --> 01:13:35.360 align:middle line:84% Vernon Street and finds a copy of a book about a family 01:13:35.360 --> 01:13:37.820 align:middle line:84% on Mount Vernon Street in the 19th century and thinks, wow, 01:13:37.820 --> 01:13:39.660 align:middle line:84% this will be great to read in my hotel room. 01:13:39.660 --> 01:13:41.480 align:middle line:84% And then you get partway through and get the suicide side, 01:13:41.480 --> 01:13:43.522 align:middle line:84% and you're like, this is not my vacation reading. 01:13:43.522 --> 01:13:45.290 align:middle line:84% And then you leave it there and then 01:13:45.290 --> 01:13:47.300 align:middle line:84% the hotel turns it back in and then sells it 01:13:47.300 --> 01:13:49.133 align:middle line:84% to the used bookstore, and then they sell it 01:13:49.133 --> 01:13:51.380 align:middle line:84% again to the next tourist or something. 01:13:51.380 --> 01:13:53.270 align:middle line:84% But anyway, but so for these books, 01:13:53.270 --> 01:13:55.580 align:middle line:84% like, the online price is not really much higher 01:13:55.580 --> 01:13:59.120 align:middle line:84% because no one is out there searching for that book. 01:13:59.120 --> 01:14:03.980 align:middle line:84% But it is a good match to the local physical bookstore. 01:14:03.980 --> 01:14:06.980 align:middle line:84% Anyway, other predictions, the online-offline gap 01:14:06.980 --> 01:14:09.080 align:middle line:84% should also be smaller for popular titles 01:14:09.080 --> 01:14:11.960 align:middle line:84% because there's a thinner upper tail for popular titles. 01:14:11.960 --> 01:14:14.100 align:middle line:84% There's not the match quality issue. 01:14:14.100 --> 01:14:17.240 align:middle line:84% And then Amazon's Buy Used feature 01:14:17.240 --> 01:14:19.790 align:middle line:84% should lead to more low prices in 2012 01:14:19.790 --> 01:14:23.330 align:middle line:84% because it increases that competition for the lower tail. 01:14:23.330 --> 01:14:26.010 align:middle line:84% Anyway, so those are the predictions we try to test. 01:14:26.010 --> 01:14:28.670 align:middle line:84% And let me just say basically the story is, 01:14:28.670 --> 01:14:31.100 align:middle line:84% almost all of these things seem to be true in the data. 01:14:31.100 --> 01:14:33.470 align:middle line:84% In each of these figures, the left 01:14:33.470 --> 01:14:37.100 align:middle line:84% is the distribution of offline prices. 01:14:37.100 --> 01:14:41.060 align:middle line:84% The right side is an estimate of the same object, an estimate 01:14:41.060 --> 01:14:44.480 align:middle line:84% of the average across titles of the distribution of prices 01:14:44.480 --> 01:14:46.160 align:middle line:90% for that title. 01:14:46.160 --> 01:14:50.120 align:middle line:84% So these should be estimates of the distribution 01:14:50.120 --> 01:14:52.100 align:middle line:84% of offline prices for these books 01:14:52.100 --> 01:14:54.480 align:middle line:84% and the distribution of online prices for these books. 01:14:54.480 --> 01:14:56.090 align:middle line:84% And if you see for the standard books, 01:14:56.090 --> 01:14:59.930 align:middle line:84% there's just this massively higher upper tail of prices 01:14:59.930 --> 01:15:04.600 align:middle line:90% online than there was offline. 01:15:04.600 --> 01:15:06.730 align:middle line:84% If you go to the local interest books, 01:15:06.730 --> 01:15:11.140 align:middle line:84% the local interest books already had something of a tail 01:15:11.140 --> 01:15:13.220 align:middle line:84% online because the ones that are well-matched. 01:15:13.220 --> 01:15:15.670 align:middle line:84% That book that finds its way to the Beacon Hill bookstore 01:15:15.670 --> 01:15:18.220 align:middle line:84% is already well-matched, can charge a very high price 01:15:18.220 --> 01:15:19.310 align:middle line:90% offline. 01:15:19.310 --> 01:15:21.830 align:middle line:90% And so the offline-- 01:15:21.830 --> 01:15:24.410 align:middle line:84% you know, more of an upper tail offline. 01:15:24.410 --> 01:15:27.170 align:middle line:84% The online-offline gap is smaller for those local interest 01:15:27.170 --> 01:15:28.820 align:middle line:90% books. 01:15:28.820 --> 01:15:33.200 align:middle line:84% Popular ones also works, but we skip that. 01:15:33.200 --> 01:15:36.690 align:middle line:90% This is a 2009-2012 comparison. 01:15:36.690 --> 01:15:42.800 align:middle line:84% 2009-2012 comparison, the gray bars are the 2009 price 01:15:42.800 --> 01:15:43.650 align:middle line:90% distribution. 01:15:43.650 --> 01:15:47.880 align:middle line:84% The white outlined bars for the 2012 distribution. 01:15:47.880 --> 01:15:51.020 align:middle line:84% And you see these distributions almost 01:15:51.020 --> 01:15:53.210 align:middle line:84% lie exactly on top of each other, 01:15:53.210 --> 01:15:56.360 align:middle line:84% with the exception that here on the standard titles, 01:15:56.360 --> 01:15:59.840 align:middle line:84% the number of standard titles sold at a very low price just 01:15:59.840 --> 01:16:04.520 align:middle line:84% shoots up in 2012 after the Amazon acquisition. 01:16:04.520 --> 01:16:06.990 align:middle line:84% And so this is if there are many more people searching. 01:16:06.990 --> 01:16:09.500 align:middle line:84% And so that left tail just gets pushed down, 01:16:09.500 --> 01:16:11.910 align:middle line:84% but it doesn't affect the upper tail at all. 01:16:11.910 --> 01:16:14.270 align:middle line:84% And you see that in all three categories of books 01:16:14.270 --> 01:16:17.583 align:middle line:84% that the lower tail gets much thicker. 01:16:17.583 --> 01:16:20.000 align:middle line:84% And it just seems to be drawing from these slightly higher 01:16:20.000 --> 01:16:24.780 align:middle line:84% prices, as if it's competition is pulling down the bottom end. 01:16:24.780 --> 01:16:28.200 align:middle line:84% And then, you know, the paper also 01:16:28.200 --> 01:16:30.810 align:middle line:84% then develops a structural model to estimate 01:16:30.810 --> 01:16:34.320 align:middle line:84% what are the welfare gains from selling books online. 01:16:34.320 --> 01:16:37.710 align:middle line:84% And our answer is, there seems to be a substantial welfare 01:16:37.710 --> 01:16:41.530 align:middle line:84% gain, but even in this used books case, it's not enormous. 01:16:41.530 --> 01:16:43.380 align:middle line:84% So we're more with Quan and Williams 01:16:43.380 --> 01:16:45.660 align:middle line:84% than we are with Brynjolfsson, Hu, and Smith 01:16:45.660 --> 01:16:47.520 align:middle line:84% in terms of what are the welfare gains. 01:16:47.520 --> 01:16:54.960 align:middle line:84% What we find is that the estimated social welfare 01:16:54.960 --> 01:16:59.100 align:middle line:84% from selling one of these 2009 books in the offline world 01:16:59.100 --> 01:17:02.100 align:middle line:90% was $7.76. 01:17:02.100 --> 01:17:07.720 align:middle line:84% Estimated social welfare from selling it online was $16.87. 01:17:07.720 --> 01:17:10.680 align:middle line:84% So it's a more than doubling of welfare 01:17:10.680 --> 01:17:14.970 align:middle line:84% that comes from selling online, but it's only 01:17:14.970 --> 01:17:16.390 align:middle line:90% a more than doubling of welfare. 01:17:16.390 --> 01:17:19.980 align:middle line:84% It's not a massive, not a 10 times increase in welfare 01:17:19.980 --> 01:17:21.730 align:middle line:84% from selling these things online. 01:17:21.730 --> 01:17:24.665 align:middle line:84% And if you look at the firm profits, 01:17:24.665 --> 01:17:26.790 align:middle line:84% the profit is coming from two different dimensions. 01:17:26.790 --> 01:17:31.370 align:middle line:84% It's like the average price is going from $12 to $16 01:17:31.370 --> 01:17:35.700 align:middle line:84% when they're sold online and the books are being sold sooner. 01:17:35.700 --> 01:17:38.450 align:middle line:84% So instead of the e to the minus rt being 0.3, 01:17:38.450 --> 01:17:41.150 align:middle line:90% the e to the minus rt is 0.48. 01:17:41.150 --> 01:17:44.750 align:middle line:84% So the books are getting sold for higher prices 01:17:44.750 --> 01:17:46.590 align:middle line:90% and they're getting sold sooner. 01:17:46.590 --> 01:17:50.000 align:middle line:84% Those two things both make profit more than double. 01:17:50.000 --> 01:17:53.660 align:middle line:84% We're also estimating that the consumer surplus, as best we 01:17:53.660 --> 01:17:56.870 align:middle line:90% can, is also more than doubling. 01:17:56.870 --> 01:17:58.430 align:middle line:84% And the consumer surplus is more than 01:17:58.430 --> 01:18:02.477 align:middle line:84% doubling because the people who are willing to buy $16 books 01:18:02.477 --> 01:18:05.060 align:middle line:84% get more utility than people who are willing to buy $12 books, 01:18:05.060 --> 01:18:06.330 align:middle line:90% and they're getting them sooner. 01:18:06.330 --> 01:18:07.872 align:middle line:84% So the consumer surplus is discounted 01:18:07.872 --> 01:18:09.740 align:middle line:90% by this instead of that. 01:18:09.740 --> 01:18:10.240 align:middle line:90% OK. 01:18:10.240 --> 01:18:16.036 align:middle line:90% 01:18:16.036 --> 01:18:17.200 align:middle line:90% Sorry, I'm out of time. 01:18:17.200 --> 01:18:18.930 align:middle line:84% So [INAUDIBLE] you say, that's what 01:18:18.930 --> 01:18:20.190 align:middle line:90% I've got on retail for today. 01:18:20.190 --> 01:18:21.950 align:middle line:90% And then-- 01:18:21.950 --> 01:18:34.000 align:middle line:90%