 Most of you, but I'm Jacques Cremere, and behind me you can see the wonderful new TSC building where I hope we can see all of you pretty soon. This seminar is actually the initiative of Julian Wright and André, who is going to be speaking. And the TSC is very happy to provide the physical hosting, but Julian and André really launched it and they're responsible. They are co-chairing the scientific committee and they're responsible for everything that goes wrong. The seminar will meet once every two weeks up to the summer and we still have to decide what we are doing afterwards. We are thinking about once every three weeks after the summer if things become more normal, but we will decide and we hope that we will have feedback from you. I won't go very long. The program is in construction, is on the website. If you've got questions which are of scientific nature, you write to André or Julian who will share them with the scientific committee. There are other things which are administrative like the mailing list and so on. You discussed with Marie-Hélène Dufour, whose address is on the website with me. And please don't hesitate to send us any comments about things which went badly. We also like to know about things which went well. For every session we will have both a speaker and an administrator of a session. And Alex, the corner is playing the role today, so I'll leave Alex to explain the rules of the seminar and so on. Now you won't hear about me again. Hi everyone, I'm Alex, just a couple of things that I want to warn you about. The first thing is that this seminar will be recorded. The way it's going to work, I guess it's still experimenting, but I'm the moderator. So if you have any questions, I guess you can either use the chat and then I'll ask the questions when relevant. I mean, I'll do my best or you can raise your hand and from time to time I'll give you the floor. So if that's okay, so we're very happy to have André as you for the first seminar today, talking about data-enabled learning network effects and competitive advantage. André, can you share your screen? Absolutely, and can everyone see it? Yes. Alex, should I get started? Yes. Okay, well, good morning, good afternoon, good evening, whatever the case may be to everyone, which is a very interesting format. As Alex mentioned, so what I'm going to do, I'll try to start to pause maybe like every two or three slides and just to give some time if people have clarifying questions during the presentation. As a reminder, the idea is 40 minutes presentation, of course, with some interruptions for questions, but then and we're hoping to have the last 20 minutes for more substantive Q&A. All right, so this is a paper with a joint with Julian Wright, who's also somewhere in the background and it's called Data-Enabled Learning Network Effects and Competitive Advantage. The background for this is relatively straightforward. We all know that there's an increasing number of products and services these days that are gathering customer data and they're relying on that data to improve those products and services. The underlying reasons are pretty clear. Again, there's lots, so there's more and more digitization, even of traditional consumer products and services. So of course, this is true for software, but it's also true for hardware products and even more traditional like Apparel and other types of products, all of which now are connected in some shape or form through cloud-based services and that enables the providers of these products to gather customer data and use that customer data to improve the products. And the second factor that makes this possible, obviously, is cheaper storage and better algorithms, artificial intelligence and machine learning that enabled that processing of that data to actually translate into improvements. So as a consequence, and I think that those are sort of the key things that matter for what I'm going to talk about is that the learning from customer data has now become a key in continuously improving the willingness to pay for these products and services. The second aspect is that the learning from customer data has become very fast. I mean, learning from customer data is not necessarily new. It's been around for decades, but it used to be very slow. What's kind of interesting now is that the cycles of learning and product improvements are very fast to the point where the product or service has actually improved during the consumption lifetime of customers. If I adopt a product, some of these products actually expect it to be improving during the time that I'm going to be consuming it. And really, because the data that is being gathered these days is can be so fine grained. The other possibilities, obviously that the products are actually highly customized and the improvements to the products and services are customized to each individual user, as opposed to just generic improvements for everyone. There's lots of examples of this I'm not going to go in through in detail in all of them. There are many more like than what I listed where we listed in the slide. The only thing that I want to point out is that so in some of these cases, most of the learning is specific to one user. So if you take something like say smart connected devices like thermostats fridges and so on. What really matters there for product improvement, the data that really matters is for each individual user their own personal data so product gets better the more I use it but it doesn't really matter how many other products they're serving. In other cases, like say recommend recommender systems, what really matters is learning across users. And of course, for a lot of other cases, both of these play a role so there's both a cross user learning and within user learning. So I want to flag this out upfront because this is something that all that we go in quite some depth in the model. So key questions we're asking. What is competitive advantage look like with data enable learning when firms are competing in context with data enable learning. What are the differences in outcomes when the learning is across users versus within users. And then finally, what are the conditions under which this type of data enable learning actually leads to meaningful network effects and I put meaningful in quotation marks because I mean some people have the view that data enable learning itself is a form of network effect. I don't have strong because basically it has a positive feedback loop, the more the more customers you serve the better the product. I don't have strong feelings I mean I tend to have a stricter definition of network effects, but be that that is it may the key message that I'll get to in the last part of the talk is that that itself is not enough for these network effects to matter and the need for them to matter is some sort of consumer customer coordination problem, just like we have with traditional network effects. So I'll talk about that in the last part. This might be maybe let me pause here for a second and see if they're Alex if they're any any burning questions. Alex, we can hear you you're on mute. No, sorry, no question. Great audience as well behaved. I was fully expecting someone would ask like what is your model. All right, so in terms of relationship with previous literature I'll go relatively quickly through this, obviously it's related to traditional learning by doing literature, which mostly tend to focus on improve so cost reductions due to serving more customers what we're looking at is willing to willingness to pay improvements, but that's not all right I mean there's a similarity at a high level between the two. And I guess that similarity holds in the simplest version of our model but we do have like quite a few new topics that we can explore and new results. What I would emphasize here is that one nice feature of our model is that we're able to fully characterize price dynamics, firm profits. And we actually can get close form solutions, and we can do some nice comparative statics in the shape of the learning functions and that's in addition to exploring some topics which are not traditionally explored by learning by doing literature. We obviously borrow so it's somewhat related to literature and network effects we borrow like the notion of consumer beliefs and coordination issues. Switching costs in the sense that when you have with within user learning in our model that creates an endogenous switching costs. There's also a number of actually recent papers that do look, you know, more or less specifically data for different forms of data enable learning. Again, I won't go into detail. I think the key to this, the key differentiation of our paper, I think, as far as we know, it's the first one, first one to vary it to have a very explicit dynamic model in which both firm choices or firm strategic choices like pricing and consumer decisions are explicitly modeled. I mean, there's some topics that these papers study that we don't and vice versa, there's certain topics that we look into like distinction between within user learning and across user learning that are better, we think are not. Again, if no questions, then I will go into I will go into the model. So I'll start with the model with a cross user learning only. There's two firms in common and entrance, and they're competing in prices over infinitely many periods. There's a discount factor firms, the two firms have the same marginal costs, and again they compete in prices they set prices every period. There's a continuum of atomistic identical consumers of measure one in every period. And you can think of these consumers as being either the same consumers in every period, but they can costlessly switch so they can make a certain choice and let's say period T, and then a period T plus one if they want they can switch from one firm to the other, but with no exogenous switching cost, or they can also be new consumers in every period. Either way works and the solutions are the same in this model with the cross user learning things are going to be different with with when we go to within user learning. And again, every period the consumers decide which firm to buy from. And so the key feature obviously of the model is how we model utility derived by consumers and we're trying to make it as simple as possible. The model offered by firm I is made up of two components. There's SI, which you can think of as a standalone value of the product or service for firm I. And then there's the learning component and the learning component takes the form of an increasing function of and I, which is the total number of consumers that firm firm I has served in the past. There's a couple of couple notes here. So, because there's a measure and equal one of consumers in every period and because we assume Bertrand competition every period in every period, one firm is going to win all consumers. So therefore the number of past customers that each individual firm wins is exactly equal to the number of past periods that each firm has won in the past. The other note is we're going to assume that the learning function is capped at an I bar what that means is basically there's a certain number past a certain number of total customers that have been served. There's a there's a maximum learning threshold that is being reached and the firm I cannot pass that threshold. So basically, you know, the learning plateaus you've learned everything you can you can learn and then the utility stays constant after that. So you can actually like relax this assumption as I'll mention in a couple of slides. So just to just to just to emphasize we don't actually we don't place any restrictions on the on the learning functions fi other than the fact that they're weekly increasing. So they can be, you know, something, something like this where, you know, which is kind of intuitive. So let's say there's in this type of learning function there's not there's no learning up to a certain point then learning is increasing and finally learning completely plateaus past the threshold and I bar. Alright, so in order to determine the equilibrium in this model. Obviously, we're going to define the value functions, which means the present discounted value present discounted value of profits for each individual firm starting given the n I and any being the starting points and I and any recall are the number. Are the number of previous periods or previous customers that each individual each firm I or E has has served in the past. So I'll show you just the easiest possible case before actually like showing you the full results. So imagine that both firms have reached their learning thresholds I both firms have have served enough customers in the past to learn everything that there was to learn. Well, in this case, things are very simple because now we basically have asymmetric Bertrand competition over infinite periods with just different utilities so each firm's utility is si plus fi of the of the n I bar which is the threshold. So obviously the firm that offers a bigger you to higher utilities going to win in every single periods. And then the value functions have the expression which is shown here which is again just Bertrand profits in one periods over infinite many periods over one minus delta. Now, so basically this is the starting points for the proof of the for the derivation of the entire equilibrium which we do using backwards induction and in fact it's going to be two dimensional induction in the n I and any. So now if we go backwards so if n I and any are smaller than n I bar and any bar, basically the way this works is that the key step I guess in the proof is that each firm is willing to offer a subsidy in the current periods, which is going to be the difference between its future value, if it wins that current period versus if it loses. And using that you know you can we can iterate and use a again backwards induction into dimensions to obtain the following proposition. So this is the first key result in the paper. So let me just explain briefly what what it means. So essentially there's a unique Markov perfect equilibrium. In which firm e wins in all periods, if it's standalone value is relative to firm I is greater than a certain threshold and the threshold is this function delta of an I and any, which is given here. So first of all it's not. I mean some sense it should be pretty intuitive that there's a threshold because again the learning functions are increasing so there's this like positive feedback loop the more you win the more likely you are to win in this period obviously you're going to win in next period because you just accumulate more learning and the opponent has stayed remain stuck. So it's not surprising that there's a threshold. What's interesting is the expression of the threshold so delta of an I and any. So you can think of this threshold as capturing firm eyes competitive advantage. Why because the higher this threshold is the higher the bar in some sense for the firm e to win in terms of in terms of its standalone utility so if delta is higher. The standalone utility of e has to be higher relative to the standalone utility of I in order for you to win. And not surprisingly, delta of any is increasing and I and decreasing and and so the more learning from I has had in the past, the higher its competitive advantage, and the more learning from he has had in the past, the lower the competitive advantage of firm I and the other key thing to notice here if you divide delta of any by one minus delta. If you look at the expression it's actually relatively simple. It's basically the difference in present discounted values of the total surplus created by I if I was to win in every period versus e. So if I was to win every period and the press discounted value is going to be f of fi of Ni this period plus delta of delta fi of Ni plus one plus so on and so forth up to infinity and just remembering that the learning caps at Ni bar. And then you basically take this difference between I and e and you obtain the threshold, which by the way means the threats threshold is socially efficient something that will come back and talk about in a couple slides. So last remark here. Again, we've done this and so this equilibrium is unique. If you rule out strategy so pricing strategies for both firm that are dominated off the equilibrium path, and we also have a tie breaking rule. So in that sense the equilibrium is unique. Now you can take the results in this proposition and also the expression that the value functions for both of the firm that I'm going to show you in the next slides. Everything holds just as well. If we take the limit when the thresholds go to infinity. So we don't necessarily need the We don't need the learning functions to be bounded. All we actually need is that they're bounded by some power function in the limit. So for example, linearly increasing learning functions works just as fine as you know, our results will go through. Let me pause here in case there are questions. Are there any questions. Not seeing any right hand. Let me know in the chat if I'm just. Yeah, take a look at the chat. Again, I'm happy to pause. I think this is the first time we're doing this. So we're also trying to figure out what the best format is here. Okay, we have a question. Why do you need the cap on and So just okay so the key the key the cap on and is key in the proof. So in the proof of proposition one we against the proof is by backwards induction. So we start with the The result is very easy when both firms have reached threshold and we work backwards from there. So if we didn't, if we didn't have the cap on and doing the proof like directly when the learning functions or is unbounded would be a little bit more difficult. Also, the one thing that doesn't necessarily go through when you go from bounded. So bounded learning to unbounded is the uniqueness of the equilibrium. So if the two learning functions have this cap then what I proposition one says basically the equilibrium is unique. Once we go to infinite. I mean, the equilibrium is still there and works very well, but it's not necessarily guaranteed that it's unique. So it's mostly like a technical technical issue, but also I also think it's quite realistic in general, you know, learning, learning is kept. Okay, so how important is linearity. There's no linearity assumed here. I'm not so maybe you can ask the person to clarify there's we haven't really assumed any linearity, or maybe I'm misunderstanding. Gary, you want to let me unmute you. I'll mute you Gary if you want to clarify. Just in the learning function it's linear. It's si plus. Oh, I see. Alright, I think you could do it becomes obviously a little bit a little bit trickier. I mean presumably you can do something more general as long as so you can have like a function of si and of the ni and just imposing that the it's increasing in both terms. I think the interaction between the two in the function might make it slightly more complicated. I think at a high level the main results will probably still be there. It's just neat. I guess that the other part is like it's just neater to do it this way because you know you have this very natural threshold right so we focus at the other parts like it allows us to focus very cleanly on the learning functions right express everything as a threshold in terms of the standalone utilities relative to the to the learning functions. Regarding the questions. I'll prioritize the questions that are more like clarifying questions and you know I see some questions that are more like discussion of related literature so I think we should keep that for later. Okay, I trust you. So that's that that's great. Okay, so let me go. So there's one corollary to this result again we can express we can get close form solutions for absolutely everything. I'm not going to show you the actual solutions in the actual expressions for the value functions for the two firms. We just want to focus on something that's kind of interesting. So there's basically four regions, depending on where the difference in in se minus si so difference in standalone values is so because he and I are pretty much symmetric up to now. So it's either I wins or he wins. So if you focus on the first two bullet points which are the cases in which I wins. So there's two cases there. If si minus se is very low. So if he's standalone values sufficiently low relative to I that basically I wins and he doesn't bother to compete very hard so he's going to price of costs. And, you know, and in this case so that's that's the first case now in the second case when I wins, but he is sufficiently close and standalone value. Then he will actually offer a subsidy in the current period because it hopes to win right if he wins the current period then actually it will win the next in the next period. It makes perfect sense that the the threshold for that to happen is delta of any plus one so if you were just to win one more period, then it could potentially win and then symmetrically for symmetrically when he wins. So we know we can push this further you can basically determine the number of periods. For which the losing firm will will subsidize so if if the standalone value of the losing firm is very low then the losing firm is never going to is never going to offer a subsidy so it prices it's cost indefinitely. If at the opposite extreme if the losing firm standalone values actually really close that there's this range of parameters that's, you know, that's, that's not empty, over which basically the losing firms offers a subsidy indefinitely. So it keeps hoping to win indefinitely even if that doesn't that never happens. And then in between these two cases there's a range in which the losing firm offers a subsidy for a finite number of periods which we can actually pin down using the using the expressions of those thresholds. Now perhaps a little bit more surprisingly, the winning firm is actually it is possible that the winning firm may also subsidize. So for example, if the winning firm. So if the winning firm is I. But in the current period, the total value offered by I so standalone value plus learning is lower than this than the current value offered by E, then actually I is going to have to subsidize at least in the current period, even though eventually of course it wins and ends up extracting positive profits. So let me talk about welfare. The key thing here. So the interesting results is that as I mentioned, given the expression of the thresholds of the delta, basically the unique, the unique Markov perfect equilibrium that I've shown you is socially optimal. Because that delta is exactly the difference in present. So it's proportional to the difference in present discounted values of the total surplus offered by each firm on its winning path. Now, at a high level, the intuition would be as follows. Well, consumers are short sided, they're not really short side. I mean, the issues like again here consumers don't have to look beyond the current period so they only care about because they can either switch or they, you know, this new consumers every period. But the firms are willing to subsidize right so in every in the current period, each firm is willing to offer as a subsidy, the present discounted value of its future profits, if it wins versus if it loses. So at some high level, there's this intuition. Well, if a firm can offer a higher PDV of total surplus, then maybe it should win. The thing is, it's actually not so obvious. So the social efficiency results in this proposition, it's actually relatively subtle and not, you know, not immediately clear. The reason is the winning firm because it's competing against against an opponent was willing to subsidize. Well, obviously the winning firm is not going to extract the full present discounted value of the surplus that it creates. And actually to see this very clearly. If we take a finite period version of our model. The threshold with finite periods is in general not socially optimal. Now this threshold when you know when we take the number of periods and make it go to infinity converges and everything's consistent converges to the threshold that I've shown you earlier. However, what's interesting about it, the threshold is not efficient, but the inefficiency in the threshold goes away as the number of periods goes to infinity. So it's somewhat, you know, it's somewhat interesting, somewhat, somewhat need that this happens and not, you know, not immediately obvious. And then finally, a final observation here related to welfare is that because of Bertrand competition, obviously the consumers are left with the surplus with a total surplus that is offered by the losing firm. Because of that, whenever the winning firm keeps keeps learning so the winning firm keeps winning in every period and increases its learning. Not only that doesn't does not help consumers, but it is quite possible that actually hurts consumers. And the reason it can hurt consumers is that learning by the winning firm actually in some sense discourages the losing firm so it's offering either less. So it's decreasing it's the subsidy it's offering consumers or actually it stops offering a subsidy altogether. So as a result, it becomes less less of a competitive threat to the winning firm and therefore consumer surplus suffers as the winning firm keeps winning. Now, what that suggests is basically there might be scope here for something like a data sharing policy in order to help consumers, in which basically what you may want what we kind of like what we want to do intuitively is. To say, well, I want to strengthen the competitive position of the losing firm or the firm that's behind. Maybe we should require the winning firm the firm that's in front to share its data or it's learning with the losing firm. The idea would be well if we do that then it keeps competitive pressure on the on the winning firm and you know it should it should lead to higher consumer surplus. The problem with that is that there is a obviously there that can be a countervailing force to this and the issue is the following if the if the losing firm is E and E is behind. If he anticipates that it's going to that even if it loses there's going to be some data sharing from I well it kind of reduces ease incentive to fight very hard and in our model this translates into offering a lower subsidy so he basically doesn't need to compete as hard because it knows. It might actually benefit from data sharing. So it's a bit of a it's a form of free writing. Now we can we actually show what we do in the paper. I'm not going to show you here the details in the interest of time but we can show the straight up formally and it comes up very neatly in the model. So basically we can do if we take identical learning functions for both firms. So there's if I and if you are the same and we assume that I has already reached its maximum learning threshold and he is somewhere behind. And we're going to say data sharing means that if he remains behind the end of the current period that there's going to be full data sharing in the sense that I is going to be forced to share all his data to eat so he catches up no matter what at the end of the current period. Well, that's like the simplest simplest version of data sharing that we can look at and we can still generate this trade off. So basically what happens is in this framework he no longer has any incentive to offer a subsidy in the in the presence of data sharing because it actually doesn't matter if it wins or loses it will pull even with I. So as a result the present discounted value of consumer surplus can actually end up being lower with data sharing because they no longer benefit from the subsidy that actually creates competitive pressure on I. It's not true for all parameter values but you know we show that on the range of parameter values this this can happen. So another effect that could happen and I'm wondering if it doesn't seem to happen in your model but usually when you think of mandated sharing you'd expect also that the firm would have less incentive to collect data so here I would have less incentive to collect data does that happen. So I think so I think that's a fair but I think Jacques report mentions that right you have some like at some point with data sharing there's something there. I don't we don't model data collection so in our case I think I mean I guess you could interpret the so it could be like efforts to efforts to collect data or like investments like you know in a very. Even with pricing you could imagine that you would have an incentive to offer lower price in order to generate more data but doesn't seem to be the case in your model. Well so actually anything to do with prices I think it's captured in our model right because that kind of takes so that's taken into everything that has to that goes through prices is taken into account in this framework right so I mean we do take into account that of course like if I lower my price I just get more data and accelerate my learning so that's definitely there. We don't model as an investment but I think so this result the fact that there's a trade off I think it's still true right I mean you can interpret it as meaning I free ride because in our case it's free riding by no longer offering low enough subsidy. If you know if the relevant variable is investment efforts or investments and connecting data sharing then you could interpret that way as well so I think it's pretty I mean that conclusion is that there's a trade off there's a very clear trade off with data sharing. Should be quite robust like we show it in very very simple way but I think that's. That's fair. Okay. Anything else. No. Alright so I have eight minutes I'm going to be like shit I think it'll be fine but let me speed up a little bit so this was all you can do you can do a bit more because it starts late. Okay no it's okay. Alright so this was up to here everything was across user learning so now I'm going to briefly show you the result for within user learning. So the model within user learning is very similar so there's two firms same thing competing over infinitely many periods. Now the consumers because with for within user learning obviously we need the consumers to be the same right so consumers. You'll see why actually in a second so the learning functions are also going to be the same except that they mean something very different. So in this case the N I the arguments for for F I is no longer the number of consumers that a firm has served but the number of times. A given so a given consumer has purchased from firm I in the past so again within user learning this is focusing entirely on the product gets better for me the more I use this product. So it's no law there's no cross user learning here so we're just focusing on within user learning. And the other key assumption we're going to make is that firms can price discriminate based on consumers individual history which basically means we can just focus here the model can focus on one representative consumer. But everything else is the same so the learning functions are the same again continuum of N equal one consumers in every period and the consumers are infinitely lived and then learning functions have the same expressions and same properties. And interestingly enough so the key reason the main first main result is that well same thing we get a markup perfect equilibrium in each in which he wins. If the difference in standalone utilities is greater than a certain threshold and the threshold is actually the exact same threshold as the one for across user learning, which obviously also means the threshold that threshold is remain socially optimal here. So one observation. It is true that the threshold is the same however here it's actually a lot easier to understand why the threshold is socially optimal because the consumers are the same in every period and because they care about within user learning. They basically they're forward looking and they think about well if I you know if I consume from this firm this period. How is that going to change my options in the next period. So it's very straightforward to see that consumers are going to look at the look for the firm that offers them the higher the higher present discounted value of future of willingness to pay which is utility plus the any subsidy that that the firms can offer. So the logic for social optimality is easier. However, the proof like the way this equilibrium works is actually quite is a little bit different and requires a different proof than the one for across user learning. And the key here is like because consumers are forward looking we also need in the derivation of the equilibrium we also need to derive the value function for consumers. So it's not enough to just look at the value functions for the firms. That's just the technical side. And then finally what I want to emphasize here is that for this here with for within user learning the pricing where the subsidies offered by the firms are no longer the main driver. And one way to see that very clearly is that we can actually remove prices here so we can assume that the two firms price at zero or sorry at cost in every period, you would still get the exact same threshold. Again, this has to do with simply the fact that consumers are just going to look for the firm that offers the highest PDV of willingness to pay. That was not true. So with the cross user learning I can't emphasize this more that was definitely not true if you remove pricing and subsidies the equilibrium is going to be very, very different and it's not going to be socially often. All right, so the other proposition that so this is the that matters a lot here. So focus on the actually so let me briefly talk about the first two bullet points so just like in the case with the cross user learning again the threshold is the same and you still have the same possibility of subsidization by the losing firm so that doesn't change at all. Now the interesting question is well because the thresholds are the same, we can actually compare. So the winning firms profits with user with across user learning versus this case with within user learning. And as it turns out the PDV of the winning firms profit is actually always lower with within the user learning. Now to understand this again it comes back to the fact that here consumers are forward looking. And the key difference is the following forward looking consumers with within user learning actually understand that if they join if they joined the winning firm this period that actually puts them. So basically that means they have a higher switching costs in dodging switching costs to switch to the loser the next period which means they're actually the losing firm is in a weaker competitive position next period which means the winning firm can actually extract more. So as a result the winning firm must compensate these forward looking consumers for the weakening of the competitor the next period. This was not true with the cross user learning because again with the cross user learning the consumers had no reason to look past the current period and making their choices. So basically as a result you get more intense competition when learning is within users than across users. And obviously that means consumers are better off with within user learning and counterpart to that is that if we look at the same if we do the same data sharing exercise that we did with the cross user learning. When we have within user learning data sharing is actually more likely to lower consumer surplus because you already it's already pretty competitive so like actually forcing data sharing is more likely to harm consumer surplus than not here. Alright so the next so this is the third and last last part of the paper we're looking at we're going to focus on network effects and the role of beliefs. So key observation here is that up to now network effects have played absolutely no role in the equilibrium analysis. Again some people may think well data enable learning because of this positive feedback loop is a form of network effects again I don't have a strong feeling like feeling about this. But even if you believe that well up to here even if say like maybe there was a network effect that network effect actually actually had no material implications there was no consumer coordination problem. And the reason is any in both of them both the model with the cross user learning and within user learning. There's no reason for any consumer in making their decision to actually look at what other consumers are doing in the same in the same period. Basically all that matters for every consumer was well which firm is offering the highest to hire you either PDV higher present utility or PDV or future utility in this period. Now we can actually generate so we can actually generate network effects that actually lead to consumer coordination problems in this model in two ways. And I think this is interesting it's an interesting exercise because it clarifies it makes it very clear why the conditions under which this kind of data enable learning actually leads to coordination issues and network effects that have you know substantive implications. So the two ways are as follows the first thing we can do is so obviously you need in order to get network effects you do need a cross user learning there's no chance. We're going to get any network effect if the learning is solely within users right because then there's never any chance that a user will care about about how many other users the firm is selling so you have to have a cross user learning. But again that by itself is not enough. So we can add one of two things we can add learning within the consumption lifetime so as I mentioned at the beginning a key sort of feature of data enable learning nowadays is that products improve so fast that when consumers adopt certain products they actually fully expect the product to improve within you know their consumption lifetime. So we can model this in two ways. I'm not going to show you the models here they're there in the paper. One way is to basically say there's a cross user learning and then there's also within period learning so we just basically add one period of learning. So the utility for a consumer in the current period is depends on how many how many consumers we served in the past but also depends on how many consumers are being served in the current period that obviously will lead to network effects with coordination issues. And the second one is a little bit different across user learning and now we're going to assume that consumers pay once and then they can enjoy utility from the product that they've chosen over multiple periods. And in this case basically the price adds acts as an endogenous switching cost and again you get coordination issues and proper network effects. And then the second way which we certainly think is more interesting one to generate network effects with coordination issues is that you can combine across user learning with within user learning. And that's it so just a bit two basic models that I've shown you combine them in the same model and there what you get is endogenous switching costs you to within user learning and then combine that with the fact that I have to care about across user learning actually gets a very nice way of generating network effects with coordination issues. So either either no matter which which of these different ways of generating network effects we choose what's interesting is as soon as obviously as soon as you get a consumer coordination problem you get scope for beliefs to matter just like in traditional models with network effects. So what we do in the paper with we contrast parater beliefs where basically everyone can coordinate on the equilibrium that's best for consumers versus beliefs that favor one firm and then we show that there's there can be a distortion from the social optimal. So I'm looking at the time. I since only have one minute so my plan I was going to show you briefly what the model with the cross user and within user learning looks like. Let me do it and maybe like 30 seconds. So it's just it's a relatively interesting. It's relatively simple but relatively interesting formulation. So what we do here is simply instead of have so the learning function has two components now there's a cross user learning that depends on how many other consumers we serve and within user learning depending on how many times we've served an individual consumer. And again I think the key message from this is that you within user learning serve the within user learning part serves to create an endogenous switching cost and then the across user learning actually creates this interdependency across different customers. Let me skip the details but basically the bottom line of this is that you do get a distinction depending on the nature of beliefs so you get a different equilibrium. And then it's the key messages that both types of learning are necessary in order to create this type of endogenous network. And I'm going to conclude very briefly. Again I think this is data learning has become essential to many products and services. And it does have very, you know, material implications for competitive advantages for competitive advantage and competitive outcomes. You know profits are a key I think a key message is profits are higher with the cross user learning versus within user learning and I think this is like a relatively general general implication. There's some again I didn't I didn't talk too much about comparative statics but we do we can actually generate some nice interesting comparative statics using both general but also like some more specific learning functions. And then the other the last key message that I want to leave you with is that favorable user expectation can actually be a source of competitive advantage. If we combine across user learning with either continued product improvement during consumption period or within user learning in action in order to actually for network effects to matter. Otherwise, you know data enable learning maybe a form of network effects but not one that you know has a material implication. So I'm done. And we're putting up with this. Okay, thank you, Andre. So we've got a few questions. So what I'll do is I'll give you the, I'll give you the mic. Okay, so those people who have questions on the chat so I'll start, you know, from the first one. So Doshin, you had a question first. So I'd like to get your mic. Doshin. Oh, yes. So it's a very trivial question. But why does the loser stay in the in the market if there is some very small cost fixed costs to pay to stay in the market then the loser will keep the market immediately. I think that's fair. So we just somehow we assume, right, you're right. So, like, a lot of shortest answer is like we don't we don't assume that we allow the we allow the loser to stay there. And thank you. Is there a more maybe there's a there's a I guess you could justify it if you know maybe there's like they're so there's maybe different market segments so that the competition that we model here is basically valid in one market segment and the loser still stays because it has other active segments that can still sustain itself. J. J. You want to say something. Sure. So my question is, what if there is some some decay in learning. In other words, the more risk that there is the more important. So I think that's a very nice extension. So Julie and I were just talking about doing this as an extension of our model right so you're right as of right now we assume that everything accumulates. I think this again the cool thing about the model is quite tractable and I think we can easily accommodate we need to think about the best formulation for this but you can make it either I know some probability that in every period you lose let's say the last period of learning or something like that basically the stock of learning the case right that's what you have in mind. Is that right J. Yeah, that was my question. Yeah, so I think it's especially a mobility data. Yeah, I think that's I think that's I think it's very that's that's that's a good point I think we're going to look at this like pretty much it's one of the next two things we're going to do. Okay, and question by Fiona is consumers are not forward looking then how do the results change some form of imperfection seems like it might be realistic. If, if consumers are not for so again in the first model so you remember with a cross user learning the consumers don't need to be forward looking so it's like actually it could be either new you can interpret it in many ways actually either it's new consumers every period, or it's like infinite to live consumers that don't actually understand what's going to happen but either way actually so with the cross user learning it doesn't matter, because they don't have to be forward looking like the only thing that matters is current period with within user learning, then it actually matters so if they're not forward looking, I guess we can. So what would happen. I mean, I guess you get a yeah we're definitely going to get an inefficiency in the results and in fact so in the model with within user learning if consumers are not forward looking. First of all, we're not going to get the social optimum and presumably going to get the same type of inefficiency that we would have gotten in the model with the cross user learning if there were no prices. I mean that's yeah that's easily easily addressed. Luis Cabral wanted to ask some questions, Luis I'll give you the mic. Right I actually had a couple questions a very small, more detailed question and then a big picture kind of question. Small detail is that could quite understand why if I'm a multi period consumer forward looking consumer I do not care what other consumers choose in the current period. Luis in which so I'm happy to go back to this learning I mean in the cross learning I mean in the internal case that's in the cross learning case I thought that I'd mentioned that there were no network effects within period. I couldn't understand why not if I'm forward looking if I if I'm a repeat consumer or whatever long live right so in this model so let's say with the cross user learning so what's happening your utility depends so it's standalone value, plus the value of learning, the value of past learning. So again here we don't we assume that there's no within period learning so basically all the learning that's relevant for today's customer is just to learn that occurred up until the previous period right. Now it is true that the product can still improve in the future right so if I buy you know I if this firm wins and I keep buying from it I'm going to get higher utility. But the key here is like you don't have to worry about this when you make your choice today because you can costlessly switch. What matters to you like if I'm looking at today today the utility that matters to me so the only learning that matters to me is the learning that occurred in the past I'm going to basically get the utility that is based on learning that occurred in the past whichever firm I choose. Then I then tomorrow I can I can revise my decision and because of that I don't have to worry about what other consumers are doing. I still don't get it. I mean the fact that I can costly switch does not mean that I don't care about different states in the future. Again so it's it's so maybe the other part is like the consumer the consumers are atomistic so like your choice like your individual choice as a consumer here does not affect the outcome for the next period. So whatever I choose whichever firm I choose in this period is not going to affect who wins this period or in the next periods. It really like it really in this case it doesn't matter like you make your choices period you're going to derive the utility which is based on learning from the past. Again you and then comes tomorrow you can revive I can switch firms I can stay with the same firm you just don't have to work like you don't have to worry about that. So basically maybe another way of putting it is your options tomorrow like the utility options you have for tomorrow are not going to be affected by your choice today. I get that what I don't get is that my maybe I got it wrong that my utility does not depend on what a measure of other consumers do today. Because again so this maybe I should have made it clear the learning if you look at the at this bullet right so the learning the and I that matters for the utility today is up until yesterday. No no no I understand that my utility today does not depend on what a measure other consumers do today, but my utility does depend on what a measure of other consumers do today. That's right and you can react to that so the thing is I can remember utility accumulates so that that's the present then that that's actually really simple because so whatever you do today. Like you can basically wait and see you don't have to like guess what other people are going to do today because you can just wait and see what they're so what's going to happen is of course what they do will matter to you. But I don't have to try to guess it I don't have to I don't have to make any like any guesses about that because that doesn't influence my current choice. So this is like a model of network effects with lag network effects lies in. Exactly. Yeah, kind of right. So what's cool about this so again this is why like when I go to actually generate proper network effects with expectations. Well obviously what I can so one easy when I say the last least interesting way is to basically say my utility today depends on learning up until yesterday. But are also going to be affected in real time by how many consumers adopt in the current period. As soon as you do that then you have to form expectations. And I don't want to take more more time but my bigger picture question is how is this really different from a learning, how is data different from just a traditional learning in terms of dynamics and welfare analysis. All right, so I mean a couple things I mean obviously there's some similarities I'm sure like there's probably like a way to sort of like map the so the pure across user learning case the one that I started with. Yeah, that probably maps relatively nicely at a high level with you know with traditional model of learning by doing in which it's costs that decreases as opposed to willingness to pay increasing. Now, as soon as we add learning within a consumption lifetime or within user learning things become very different because this is something that was not there and traditional learning by doing models. The other thing I would say is that, because the way we formulate it in a relative to traditional models I guess relative to especially your paper with with my career then we can also get. So actually get close form solutions we get full characterization of the price dynamics, we can also do I guess in comparative statics on the on the learning functions but more substantively, I think it's you know we can look at a cross user learning versus within user learning in terms of actual beliefs and actual network effects that are there are being introduced in various different ways. Okay, thanks. Okay, now we have a question by Earthland. Okay, thank you. I wanted to ask say you said that firms can price discriminate based on history in the between user learning case. What if they come. Can you repeat that? I didn't get the first part in the within user case between user learning firms can price discriminate based on history. Yeah, what if they can't price discriminate perfectly. So that's a good question so basically the reason we need this in some senses they because they're what happens if you can price discriminate then is conceivable you'll get a pretty complicated might get a more complicated analysis in which some consumers are like, different consumers have bought different number of times from the two firms, you can't rule that out right so you might actually get nonexistence of pure strategy pricing equilibria in every period, which makes things, you know, very very messy. Okay, and obviously you can justify sorry just one as I would say that you can easily justify this right I mean the whole point actually of within user learning of gathering data is that well firms actually can offer. You know can customize the offerings to to different consumers. Okay, and actually in practice they do this they offer they can they can easily say okay well show me that how many times you bought from me and I'll give you a discount or something like that. I mean it's true but you can also imagine that there are some users as firm can do this better than the others for instance some need and the old users at the same time. So, I totally agree so I mean that's honestly I think that's probably like an interesting extension I think you know what they're like you have to you get into like more complicated pricing issues so there may be again because users have different histories. You might get it you might actually have to get into like mixed strategy pricing equilibria. It could be it could be an interesting extension. Next question from. Hi. Yes, my question is more on whether we can translate some of the learnings of these modeling to an empirical test of which type of learning matters more, more across different industries, because your implications are sort of similar for within and across user learning but they're not identical. That's right. So I mean that's it's an interesting thing to think about so empirically I guess what would we I guess let me go to you go to conclusions can think a little bit about like what's what you could test for like in terms of main differences. So I'd say to the main difference to us if you look across across user learning versus within user learning is that competition is fiercer. When there's more within user learning and so the firm should be lower I think that that should be I think that's sort of like the key implication, which presumably I mean you're you're much better than I'm like you might think about like taking this to the data and try to find a way to well to test for that right. Yeah, there's no sort of direct implications of sort of our entry or. So we don't have I mean that's a good question we don't have I mean here it's like. Yeah, you have two firms we don't we don't have an entry there. Okay. Yeah. Thanks. But I think that's by the way that's actually an interesting. So again there are other papers and Alex has a current paper there's a couple other papers they're looking at different aspects of data enable learning. I think it would be kind of interesting to so we haven't seen this distinction like between a cross user and within user learning I we haven't seen this anywhere else I think it's one of the key novelties of this paper. So there are probably ways to combine this with maybe a model of entry right we have multiple firms and there's like entry and exit of firms and see what presumably that that will matter there as well. So that might also lead to other important empirical implications like that. Ben Cassner. Yeah, thanks. So you had this data sharing results that in a lot of circumstances ended up being negative for welfare when you have total data sharing. I was wondering if partial data sharing is more broadly helpful, basically, letting the entrance catch up, but not too much conceivably, but I mean, I think the thing is that the straight up will still remain right, unless you can find the way I think it's obvious that more partial or like less data sharing necessarily means less free writing in the free writing is there. It's a matter of like, which data I mean what you would want is like I want, I want data sharing that minimizes free writing and maximizes the, how to say the competitive pressure that you put on the leading firm on the winning firm in order to have the highest chance of making sure that doesn't harm consumer welfare. I'm not sure that it's just, it's just a matter of making the data sharing partial like free writing is necessarily going to happen. So there's another question on Cabral and Riordan, but I guess this has already been discussed. So, the last question is by Jacques, so that's a way to close the loop. Jacques, can you unmute yourself? Jacques, you promised you would not talk again. Jacques, you're still here? Yeah, I couldn't have done that. This is really interesting. One of the big issues in policy is the question of multi-product firms in the case of data. People are very worried about, and we speak quite a bit about it in our report. I mean, one could see that we could extend your model in order to take into account a cross product learning. Do you have any idea what would happen? In particular, I wonder if your results on the effects on competition wouldn't be at least partially reversed. Which ones do you mean? The comparison between within user and across user learning? Yeah, the fact that letting data sharing would reduce the intensity of competition. Oh, I see what you're saying. They say in one product that's not on the others. And I think from a policy viewpoint, it would be really interesting to try to extend your model in the direct. So I absolutely agree. We haven't done anything in terms of, so this is obviously within the same product. I mean, I think high level, you probably will still have some sort of like, there's still going to be some sort of free writing. But I'm sure there's some interesting novel effects, right? Like when you think about data sharing across multiple products, right? I mean, there's an interesting question like, does data sharing, when it's a cross, so in this case, there's no question. Well, actually, you know, even here, right? I mean, what's interesting, even if it's within the same product, data sharing has, well, it can intensify or it can make a competition where it can make it less severe. I think it's, I mean, it presumably it's even more complicated when it's across products, like whether data sharing is pro or, I mean, I guess this is something Alex and Greg look in their paper whether data sharing is pro or anti-competitive depending on. Alex, do you guys have, so you do have something on mergers, right? Data-driven mergers, presumably as data in different markets. Yes, yeah. Data is generated on one market and can be used in another one. So I think maybe that's a more suitable framework for addressing that. I'm sure you can take some, like you can take hours and extend it to multiple products. We haven't looked at that. Thank you. Okay. And so there are two, two last questions. One thing as also like the core organizer of the seminar. So I don't know, I think we may, I would love to answer more questions, but we may also want to send a strong signal of like being on time. Although I guess the good thing here is like people have the option of dropping out. That's what I was going to say. People can leave. Okay, then everyone, well, thanks everyone who is here all the way up to now. If you want to drop out, you're more than welcome to. I'm happy to stay until we exhaust all questions. Okay. So there's another question from someone whose name unfortunately I can't pronounce. You can't read. I can't read those characters. Is it Korean? I saw them. Yeah. I haven't used you. And I can't even dress. Okay. So in the meantime, Jay, oh, it's Yoshi. Okay. So you have a new view. That's funny. Yeah, hi. So my question was about the precise relationship between a cross users and within users. So if I understood correctly, in the within users model expectations matter, because if I join one of the providers, then I learn, but if somehow I switch, I lose the learning. Is that the idea? So even with within user learning, you do not have to form expectations. It's basically it's forward looking, but with obviously with within user learning with pure within user learning, I don't care what other people do. I just have to basically look at the present discounted value of utility offered by each firm. But I don't have to form expectations about how many other users are going to are going to use each firm because it doesn't matter to me. The only thing that matters is how many times I use the same firm. Of course, the key is once I combine and this is like, I can't emphasize this enough what's cool about this is like once you combine within user learning. So within you, only across user learning, no network effects. This was the discussion with Louise earlier within user learning perhaps even more obviously no network effects. When you combine the two, then you actually get a network effect, which is cool because then you actually you do have to care. I there's a switching cost because because if I make, you know, if I use a certain firm for a certain number of periods now it's costly for me to switch. And then secondly, because there's a switching cost now I actually have to worry about, well, how many consumers is each firm going to serve because that actually determines my across user learning component. If you combine the two, yes, expectations matter and you do have a network effect. But if you just have one, one or the other, you're not going to get any expectations don't matter at all. Yes, I mean, there have been some policy discussion on consumers having control over their data. So what if the consumer can bring on their own data and give to another firm within the user framework. So that's interesting. So I'm going back to where we discussed this right so let me see if I can reinterpret this as some sort of form of data sharing right so I think that so the key message here is that while learning by the winning firm decreases consumer surplus. And in some sense, if you're the consumer, like if consumers could coordinate, like I know that this firm is we know that this firm is going to win. But the problem is, because it wins every period actually that can hurt us so if we can somehow all agree. Well, let's join the loser for a period in order to make the loser stronger that actually helps our surplus because we keep the we keep the loser in play. So I think this is where you're right there may be scope there if consumers own their data, maybe they say okay I'm going to give my data to the loser or whatever to the other firms just to you know just to just to make that firm more competitive and put pressure on the winner. So in some sense it's really I think it's related to what we're talking about with data sharing here. Does that make sense to appeal. Yes, yes. So that's a real occasion to data sharing. Yeah, I guess that may be an interesting question. Like, can you because again we model this as data sharing and we chosen what a specific form of data sharing extreme. Maybe so there's probably an interesting way to model this idea that consumers own like you sort of micro found this like consumers own their data and they can choose how much data to share with how many firms in every period. I'm sure there's probably some something interesting to be done there. Jennifer has a question. So yeah I just wanted to follow up on that. Thanks so much for a nice presentation. I guess I was wondering if privacy might be a second dimension of quality. And so, then you might give up search result quality, your main dimension of quality for this second dimension of quality. So that's interesting right so this would be again would record this is actually very probably an interesting topic it would be interesting to microphone this let me see if I can translate this into our framework right so here we're basically saying the more consumers you've served in the past there's an auto like we basically black box everything into this f into the fi functions. And I guess what you're saying is like if you have privacy. Then maybe there's a, there's a second components like this it could be privacy but basically privacy is negatively correlated amount of data something like that right. Right. That could be an interesting extension right so maybe again you would have to like just like with JP I'm trying to think and I'm sure there's probably very interesting extension there like a variation of this model, or you can add privacy as a dimension and then there's you can also add consumer choice in every period whether to actually leave the data or like choose to not share the data or something like that. Right. And you do that so actually would be quite interesting, especially if consumers infinitely live because then they have to make calculations okay if I share my data will I get better quality but then, you know, it decreases my privacy, I share more data that actually allows the winner to extract more value, I think it's it is quite promising to go into it only affect one type of learning so it might only like the privacy thing might only affect the within person learning. That's right that's a good point yeah because otherwise you're right you're right because the other one is well if it's just for other people. Well, in principle you don't care that's fair. Thanks. Yeah, thank you. Gary. So, I was reminded by a paper by Kurt Taylor on supplier surfing. So it's a ran paper thing late 90s that consumers might have private information about their valuations and as the firm learns more they can extract more surplus for the consumer. So, think about this with the within consumer model that you have that consumers have incentives potentially to be switching firms in order to try to hide their true valuations, even though they might be generating more surplus with the firm. They can extract more of the surplus from them. And so I have to set up to not always buy everything on Amazon if they can personalize my pricing. So I think that's a good point so this is where we sidestepped all these issues right this is always them as this question by assuming. This is the assumption here that firms can price discriminate based on perfectly price discriminate now of course, if that's not true that's another interesting dimension right because consumers can try to hide essentially, like I can pull myself with other consumers in order to to face a lower price or something like that. That's fair. Actually, I think if I'm not mistaken. There's a paper. So when I first presented this at API OC. I'm blanking on the name someone presented a paper also related to data, which had a flavor what you're describing Gary. Oh, I think Simon sorry. So it's Simon, Simon Anderson and a co after I think they, they have something like that. I don't fully recall but it feels like that there's there's something along these lines across different there's something across different products and I thought it was something about consumers trying to like hide behind like not revealing their data in order to benefit from lower prices. Okay. Well, I think so there are no more questions on the on the chat. So, thank you, Andre. Thank you all for, for being here for joining us today. And so the, the next seminar is going to be in two weeks. So may fifth and we'll have really from Carnegie Mellon. Thank you. Thank you so much that this was a great DJing by Alex. If I mean obviously I want to emphasize if people have comments you know what we could do to make it you know I guess more more informative looking to improve the formatting do let us know by by email but those I mean certainly from my perspective those super helpful. So thanks everyone. Okay, right. Thank you everybody. Thank you and thanks for organizing. Thank you. Okay, bye bye.