 But after an hour, we'll have a very short break and we'll continue the discussion, like you know, more informal and not recorded. Okay, so please remind, let me remind you to please mute yourself. And if you have any questions, you can either use the chat for everyone or just send me, but I think it's better if it's the chat for everyone. So when Mayam is speaking, I think we'll only take clarification questions and you know, more substantive questions will be kept for after. So you mute yourself, you can control the video though, so no one is sure, I mean there's David and only he's showing the video. So I think it would be nicer and maybe if we can see some of the faces, it you know, it gives me the impression that you're speaking to a war. Hi everyone. So Mayam, I guess I've said all I wanted to say and the floor is yours, you have 40 minutes. Okay, okay, thanks a lot. So let me share my screen practice this yesterday. So hopefully it's working. Let's see, make it, the screen. So can everyone see my screen now? Okay, great. Everything's fine. Okay, okay, let me start. Thanks a lot for inviting me. It would have been better if it wasn't to lose. I love dessert, but it's great to be able to present my work. So today what I'm gonna be doing is talking about optimal rating system on platforms and I'm gonna do this, try to do this impossible task of presenting three papers in 40 minutes. We'll see how it goes. So this presentation would be based on three papers. The first one is an empirical work with Sean Cui, Giancarlo Staniolo and Steve Tedelus. So this is about, it's an empirical work where I'm looking at what the impact of certification and what happens when you change a certification to threshold on different parts of the market. So I'm gonna present only the results of these paper and I'm not gonna go through the details. And then the second paper, which I'm gonna be spending most of my time is about optimal quality rating. So in that paper, we are gonna be looking at theoretically what would be optimal when you're designing a reputation system and a rating system. And if I have time, I'm gonna go present some results from a third paper. So the difference between the second and the third is that the second one, we are only looking at adverse selection and we are assuming that the quality of the sellers are fixed. And but in the third one, we allow for moral hazard. So I'm sort of, I will be just presenting you some of the results to see how the results are different when you have to think about moral hazard as well as adverse selection. Okay, so let me get into the presentation. So I guess the audience here know that, all know that asymmetric information is can be very tricky and information design can help when we have markets with asymmetric information. And there are many examples, especially on platforms and two-sided markets. So eBay and Airbnb are some examples, but there are many others like Uber, Upwork, if you're hiring someone from internet, there are a lot of these markets with asymmetric information. And but this is not just a new thing. It has been around for a long time. I can love papers with coins long time ago. And you have this problem in credit ratings for consumers and corporate debt markets. Also you have the certification of doctors and restaurants. The common features for these markets is that you have adverse selection as well as moral hazard and the intermediary. So here you have an intermediary who usually observes more information than one side of the market. So they have a bit more information and they wanna decide what to transmit to the other side. So what I'm gonna be answering in these papers these are some of the main questions. One is that, is there a thing is too much information? Why not just intermediary transmit all the information to the other side and let them decide? Is there a thing that they actually, the other side actually prefers to not see some information? Is there some benefit of hiding some information? And then the question of optimal design comes into play. The second question is, are the both side of the market value the information the same way? So if for example, on eBay you have buyers and sellers are sellers better off sometimes at hiding where buyers prefer to see the information or vice versa. And the same goes through for the other kind of markets. And another question which we answer actually mostly in the first paper is to see how reputation mechanism information system impacts in comments and in trends differently. And we wanna see how these different part segments of the market are impacted. Should we think about them separately or they're sort of impacted the same? And the other one that I'm gonna give you some intuition for is that why we see a lot of simple signals and certifications in a lot of platforms. So a lot of times we see platforms giving different badges or they have like five, one to five ratings. So it's not very elaborate certification mechanism and I'm giving you some intuition of why that might be the case. Okay, so let me start with the first paper. So the first paper is mainly about certification. So there are a lot of market places that labor sellers who meet some minimum quality threshold. So they sort of do give some certification badge. So for eBay, which we are looking at this paper, they have this eBay top-rated seller or ETRS badge. Airbnb had this ABAB Superhost badge that you might have seen when you used to travel back in ancient history now and hopefully sometime in future. And then in Upwork, they have this Upwork top-rated badge. They give two top-rated sellers. Okay, so the benefit of the badge is that it mitigates some of the asymmetric information problem, but on the other hand, it can be a barrier of entry, especially if getting the badge is very hard and it's very valuable, then you might have sellers not wanting to enter the market and that can be a problem. So what do we do in this paper? In this paper, we are going to be looking at some policy that happened, a change of policy and happened on eBay, which they make it seem in this badge much harder. So beforehand, as you can see, about 10% of the sellers on eBay had the badge and afterwards it dropped sharply to 4%. So you might think, okay, 10% is not a lot, but these sellers or 4% afterwards is not a lot, but these sellers are disproportionately selling a lot of items on eBay. So these 10% of sellers beforehand were selling more than 50% of the items on eBay. So this drop was pretty sharp afterwards. So we wanna see what's the impact of this on incumbents, on interns, and also on total vector, especially consumers welfare. So what we see, okay, these are the questions that we try to answer in this paper. So we wanna see how the incentive of new sellers to enter the market has changed. So when you make this, getting the badge much harder, does that pay as a barrier to entry? And does that make it, does it make it harder for new sellers to enter the market? So what we see is actually, it really depends on the quality of the sellers. So the effect is heterogeneous in quality of the sellers. So sellers who which have a high quality actually have now more incentive to enter, the idea here is that these sellers whose quality is higher, now they get the badge still before and after this policy change, but afterwards, given that very few people have this signal, they get much more value from this badge. And as a result, they have more incentive actually to enter the market. And interestingly, the opposite is also true. And we see a little bit not as a strong evidence, but a little bit of evidence of some improvement in incentive on the lower end of the qualities as well. So this is now the case that these very low quality sellers are now pulled with better sellers. So we had these, let me go back to these pictures, we have these sellers who lost their badge, their badge before that are not badge anymore. So now the sellers who were not bad beforehand are pulled with these sellers who are sort of middle quality, high end of quality, and they can be better off. And this has increased their incentive to enter the market. And also we see a bit of disincentive for the sellers in the middle of quality because these guys had higher incentive and higher probability of getting the badge before, but now they don't. And so they are not gonna be entering as much. And we wanna see if we can see the same kind of impact for the incumbents about who are on the exit side. And actually what we see is the mirror effect here. And we see less exit from the two tails of the distribution. So for the higher quality guys who exit less often and we see more exit from the middle. So just exact opposite of what we see above. So we see a lot of actions happening on the selection part here. So for both sellers entering into the market and exiting the market. And we wanna see what kind of impact we see on the quality. So like if people start to improve their quality, as now getting the badges more valuable and the fewer people who are getting the badge, but we don't see a lot of impact. We see some, but the impact is not a lot. So what we see is that the sellers who were very close to getting the badge, but they were just below the threshold to get the badge, they sort of pushed their, some of them pushed their quality a bit higher to get the badge. But we don't see a lot of actions in other parts of the quality spectrum. So one thing, for example, that I expected was to see some more actions on the higher end, but we didn't see a lot of impact there. For the VFR interrogations, we see some improvement for buyers on average, but about two to three percent. Okay, so I think I'm going a bit too slow to make sure that I can get to the second paper, I have enough time for the second paper. So what the main takeaways here is that for digital platform is that the certification policies affects quality distribution. It's not just, you shouldn't just think about the average quality, you have to think about the quality distribution. It can broaden or contract the quality range depending on what you do. And so it depends like if you're a marketplace it really depends on what you want to maximize, what your preference is here. If you care more about the sellers in the middle or sellers on the high end or low end, then it impacts the optimal choice of certification threshold for you. So if you want to have more of the high end sellers, so for example, like if you pick a study, you want to have like this, products that have like a very high number of supporters, you want to be very specific on the high end and try to incentivize those guys to come. But if you're Uber and you don't care about people driving others with Tesla's and you just care about the average quality, you don't want to be giving some signal that gives or like incentives to very high quality drivers. And the other thing was that it seems that the quality control policy seems to be more about affecting selection rather than changing quality. If you saw some impact, but very little. Any questions here or should I continue to the second later? Okay. No, there's no question. So now let me go to this paper that I'm gonna go actually into the model and give you a little bit more detail. So what you have noticed is that in that paper, we had a simple model that sort of guide us into what kind of empirical work we wanted to do, but we didn't have like a very strong, complicated model that we can look at what's the optimal level of information this course. And the model was not, we didn't actually identify the model. So it wasn't, it was a reduced form analysis. So what we didn't answer this question, okay, what is the optimal and as a function of different parts of the assumptions on the market. So what I'm gonna do in the second paper and what we did in the second paper is actually doing that. So we wanna see what's the impact of improving information on welfare and then you can define welfare about consumer, producer, total welfare or some base on consumer and producer and we wanna answer what's gonna be the optimal level of information disclosure. And we wanna see what's the role of supply especially and demand considerations and also on your choice of optimal information disclosure. And also I'm gonna go over some results about what happens when you wanna give four or five signals and what would be the optimal. So for example, if you're just giving one signal, so if you're certifying a set of sellers, so like just this badges that we just talked about what would be the optimal thresholds of food and how much of the gap in the total welfare you can pose by just including one, two, three signals. Okay, so let me start on the model. Let me slow down a little bit. Okay, so here we have a unit mass of firms and we assume that they have a quality Z that's come from distribution F of C and we're assuming that this distribution of quality is fixed. So here we only have other selection no more hazard in this paper. As I said, the third one is the one that includes more hazard. And we're assuming that we have competitive market so you can think of the supply function which is just gonna be first derivative of cost function to be the deciding factor here for the sellers. And you're assuming that the cost function is the same for all producers. So that's the supply side of the market. So for the sellers or producers they have a quality and the cost function is the same for everyone. So depending on the price they're getting they're gonna be deciding what quantity to produce which is going to be S of P and that's just the first order of margin cost. That's the margin. Okay, so the demand side we have a mass M of consumers and they have a unique demand and they have to choose between all the options present to them. And the utility they have is gonna be Z plus theta minus B. So this theta is gonna come from a distribution psi of theta. So that's the type of the buyer so that determines the demand function. And so they Z you can see is the quality of the item they're getting or the expected quality of the item they think they're getting depending on the signal they see and then they have to pay a price of P so that will be subtracted from their utility. And we are assuming that outside good is normalized to zero. So this is just a fancy way of saying that they have a demand function. So for example, if the quality Z is zero this is gonna be their demand function. And if they know the average quality of a group is ZL this will shift the demand function for them up. And if for example, you have two groups. So for example, you are certifying a group of one group and not the other. And the group that are certified have the average quality on ZH. They will have the red, the upper demand function and the others, the low guys will have the lower, low demand function. So one thing to note here is that in equilibrium these buyers should be in different between the choices. So if you have different sellers with different qualities selling at different prices the Z minus P for everyone should be the same. And the way that we have given that the way that we have modeled this they sort of actually the value of this additional Z would be all consumed in P. So it's not gonna be a big factor for the consumers but it will be important for the whole market. So here we wanna talk about the optimal information disclosure. Okay, so what do we have here? So we have the buyers share a common prior. So let's say they all know the average quality of the sellers. And here we wanted as a planner the planner has to choose what information to these codes. Let's assume that the planner knows F. F was the distribution of quality of the seller. So it can be coming a function of that for simplicity just assume the planner knows all the information they know exactly the type of the seller. And then they have to decide what information to reveal and the information that they're gonna be revealing would be has to have the same mean as F but they can hide some information. So it can be coming any mean preserving contraction of F can be an information that they're gonna be revealing. So they can do some garbling of information. So they can hide some information but they can potentially give exactly F or the other extreme they can just give no information and just give the mean of quality which is known by buyers anyway. So it's like giving no information. Okay, so what's the problem that they're gonna be solving if so they're gonna be choosing a G from this script G set. So they're gonna be maximizing this expression. So here we are assuming that the rate they put on consumers at once is gamma and this Q of G is the quality the equilibrium output quantity set. So the planner problem is gonna be maximizing here. They wanna choose a G member of this script G script G to maximize this weighted welfare. So this is one minus gamma is the weight they put on the profit of the sellers. And here this would be the welfare for a consumer cell phone. Okay, so this is the demand function, the English demand function for the consumers and this P is the price in the market for quality zero. Okay, so let me stop here, see if there is any questions or can I continue? So what do the buyers know? So the buyers see the signal G and let's assume that beforehand they only know the average which is also in G so they don't know anything else. But we have some extensions when they have some prior information but we need to have that to be common prior everyone share that information. It will be sort of similar. But there's something I probably misunderstood. I thought that so when you introduce the originating among sellers, so the buyers observe a seller's type? When? Well, no, so the buyers only see this signal of the seller. So this G can be, okay, badging some sellers. This G can be telling you, okay, one a star, two a star, three a star, four a star or it can be exactly the quality of the seller. So it depends on what the planner is going to disclose. So what I had here was, okay, this Z for the utility that I had here was what was the expected quality that buyers were observing? So it's a function of Q, so it can be the same for everyone if there's only one signal. I mean, if there's no signal and they say, okay, it's just average quality. So it depends on the signal, the G signal that is transmitted. Okay, so any other questions? No. Okay, so here, before I go to the results, let me give you some intuition here what happens in this market. So here when, for example, you only care about the consumers, so here you just, you want to maximize only this part. What happens is that this, you have these two different effects of giving more information about the sellers. One is that as you give more information about the sellers, you're gonna be making more dispersions on the prices in the market. So for example, if you don't give any information, there is only one price in the market and every seller in the market is getting that price. What does that mean? It means that all the sellers in the market are gonna be producing the same level. So high-quality sellers and low-quality sellers are producing exactly the same level. Okay, so when that happens, you, so you have the same level of quantity for each seller, but high-quality and low-quality. So you have a lot of mixing of qualities. So this can be bad for the buyers because the average quality of the items on the market is going down, but what happens actually here is that given that the buyers are paying that extra value on quality, that might not be actually that important. What is more important for the buyers is gonna be coming from this quantity in the market. So when you have no information, we're back to again to no information about the sellers, the total quantity is gonna be a function of this average price, which this average price is coming from the average quality in the market. And the level of this quantity depends on the supply function. So for example, if your supply function is convex, that actually giving no information would reduce the quantity in the market. And the reason here is that, okay, the high-quality and low-quality, they're all getting a price that is on average sort of low and everyone is producing a little bit. When the supply function is convex, if you give some information, now you will have this high-quality sellers getting higher prices and producing a lot more items. And that will push the quantity in the market higher and that would be beneficial for the consumers. So then you will get to a case where it's better to give more information for the consumers and otherwise not. So let me actually go with these effects of information a bit more slowly here because that's sort of the main idea of the paper is that if you give more information, it spreads out the sellers according to their quality. And so here the direct effects is you have this reallocation of output from low to high-quality producers which actually the benefits of this extra quality is arbitrage is a way and it's sort of translated to higher profits because the prices are going to be higher because all the consumers who wanna get this higher quality has to pay higher prices and because they need to be in different, the consumers need to be in different on the margin, they're not getting the benefits actually. The indirect effect is that the total output can change. So the total output is gonna be increased if S is conveyed. So that's actually beneficial for the consumer. What is gonna be bad for producers? And it's bad for the producers because it will decrease the price as a function of Q. So that if you look at the price as a function of Q when you have no quality, so for quality zero that's gonna be going lower and it sort of pushes the prices down for the consumer, for the sellers as bad for the sellers but good for the consumer. So here the indirect effect is sort of goes in two opposite effects for the consumers versus producers and as a result we get some mixing of the result. So here's this picture shows the whole results that we find here and that when S is convex and we care about the consumers more, you wanna do full disclosure and the same is true when S is concave and gamma is less than half but then gamma is, you care more about the consumers and the supply is concave. It depends on if S prime over S double prime is increasing. So for example, if it's increasing you wanna do full disclosure at the beginning and then cooling. And otherwise you wanna do cooling and then full disclosure. So here you wanna disclose some information but you wanna see when you wanna disclose some information and it depends on if the first effect is dominant versus the second effect is dominant. So the first effect when you have the reallocation is more determined by S prime and the second effect is more determined by S double prime. And depending on which direction is, so this would be a lot of positive and negatives here and not gonna go through the details but the disclosure result depends on if S over S prime over S double prime is higher low. So you wanna be disclosing. So for example, here, if you wanna just look at this side when you care about the consumers more but the supply function is concave, you wanna disclose at the beginning and then pull all the sellers. So this is the point that you wanna be when S prime over S double prime is increasing you wanna be pull at the higher end and this goes at the lower end. Okay, so let me, don't have a lot of time. Let me give you some overview of this second result of the paper. So here when you wanna actually maximize total surplus given that the second effect sort of arbitrage is the way you always wanna give all the information but we see a lot of times that you're just giving limited number of information and what we wanna see is that, okay, what happens if you can only give a few signals? First, we show that if you wanna give a few signals, we wanna give partitions. So the best thing to do is to find the Z1, Z2, Z3 and say give the signal that the sellers are between the Z lower bar and Z1, between Z1 and Z2 and so on. So it will give you what is the average quality in between, so what buyers will expect is the average quality of the sellers that are in within each of these intervals. So what we show here is that when you wanna choose this Zs to maximize the total surplus, you have this necessary condition, just forget about the top one for now and what it tells you, this is the average cost and benefit of increasing this Zk, the tertial Zk. The cost is, okay, these guys are gonna be producing higher quantity, but the benefit is how you have some extra production, which is beneficial when you're looking at total surplus because the prices just won't get to the other. So this is the condition that is quite intuitive, but the problem is that all this quantity, the total quantity in the market and the average, the quantity for each seller is gonna be in dodgeness and we show in another proposition how we will find this. So given that I don't have a lot of plans, the main idea here is that this is the k-mean criteria for optimal cost rate and the main idea here is that you wanna put this thresholds at different levels such that you give the most amount of information possible. So this is like the cost freeing thing and that people do all the time. So you wanna sort of cluster the sellers and give clusters of the sellers which are most similar to each other and the different clusters you want to be as far away as possible. So you're sort of giving these signals such that you give the most amount of information and what we show later is about the performance and wanna see how much of these welfare gap even total surplus of total surplus you can give when you give, you move from like how much, like the total of effort gap being between the total of effort gap of full versus no information. I wanna see how much of these welfare gap you cause when with different number of thresholds. And here we show that for example, this is for some different distributions you can see with only one, so when you put only have two groups or one signal sort of like certification with even just one you can get quite a bit of this gap cause with just one signal. And when you get to the point signal so it would be like one to five a star you get to very like more than like 90% and above. And one thing to note here is this is the amount of number of sellers who are gonna be batched and you can see that this number of sellers can be actually quite low. So it can be 5%, 9%, 20% so this is the share of the sellers who are gonna be batched. And so just having one threshold so it can be quite selected. So I think this is a very cool result because it sort of tells you that, okay in this model we didn't have any cause for acquiring information or for the buyers who are seeing these signals trying to understand what the signal means. So we don't have any of those costs in the model but if you include any of them then it might be actually optimal to just give one or two signals because you can see that you can get quite far but just giving one signal. So it can be sort of give you some intuition why you have this way often in a lot of market places and you give just very limited number of signals about the sellers. So the take away of this paper is that the optimal disclosure depends on the curvature of the supply function as well as the slope of supply function. So and when you're maximizing total sales process best disclosure or information because you reallocate market share from low quality to high quality sellers and this increase average quality and overall it's better but when you care for example more for the sellers or more for buyers depending on the supply function you sometimes wanna hide from information. So even for the consumers sometimes it's better they prefer not to see some information. And we show that there is a for the linear supply case there is a simple solution when you wanna give only limited number of signals so this is the came in clustering technique actually and it's very easy to find and the gains from giving this signals decreases fast in the number of thresholds. So when you have more like with only one you get about 50% and with five you get close to 90% most of the time. Any questions? Thank you Maian. Now the floor is open for questions. So if any of you has any questions I guess you can put it in the chat and I'll give you the mic. So I have a couple. So regarding this first bullet point on the takeaway slide I mean this is a broad question but I'm intrigued by the sort of the connection that your work has with the connection on third degree price discrimination and demand curvature. So it seems like you're looking at the sort of the dual problem in a sense that I sort of I can't make precise but I wonder if you have you thought about the connection there? Is there something to say? Actually not, I have to think about that. Thank you. Yeah I think it's sort of is yeah I guess it is sort of similar because like with third degree price discrimination you're sort of thinking about the buyer side and you're doing sort of similar actions here, right? So you're deciding how to prove buyer so it's sort of similar to this one exactly. Yeah, I have to think about this. That might be actually sort of a very dual problem of that. So also another point that I had is that so when you discuss this simple solution for the supply case, the linear supply case, sorry, you assume that basically the platform can really and I mean that this is the assumption of your work that the platform can sort of partition the signals whatever way it wants. So in practice I guess one challenge is that you know even if eBay say gives a five-star scale, the empirical distribution that you observe is not at all sort of uniformly distributed or even normally it's like everyone is like five stars. Yeah, yeah. So is there a way for me? Yeah, so I mean here what I have in mind is not just like for the platform to observe the information and decide what information they're gonna be transmitting. That's one thing that I've noticed or realized is that a lot of times now even Amazon doesn't give you exactly the average ratings of the item. So even the five star rating that you see might be a function of, it is a function of feedbacks that consumers have received but it's not exactly the average. So you can be sort of getting all the information but what you're gonna be transmitting is not just a simple average of them. It can be okay, I'm gonna give four and above to just 10% of the sellers based on these criteria and so on. So it's not just transmitting the information that you get. So one thing actually in all of these papers at the moment, this is something for future research is that for now I have assumed that the platform exactly sees the information about the quality of the sellers and but it needs one other feedback level which is not modeled yet and that's the part that you see some signal from the buyers and then how to translate that to the quality of the sale. So that layer of it is actually missing in all this research at the moment. So that's dynamic part of actually how to get the information, how to solicit this quality information is actually missing here. So does anyone else have a question? So I'm not seeing anything on the chat. So I don't know if my chat is not working or if... I was very clear. Yes. I have a couple of questions. This is Bruno Julien speaking. Do you hear me first? Yeah, we hear you. Yeah, I have a poor microphone. Now I have a question to link the first part and the second part because the first part was a lot about exit and entry of the sellers and the second part is really an intensive margin but there is no extensive margin. So what would happen if you let's say introduce fixed costs in the second part and perfect signal or something like that? Yeah, so we have some extensions for that in this paper as well. So it actually makes it so when we get there, we have to make some further assumptions about supply and we have this linear supply but the results are... Let me think about that. So when we have entry into the market, so here we're assuming that for all the sellers, they're in the market here and there is no entry, but when you do have entry margin, let's see. So the result that we have is mostly for the few thresholds and we don't have a result for the full disclosure part. So for the full disclosure part, you will have a similar thing with wanting to, like when you want to disclose more, you're gonna be having the same kind of results, you're gonna wanting to disclose when you have the entry condition as well. The results probably will be different when you are not gonna be the result without entry margin was no disclosure at some points because when you do that, then you might be impacting entry. And that's the part that we have to work on. So we have some preliminary results but they're not actually very complete, but it sort of makes it a bit harder. Let me see if I remember exactly what's, okay. So here the point is like when you have entry as well as sellers producing, when you give, for example, if you give no information, that will make the average quality a bit higher, average price a bit higher for lower quality sellers. So it will sort of give this lower quality sellers entering the markets in higher rates and increasing your total quantity in the market. So it sort of makes this impacts the results actually make it a bit natural. Stop sharing, so we can see everyone. So it will make the results, the effects a bit stronger. So it sort of goes to the same direction and makes it a bit stronger. But that's for the linear supply case. And we have to do it for the full disclosure for the disclosure condition actually and for the full supply function, we haven't done that. Thank you. Anyone else? Okay, so I guess that was all very clear, Mariam. So thanks a lot. We will have another seminar in two weeks' time. This time it will be André Fratkin. So thanks for joining us and see you in a couple of weeks. Bye-bye. Bye, thank you.