Tony Ke (The Chinese University of Hong Kong) - Information Design of Online Platforms

[Music] recording the floor is yours thank you very much Alex um thank you for having our paper so the title of my presentation today is uh information design of online platforms so we are motivated by the observation that online platforms have gather a lot of consumer data and this data and data analytics enable platforms to learn consumer preference and this has two important applications first uh the platforms can use this information to personalize product recommendation for buyers and second uh the platforms can use this information while data analytics to um offer Target advertising for the sellers so I'm going to offer a little bit motivation for each of these two applications uh for Port recommendation it's getting more and more important actually become one of the most important Revenue engine for e-commerce platforms for example for Amazon it is claimed that product recommendation is responsible for more than uh onethird of the revenue generated um and for streaming services such as Netflix is even more uh on the other hand uh the target advertising so here I'm showing you a chart uh that basically list the biggest advertising Platforms in the world uh so if you look at uh the number four number three and number four are Alibaba which is the um uh largest e-commerce platform in China uh and also Amazon so Alibaba and Amazon these online retail platforms have become um uh the one of the largest advertising platforms only after Google and meta so um our research questions are based on these two observations uh we're wondering how we put our recommendation interact with talkar advertising are they to stand alone decisions or there is some inherent tradeoff uh so our take is that there is an inherent tradeoff between them because both of them can be seen as monetization of prominence the idea is that uh consumers may have some limited attention and therefore there's some prominent position on this platform uh and the platform can either recommend a product to take the prominent position or it can sell this prominent position through uh advertising uh to Sellers and due to this inherent tra off we're going to illustrate using a model that the platform need to balance the sales commission and advertising Revenue okay and lastly given we see so many ads these days on these e-commerce platforms were wondering about the social welfare implications of the sponsored ads uh on uh this Ecommerce uh platforms uh we're going to answer this research questions by recasting our question into a public information design problem uh to be more specific the platform is going to uh collect information collect data from the uh about the match information the match between the sellers and the buyers and I'm going to use this match information to communicate with uh buyers about which product is likely a match and also communicate with the sellers about whether a buyer is likely the target so you can uh think about this communication process as a product recommendation algorithm or a Target advertising algorithm uh so uh so we're going to particularly uh consider a a information design problem um where the platform communicate this information to both the buyers and the Sellers and given the platform's information design the buyers will decide should I follow the platform's recommendation to buy the product or should I continue to search other products and and the sellers uh will decide how much to advertise to become prominent uh as mentioned you can visualize this information design in practice as a recommendation algorithm a targeted or a Target advertising algorithm and because of the automated nature of the algorithms uh we feel that the um commitment assumption that is important for Bas and persuasion it's uh more natural to hold in our setting uh here are the preview of the main results uh so by using this uh uh public information design framework uh this is public information design because the platform only Design One information there's only one information design that is the same for both the buyer and the seller uh so we by using this framework we unify the analysis of personalized product recommendation and targeted advertising so we're going to show that there're basically uh two size of one coin and we are also going to show that we can reformulate this uh platform's information design problem as a optimal control problem of consumer search length uh the idea is that by deciding how much information and what kind of information to reveal the platform essentially will influence how much the consumer will search so uh the information design problem could will essentially can be formulated as a uh control problem of consumer search LS we show that um there is a tradeoff between match efficiency and surplus extraction the Surplus extraction is from the sellers um by inducing seller competing for the U prominence and we also show that the uh optimal design may be socially inefficient um because the optimal design May limit consumer search and makes the Mass match the product with a long tail of and matched products for product recommendation uh so due to the uh time I will skip the literature review and jump directly to the model okay uh so the model uh will consider consider a two-sided Market uh so a platform um the commission rate is exogenous and there's some commission rates Alpha uh their n number of sellers uh their production cost normalized a zero and we use a index n under I consumers uh index I now this is one of the most important assumption we make um for the model that is where we assume that there is one and only one seller that matches with each consumer okay um particularly for a matched seller for the Matched seller for a particular consumer there is a elastic demand uh and for the unmatch salaries this zero demand uh so this is a a binary match um uh this can be S thought of as the extreme case where consumer have very strong preference so they like some Sellers and they don't like others okay and uh uh the reason we make this assump is uh because uh the match will actually be the state of the world uh where the platform will Design information upon and this assumption basically simplify the state of the world a lot uh because there's only one seller so the state of world just become which seller is the match that's the uh that's the state of the world uh consumers outside option of not buying anything from the platform is uh normalized at zero uh so uh a prior uh we assume that all saries are equally likely to be a match uh this also simplify the analysis um uh a lot so I will touch upon this point later uh when uh we look at specific model analysis so we have a uniform prior for order uh consumers okay so this is the model setting for the uh two-sided markets so uh I will uh give a little bit more details about how consumer make a search decision and how sellers make advertising decision so for consumer search it's a sequential search uh for match values and prices so it's it's essentially a wiin spe model uh search cost C uh a prominent position with zero search cost uh so you can also make this a positive search cost uh so what we require here is that the consumers start their searching with the prominent position uh the platform holds a uh second price auction separately for each consumer and we show that it can be robust if we add a reservation price and the sellers will set the price and also the beat for each consumer so here the price is the same for all the consumers but the beat could be different for different consumers so therefore this is targeted advertising uh lastly about the information design uh the state of the world assignation can be uh represented by this Vector uh where n one is the index of the mat seller for Consumer one and so on so we have a n dimensional uh Vector that represents the state of the world uh the platform designs the information structure so uh uh k um and the sellers and consumers they observe uh high and also a signal realization s um so just to help us understand this a bit more we can consider two special cases uh the first case is a perfect information that is uh the signal is just equal to the state of the world uh in this case you can think about this is like recommend the perfect match to each consumer uh the second uh is a total and informative so you have the S does not depend on Omega at all in this case you have a random recommendation okay all right uh here is the model it's a summary of the model and also the timeline so uh the platform uh commits uh to information structure pi and then a signal as realized according to Pi and sellers and consumers obser the platform's choice of information structure and the signal realization uh the sellers set the price and beat and then consumers observe who are who which seller take the prominent position uh and then they decide should I continue to search or should I just uh purchase from this seller or should I take the outside option okay uh so that's the timeline uh uh so next I'm going to uh move to the equilibrium analysis um so um first um we have the following llama that basically pinned down the equilibrium price we show that in any perect Bas in equilibrium every seller with a positive demand will set the price at the Monopoly level uh this is a direct implication because we have this binary match so for any model with a binary match is is easy to argue that uh basically in in the similar uh manner of diamond Paradox you can argue that uh um the seller will set a Monopoly price uh so we have so we have this on purpose because we want to simplify the pricing decisions so essentially the pricing decisions become inactive in our model now given this U Monopoly price um then in a transaction between a match party uh we can calculate the consumer will always expect the following Surplus U and the seller will uh uh get a profit V okay so because that there is a elastic demand uh so it will be convenient to work just with you and V instead of D directly okay so in the following analysis I'm going to work Direct directly with uh u and v uh so that is the uh how they split the pi uh in the uh match party okay uh the search Cost needs to be positive but not uh higher than consumer surplus otherwise no consumer wants to search um and second is uh we have the following Lama that characterizes the consumer's optimal search strategy uh so um it's quite intuitive consumer should search by the descending order of her poster belief so the platform first uh designed the information and then there is a signal arrive now based on this uh uh signal the consumer will have a posterior belief about which salary will be the match okay so you can s let's say the consumer believe uh Sal two has a 50% with the match Sal three has a 20% with the match and so on okay and then if the consumer decide to search and what is his optimal search strategy well she will search by the descending order of her posterior belief now this sounds a little bit uh straightforward but there actually there is some complication behind it uh because in our setting all the sellers are not independent um because there's one and only one seller um so it means that so let's say if I find out that seller one is is not a match then the consumer will know that the rest of the sellers will be more likely to to be the a match because there will be one match always be one match so it means that when a consumer searches one seller she will update her belief about all other sellers uh this is this will violate uh the independence uh Assumption of Wis men's model so we do not we cannot apply wisan directly but we can essentially but because model in our setting the the setting is simple enough you can still characterize the consumer search strategy in the following way okay so given that um in the following I'm going to analyze two benchmarks uh these two benchmarks will help us understand the tradeoff or the mechanism behind the model and then I will look at the general uh information design problem so the first benchmark I'm looking at is a full information design so in this Benchmark the platform just uh reveal the perfect information to the market um so given this is given this a full information uh design um then we know that um the consumer then we can show that all salaries will be zero um and then the uh platform will actually let the match seller win the auction uh but of course the match seller does not need to pay anything this is a second price auction uh because all other sellers be zero uh so the the master seller will earn a profit of V uh minus the commission so it's a one minus Alpha time V uh consumers do not need to search because the match seller is takes already takes a prominent position so they just a buy from the prominent seller and consumer expect a surplus of you and the platform obtains a profit of alpha times uh the number of uh consumers times V uh basically the platform only get commission there's no advertising Revenue okay so uh therefore under this full information design the auction is as if inactive um even though the Market is organized by auctions but uh the auction is aive in inactive and we essentially end up with a pure um personalized recommendation strategy okay so basically our full information design replicate a pure recommendation strategy uh next I'll talk about no information design so uh on the no information design basically the platform just recommend randomly so they designed the information uh uh structure to be a random uh totally random and in this case uh consumers posterior belief will be the same as the prior belief um so they believe all the salaries are equally likely to be a match so we can show that uh consumer has the following search strategy is n or all if the search cost is very low then the consumer will search until she find a match otherwise the consumer will never search Beyond The prominent seller The Advertiser seller the intuition is the following because during the search process the consumer actually become more and more optimistic so when I see the seller one is not a match I update all the match probabilities for the other sellers higher and because all other sellers are equally likely to be a match so it means that if I want to search Sal one I will be willing to search Sal two because search Sal two is more likely to be a match than seller one and if I find out Sal two is not a match I'm more likely to search I'm more willing to search Sal 3 uh so therefore it will becomes a n or all strategy uh for consumer search uh so um then uh we can show that this is the market equilibrium so if the cost is very low um essentially consumer will keep searching until find until she finds a match then we have basically complete information right because consumers is going to find out who is the match so this is as if there is complete there is perfect information the only difference is that consumer need to pay a suchar cost so we have consumer surplus is you minus some expected search cost now otherwise uh when the search cost is not so low then we show that all saries will beat uh their expected uh match probability times the uh payoff Upon A match so we know that they're equally likely to be a match so that's why we have one over n here and so one minus Alpha is is after commission and V is the U payoff upon match okay and then um because in this case the platform commins not to use information so therefore um a random seller will win and because this a second price option so all seller be the same so therefore this seller will get zero profit um consumers will only consider the uh advertized product or The prominent uh product um and because each product is equally likely to be a match so uh the match probability will be 1 over n times the consumer utility you upon a match the platform get profit from both um the this is the commission and uh the advertising Revenue so the advertising Revenue come from seller speed and the commission come from a consumers expected match so if you add them together this is I I divided by m uh so um so we can see that so the no information design basically replicates a pure and targeted sponsor search ad strategy so that that was the strategy used by Amazon before 19 uh sorry before 20 uh 2019 so after that Amon started uh allow targeted advertising Okay so uh so we basically have two benchmarks uh and we can compare them so if we compare them um when just compare these two extreme cases when the search cost is very low uh the platform will be indifferent between a no information design or a full information design um because consumer is going to search anyway under no information design on the other hand as long as suchar cost is not so low and then there is a meaningful uh uh uh distinction between these two design particularly we show that when Alpha that is a commission times n that is a number of sellers when Alpha time times n is greater than one uh the platform prefers full information otherwise it prefers no information so the trade-off is quite uh uh straightforward so there is a tradeoff between U match efficiency and surplus extraction so from the platforms perspective by uh Supply supplying more information you can get a better match uh a better match means a higher commission but on the other hand more information will make the Sellar more heterogeneous uh when seller become more heterogeneous they will reduce their incentive to compete and lower the advertising Revenue so basically on the platforms tradeoff between sales commission and advertising Revenue imply is is due to this match efficiency and um s plustic extraction tradeoff okay uh so we have analyzed the two Benchmark cases so next uh I will give you a little bit more ideas how we uh formalize the platform's information design problem so we see that the an auction is personalized um and also the price is inactive so it's always Monopoly price so this basically allow us to separate consumer problem for each consumer so essentially uh we have a we we only need to consider one consumer so let's just consider a representative consumer eye this consumer uh has a posterior belief upon receiver signal si uh this M and M n is basically denotes what is the probability that the match seller is n so the seller n is the match seller because ni by definition is the matches seller for Consumer I okay so ni equal to n is the match seller is seller n you notice that when we write mu n we do not have a s here uh this is on purpose uh so uh if you're familiar with basian persuasion essentially uh in basian Persuasion you deal with the um uh distribution of posterior belief but here we can deal with posterior belief directly uh the reason is that all the cellar are exanti symmetric so then basically different signals essentially induce a permutation right you can think about if I reveal uh one kind of signal then solar one will more likely to be a match another type will maybe solar two will be more likely to be a match so uh in the paper we basically Pro that we can deal with the posterior belief directly where what really matters is mu1 mu2 up to Mu n so without loss of generality we can rank them in a uh descending order so basically what really in the end what really matters is um the distribution the posterior belief distribution okay um so the platform's problem essentially becomes trying to design this posterior belief uh uh subjected to the normalization condition to maximize The Profit okay so next I will uh basically show how we calculate the profit this Pi mu function sorry just a quick clarifying question so for that do we need that Pi is the concave function uh this profit function or because otherwise one would think that you couldn't benefit from some randomization or beliefs uh that's a good question is it a concave function uh what do you mean by conf this is a multi-dimensional uh mean like no I mean the pi Capital pi profit function are the capital uh Capital the one in the objective maximum Pi yes but this is a the the MU here is a vector right and and the pi is a function of that Vector is it the concave function a concave uh it when we have when we have uh multi Dimension do we still have concave uh I'm not exactly sure um okay we can postpone the the question the yeah um yeah uh so so you maybe you can you can see later because we do not exactly use U when we we think about concavity we think about optimality condition we do not use that directly to solve these problem so I have never thought about that yeah okay yeah okay okay um okay how we calculate this uh Pi mu um uh to calculate this P mu so we introduce a uh a definition called search length this is defined as the number of sellers uh including the prominent one uh this consumer will visit if she has not find a match yet okay so if she has not find a match yet the number of Sellar she will visit now of course this number will depend on two thing first is the consumer's posterior belief that is the me here second is who takes the prominent position so that is n so we basically we assume that given consumer posterior belief as Mu and given seller n takes the prominent position uh then we can calculate uh the uh search lens um by solving consumer's optimal search problem uh that is the L okay and we int sorry to introduce so many different notations so uh there's another this is a uh important because we trying to reform malize this problem so this uh there's another pie so in previously uh here this piie is defined on the posterior belief okay but here this is a different this is a different Pi this this Pi depend on the posterior belief and also l so this is defined as platform's Revenue given the posterior belief and a exogenously given search lens l so you can think about um if the platform can exogenously fix consumer search lens at L and given consumer believe me what would be the resulting consumer sech strategy and the resulting platform revenue and in the paper we basically show that um the pi mu which is something we are interested can be linked to this new definition by the following equation uh so it will be on if we put a equilibrium search lens into this uh uh uh into this function then we get these two are equal so um this is not as straightforward as it seems but due to the time I will not dig into the details uh so the complexity lies off on the off equilibrium path um so um it's but but just let's just take this as a as as it is um and then uh we can actually write down the uh Pi mu L as the following right so because if we if the platform can exogenously fix consumer search lens at L then we know that seller one up to seller L minus one will be zero um because uh they will be searched anyway right so if they are the match they will be searched so there's no point to take the prominent position uh therefore there will be zero um and therefore uh actually seller L will be the winner um remember here we the seller one to seller seller n their belief is in a descending order okay so uh so seller L will be the winner uh because then seller L speed will be higher than seller L plus one so the we can write down the platforms uh profit as a following there is a commission fee that come from all the possible match from all the L Sellers and advertising Revenue okay so the commission will only conf from the L sellers because the consumer will only search L sellers um and the advertising Revenue come from seller l l plus one speed okay uh so then um we can write down the we can rewrite the platform's problem as the following uh so the platform is essentially trying to uh choose the search lens um of course this search lens must be the equilibrium choice of the consumer so we have this uh um uh constraint here but now because given this we can actually write down uh this uh Pi mu L function we can actually solve this uh uh one by one so we essentially we can solve the uh problem for each given search lens L okay so hopefully uh this give you a little bit taste uh of uh the analysis so I'm not going to bore you with more details so this this is the equilibrium okay uh the equilibrium is that if the search cost is low uh in the previous comparison of the two benchmarks we already see that the platform will be indifferent between no information and full information here we basically show that this result is more general actually any posterior belief generates the same platform revenues now otherwise if the search cost is not so low we find that there are two cases depending on the um uh the commission rate so if the commission rate is relatively High then we show that the optimal design is a full information so this is a full information design otherwise um the design is a little bit more complex so the platforms will uh trying to uh um make the Cons trying to recommend two products with relatively high probability and a long tail of products with equal but low probability so I give a little bit more intuition about this case in the next slide okay so uh under this information design the consumers will only visit the prominent seller so the consumers will just search The prominent seller will not continue will never continue to search and we show that this is a Bas imp plausible okay so um so let me give you a little bit in uh intuition about the the the optimal information design in the second case so this is this is basic so the platforms optimal information design is the following so suppose n the true state of the world is n that is seller n is the Matched seller then the uh the information structure is that then the platform um upon receiving n uh s will be with some probability will be equal to n so there's some probability that the platform will in recommend the match seller but there are also some probability that the platform will match will recommend a a unmatch seller with high probability and there's also a a bunch of all other sellers the platform will recommend with low probability now upon receiving this information structure and Signal the consumer will have the following belief that seller and has this probability to be the match seller and seller M plus1 has this probability to be the match seller and all other sellers have a relatively low probability to be a match other so why this is the optimal information design well if you think about the platform's revenue it come from two parts uh first is a commission and second is advertising Revenue so we want to recommend the match seller with a high probability because because this will ensure there is a a good match uh with there is a match with good probability than therefore this will uh guarantee we have good Commission on the other hand for advertising Revenue it's a little bit more complicated we basically want um so we call this Contender we want the contender we want to maximize the contender's reward upon winning and also minimize the contender uh uh reward upon losing this will incentivize the contender to uh beat uh higher the higher the possible so uh this by um recommending uh this Contender with a high probability this give this uh Contender a a a high reward upon winning because he knows that if he wins he has this probability to be the match so he is more likely to beat high on the other hand we also want to keep a long tail of m match sellers uh because uh if we don't do that then after consumers let's say the consumer visit the mat The prominent position first uh and find out the prominent position is not the match uh now now if we don't have this long tail then the consumer will be sure that the next seller will be the match uh because then he update uh she update her belief uh upon the uh upon seeing not a match from The prominent position she will infer that the next seller will be 100% to be the match then knowing that this seller has no incentive to beat uh because she knows that uh even if uh I lose this I will still get the consumer right so U by keeping this long tail we ensure that after consumer seeing the prominent position is not a match the consumer still does not want to search so then this seller will know that okay so if I don't win the position I will not be searched so then I have a strong incentive to win uh so that's why we want to keep a long tail this long tail give this seller a incentive to beat high and okay I think I'm a little bit uh out of you have two two more minutes two more minutes oh okay that's great so uh um so we know that the socially optimal design is information uh full information design uh this means that and our analysis implies that uh the sponsored ads so this incentive uh the platform's revenue from advertising May introduce some social efficiency because it involves noisy matching uh and we show that as such cost gets higher or when there are more sellers on the platform uh the noisy matching is more likely to arise and the platform profit can increase uh just to conclude um so we offer a unified understanding of personal recommendation and and Target advertising so these two things um in our framework can be um captured by the public information design framework uh we show that the optimal information design strategy may be socially efficient uh when the commission rate is uh relatively low um and it is socially efficient because it entails limiting consumer search and mixing the match product with a long tail of and match ons for recommendation so thank you very much thank you Tony you're perfect on time and and now we will have discussion by Andrew Rose Andrew you have five minutes okay thanks can you hear me fine yep uh okay so first of all thanks to the organizers for the opportunity to read and discuss this paper which uh I enjoyed a lot um so as you saw in the presentation it's very topical it's taking um issues from IO like platform design and consumer search uh and then merging them with tools from Theory like uh information design um and it does this using what I think is is quite a simple but very elegant and parsimonious model um and even though the model is quite simple the analysis is is certainly non-trivial and it leads to quite Rich interesting uh findings okay so I just have a few minor comments and my first kind of uh most important comment is that if I understood correctly this is already RNR anyway so you may not want to change too much uh but with that said let me go through a few or give a few suggestions so my first kind of comment or suggestion is about this um match technology that you employ so you basically assume that there's exactly one product in the market which matches so this is kind of going back to this Chen and her paper the EJ paper maybe even before then y men's work um so the advantage of that is it makes the pricing very simple um which I which I think is a good thing to do here otherwise things would be very complicated um I guess you know you might you might question the assumption that consumers match with exactly one firm so as you saying the conclusion obviously if you match with for sure with multiple firms and things are a little bit trivial but I wondered what happens if there's some chance that a consumer actually doesn't match with any firm um because obviously pricing will be exactly the same pay to be simple um but consumers have less search incentive which maybe gives the platform more power on the advertising side so I wondered if you'd thought about that something a little bit related is at some point in the paper you have comparative Statics in the number of firms which I think is fine and interesting but of course it's a little bit special because as you increase n you're reducing the probability that each existing firm matches so like you're you're basically changing two things at the same time um so that was my kind of my first comment about the match technology another comment is that um in the model all the firms are symmetric exanti in the sense that they all match with equal probability and uh I think we're all guilty of assuming this in models because it makes things uh makes the pricing much more tracable but here of course the pricing is already quite simple and tractable so I wondered if you thought about what would happen if you had some asymmetries in match probabilities can you say anything interesting right because obviously if some if there are some firms where the match probability is very high then again consumers are probably going to go and search them anyway which means that they don't have to bid very aggressively so I'm wondering if you want to if if if you can characterize what you should do presumably the platform should try and handicap those firms uh so that was another comment another comment about participation so you assume that all the consumers are already participating on the platform um and I just wondered um you know so what happens what would happen if they had maybe an outside option um you know or maybe some stochastic outside option can you think about how the platform design should be changed the information design should be changed to encourage consumers to participate and in the case where you have a have um a low search cost I just wonder if this could act as a tie break right because technically you show that any design is going to lead to the same outcome so can you basic can you maybe use participation as a tiebreaker um another comment so Alex how much time do I have left two minutes okay one two minutes okay I think it should be fine so another comment um which you didn't touch much upon in the talk is this assumption that you basically have symmetric signals for both the consumers uh and the sellers so as you say somewhere in the paper of course that would be complicated if you did it if you had different signals because each side would have to update given what they saw about what the other side might have seen um you know I wondered if there anything interesting to be said maybe at some point in the paper by looking at say one-sided Benchmark so what if you only give a signal to the consumer side I guess you want to give them right if the search cost is small then it doesn't really matter but if the search cost is large I guess you want to give them a perfect recommendation so that you maximize commissions if you're only giving a signal to the seller side I guess you don't want to give perfect recommendations otherwise you're not going to get any bidding so it seems like those two could be interesting benchmarks and you can compare that with what happens when you give the symmetric signal uh to both sides maybe let me just finish with a couple of other minor things um you know so again there's some discussion this in the paper maybe you can make it more explicit you could think about the platform's incentives to obis skate so basically choose the search cost maybe it's maybe the maybe the platform can choose at some cost to have a higher search cost and it seems like it's weekly better off uh doing that of course it's a cheap point but the the alpha here is taken as given um and you could discuss that one thing if I understood correctly is that there could be actually a perverse effect of a higher Alpha right in the sense that if you compare the full and the no information benchmarks if the search cost is small there's nothing interesting happening anyway it's the same outcome but if the search cost is large it seemed to me that if you increase Alpha so the share that platform is taking at some point you're going to split from no information to full information it seems like that's actually going to benefit the sellers and the consumers right so the fact that you have this Distortion due to the information design means you get this kind of perverse compar static with respect to the alpha I'm not sure if that's interesting or not but it may be worth pointing that out somewhere and let me just finish with a couple of small presentational things so personally I found it a little bit annoying reading the paper when you had this I everywhere in the platform profit expression so why why not just set I equals 1 given that you don't have any participation and another thing is that you have this you introduce at some point this notation X plus or minus Z so again you didn't go through this in presentation in the slides at some point you had plus or minus zero so if I understood correctly this is basically Epsilon tending to zero so I'm just wondering can you say a little bit more about that is this basically if you if you broke ties would you not need that or given that this is an Epsilon and you're taking the limit does that mean that the that the optimum is not well defined that wasn't that wasn't clear to me so I think maybe some discussion of that you know before you introduce that in the paper would be good and Alex I'm done so I hand back to you great perect