CS 286r Comments
10/20/2008
Sagar Mehta |
Betting on Permutations The permutations paper provides important results regarding the computational complexity of auctioneer problem of risklessly matching up wagers in permutation betting. The matching problem for subset betting and pair betting result in different complexities, where the first can be solved in polynomial time, but the second cannot. It is important to point out that the state space for subset betting grows much larger than for pair betting. If I have n horses, then there are 2*(n choose 2) possible pairwise comparisons, but there are a much greater number of subset bets I can make. I also find it interesting that the idea of A>B cannot be expressed in terms of some combination of subset bets. I think a possible direction for future research, empirical or theoretical, would be to examine how well these markets work with regard to efficiency of information aggregation. I also wonder how a bidding language that allowed both types of bets would perform under the same metrics. Pricing Combinatorial Markets For Tournaments The second paper considers combinatorial markets for tournaments and uses a Bayesian network to represent market distributions. I found it interesting that the paper mentioned the NCAA tournament. I wonder whether it would be practical to allow for combinatorial bets in such a tournament. The reason is that assuming players have a fixed budget and limited time to process information before submitting bids for such a tournament, it is possible that a larger state space of events will result in each estimate being less likely. For example, if I believe Stanford and Duke will both lose in round 1 and are on separate sides of the draw, why would I even bother betting on Stanford vs Duke? Does this limit the potential efficiency/information aggregation gains of this type of market? If they are allowed to place odds on every combination, they might be less likely to make an accurate prediction of each individual combination. |
Nick Wells |
Betting on Permutation Basically, subset betting is betting on the specific rankings for a few horses, and pair betting is betting that horse X will beat horse Y. This paper looks at pair betting and subset betting proving that it is NP-hard for the matching problem. The verification can be done in polynomial time. I think further explanation of the intuition and implication of the results of this paper would be helpful. Something that potentially interests me about these two papers, is that if the implication is that the computational complexity is unrealistic for these types of betting situations, what types of approximation methods are generally used? I would be interested in learning more about those. The paper mentioned something about a natural greedy alorithm being ineffective, but what, if any, other methods are there? Pricing Combinatorial Markets for Tournaments Combinatorial markets are able to approximate an entire joint distribution. This paper claims that they are NP-hard in computational complexity, meaning that they are at least as hard as the hardest non-deterministic polynomial problem. The paper shows that one can represent the distributions as Bayesian networks. For any change in distribution following from a purchase of bets, any resulting distribution can still be modeled by a Bayesian network. This paper shows that single-elimination tournaments can be represented as stated. I am not sure that I fully understand the full breadth of the results presented. Perhas that would be useful to discuss more. |
Subhash Arja |
The first paper "Betting on Permutations" by Y. Chen, et al, seeks to analyze the auctioneer problem of matching up wagers in permutation betting. The paper looks at subset betting and permutation betting. Both of these scenarios are important because they directly correlate to real-world applications of the stock market and horse race betting, to name a few examples. The paper states that buying or selling a stock is equivalent to betting on the stock's value. In the case of subset betting, where participants decide which positions a candidate will fall, the authors show that there is a polynomial time algorithm for the subset matching problem. The key insight is to show the reduction of the problem to a maximum weighted bipartite graph, and, since this solution has a polynomial time result, the original problem can be solved similarly. In the scenario of pair betting, however, optimal matching of pair bets is shown to be NP-hard. One limitation in the paper is the bound set on the bets of any property and ordering. It is stated that allowing this would increase the computational complexity too drastically. A previous paper on the market scoring rule is used as a basis along with previous work on combinatorial auctions. In an attempt to solve the problem in pair betting, a greedy algorithm is used but is shown to be a poor approximation. This idea can be extended beyond this paper by conducting research to find an algorithm that gives a closer approximation. One project idea that could result from this paper is to actually apply the concepts into a software implementation and simulate the stock market. This will verify the usefulness and accuracy of this model to real life scenarios and could provide insight into flaws or inadequacies. The second paper, "Pricing Combinatorial Markets for Tournaments" by Y. Chen, et. al., also covers combinatorial market concepts by applies them to the specific case of tournaments. The authors consider single-elimination tournaments, which somewhat limits the complexity of the calculations. They reach the conclusion that the solution to the pricing problem in this is NP-hard. This leads to the conclusion that considering non-single elimination tournaments would yield the same result. By choosing to represent the market distributions as Bayesian networks, the authors assume that at each stage, the available and past moves are known. Through extensive mathematical proofs, the solution is shown to be NP-hard. One idea that the paper uses that I found interesting was representing each game as if it were an independent market. This assumes, of course, that the result of another game does not affect the outcome of another game in the same round. One project idea from this paper is the application of this model to real sports tournaments. In those tournaments, the victory of one team over another has minimal impact on other games in the same round. This would give insight into the accuracy of this model and possibly give way to approximate solutions that are at least polynomial in complexity. |
Ziyad Aljarboua |
Betting on Permutations: This paper discusses the case of permutation betting. This type of betting is applicable to most types of markets where people wager including the financial market where people wager on the security's value. This paper also examines a scheme called Subset betting that simplifies the the complexity of the auctioneer problem where bidders have large number of subsets of orderings. The scheme proposed tries to offer an expressive bidding language that is still feasible to compute. Having bidders bid on sets rather than individual positions simplifies the computation tasks and minimizes the number of orderings subsets. The other proposed scheme is Pair betting where bidders bid on final outcome. In general permutation betting markets, an auctioneer needs to solve both matching, a decision problem, and optimal matching problems. In other words, the matching problem, also called the existence of a match, is fact that the auctioneer needs to find a subset of orders submitted by bidders that can be matched in a risk free manner that will give a non negative profit. Te other problem that the auctioneer needs to address is the optimal matching where he/she needs to take into consideration other factors such as profit or volume. On the other hand, in the Subset betting markets, traders can bet on subsets of positions a player might end up at or bet on subset of players to occupy a particular position. In this type of market, the matching problem is more complex than in the general permutation betting markets due to the fact that the number of available securities to bet on is exponential. So, the matching problems in this kind of markets have large number of constraints.Finally, pair betting markets allow traders to bet on weather a candidate will rank higher than the other. This type of betting is very complex. In conclusion, we can see from this paper that an auctioneer can find the optimal set in the Subset betting. An unexplored area left in this paper is the computational complexity of optimal indivisible matching for subset betting. -------------------------------------- Pricing Combinatorial Markets for Tournaments: This paper discusses the problem of pricing combinatorial markets for single-elimination tournaments arising from the fact that the probability distribution over a set is exponentially larger than the number of base bets. This paper also demonstrates a pricing algorithm by restricting to a natural betting language. Consider a n team tournaments represented as a set of binary vectors of length n-1 where agents can bet on the 2^2n-1 subsets of the entire outcome space. This paper shows that in this example, the pricing problem is #p hard. When types of bets agents are restricted, Bayesian network structure is not preserved. This paper also discusses other problems but i found it hard to follow. |
Haoqi Zhang |
Betting on Permutations (1) and Pricing Combinatorial Markets for Tournaments (2) The main contribution of the first paper is in establishing the hardness of two bidding languages that allow for fairly expressive bids in a market where the outcome is a permutation. In subset betting, where a bidder is allowed to bet on multiple candidates landing at a particular location or a particular candidate landing at multiple locations, the authors prove that an auctioneer can compute the optimal match (e.g., maximize worst case profit) in polynomial time. The result is surprising in that there are exponential number of possible bids that a bidder can submit. By using the ellipsoid method and a theorem on its polynomial time complexity as long as there exists a polynomial time algorithm to solve the separation problem may give non-positive outcome. Here the authors give an ingenious reduction via bipartite matching that uses weights depicting payout to link candidates to orderings, which has a polynomial time solution despite an exponential number of possible orderings. The authors then considered pair betting and showed that it was surprisingly NP-hard. I find this to be quite a negative result because it seems fairly natural for bidders to have beliefs about relative likelihoods of victory over candidates than positional estimates. The main contribution of the second paper is in establishing the hardness of the general pricing problem for tournaments and providing a tractable bidding language for expressive bidding in tournaments. The bidding language allows for bids over teams winning particular games, possibly conditioned on reaching that game, as well as pairwise bids over winners, given that two teams face off. The authors show that under this bidding language, updating probabilities via the log market scoring rule can be done by updating the Bayesian network. Given the tournament tree structure of the Bayes net, updates can be done in polynomial time. The authors then show that the expressiveness of the language leads to voiding independence assumptions among different branches of the tournament, but that these assumptions are preserved by restricting the language to be conditional on teams winning. In both papers, I found the application of known theoretical results relating to LP, bipartite matching, and Bayes Nets to be crucial to finding tractable expressive languages. It seems that future progress may also rely on breaking down the combinatorial problem via the considerations of known effective restrictions of the known, e.g., tree structures and bipartite graphs. One part that strikes me is how pairwise bidding is NP complete in the permutations case but similar pairwise bidding in tournaments is not, which seems intuitive in that we need not consider all permutations in tournaments but still baffling at the same time. One question I have for future work is whether any of the results can be extended to non-permutation/tournament settings, whether seemingly general settings via restricted languages or other games where the outcome space is structured. |
Angela Ying |
Pricing Combinatorial Markets for Tournaments This paper discussed combinatorial betting, such as in a single-elimination tournament. The paper was interesting in that it uses the framework of a Bayesian network to represent the tournament and how parent games relate to child games. All in all, the paper had two main contributions - it determined that generally, pricing for this tournament was #P Hard because a game of n players can have up to 2^(n-1) different ways to bet. In addition, it developed a polynomial-time algorithm to price the tournament bets on a restricted betting space. Furthermore, it determined that it is necessary to limit the space of possible bets to those that allow independent events to remain independent - thus, the only type of conditionally dependent bets that are allowed are of the form "x will beat y given that they face off." This way, the Bayesian network relationships and unrelated probabilities remain unchanged, while overall probabilities change slightly. This paper was different from the other one we read because it uses Bayesian networks rather than graphs to match bets / make the market. However, they are similar in that they both concern a betting space that is exponentially large. One difficulty I had while reading this paper was understanding some of the terminology - specifically, I had no idea what #P hard meant and so I did not know why having a #P hard problem was a bad thing. However, overall the structure of the paper was good and it presented an interesting approach, the Bayesian network, to solving this problem. A possible extension to this paper may be to look at tournaments with different elimination rules, such as double elimination or round robin, and see if we apply the same methods to those types of tournaments. Betting on Permutations I thought this paper was very interesting - it addresses the matching problem that auctioneers face when people bet on permutations of an outcome. For example, a horse race will have many different outcomes depending on which horse places where, and the auctioneer will want to aggregate all these bets to accept or reject them in order to minimize the maximum loss, and hopefully always make a profit. The main contribution of this paper was proving that matching subset bets with divisible amounts can be done in polynomial time, while matching pair bets (both divisible and indivisible amounts) is an NP-hard problem, which is interesting because the space of possible matches is much smaller in pair matching than subset matching. Finally, this paper proved that figuring out if a solution exists can be done in polynomial time. I thought it was particularly interesting that most of the proofs were done by reducing problems in graph theory and changing the problem into a problem that we already know is NP-hard or NP-complete. I guess I had never thought to have bets placed in a graph in order to match them. An aspect of the problem that I think would be interesting to examine is the optimizations. In all the problems examined in this paper it has attempted to maximize the auctioneer's profit and attempted to minimize the worst case scenario loss. Even though these are NP-hard problems, I am wondering if there are randomized algorithms and/or optimizations that can be done while actually computing the solution in order to reach a good, but not perfect solution in polynomial time. In addition, it would be interesting to compare the results of these optimizations to auctioneers in real life who may not have the time to figure out how to best match different bets. How do auctioneers decide on the spot whether to accept or reject a bid? |
Zhenming Liu |
The papers [CFNP] and [CGP] both studied prediction market on permutations. In markets of such setting, the only information to be aggregated is the ranking of a set of items. It is slightly surprising that in this restricted problem, we still cannot do much to design a computational feasible mechanism (or maybe the emphasis on the negative results on both papers gives this illusion…). In [CFNP], both the subset betting and the pair betting problems are reduced to optimization problems in more conventional form such as linear programming or finding some specific properties over the graph. Some of the techniques appeared in the reduction are quite neat. In [CGP], they studied the complexity of logarithm scoring rule and obtained a few immediate impossible results. When imposing further constraints in betting language, polynomial running time algorithm becomes possible. After reading an array of “theory” papers in prediction markets, I start to be confused about some high level questions. Perhaps the most important question is that I am not sure that whether the model of the market or the technical part in proving the theorems is more important in papers of such type. When facing a computational complexity paper, the techniques used in the proofs are arguably the soul of the paper and ultimately we would like to build up a systematic toolset to approach complexity problems. And we expected the invented techniques can be reused. For these theory paper, on the other hand, the techniques appeared in the proofs seem to be isolated and it is not clear how we shall treat these proofs. A second confusion is that the distance between theory and practice in the prediction market seems to be even larger than such distance in computer science. Even using the most abstract model, a theoretical computer science result has a direct and clear implication in the “real world”. Meanwhile, perhaps it is not even easy to validate any model in real market, the connection between theory and practice seems to be quite remote in prediction market studies. [CFNP] Y. Chen, L.Fortnow, E.Nikolova, D.Pennock “Betting on Prediction Markets” [CGP] Y.Chen, S.Goel, and D. Pennock, “Pricing Combinatorial Markets for Tournaments”. |
Brett Harrison |
Betting on Permutations By Chen et. al. Pricing Combinatorial Markets for Tournaments By Chen, Goel, and Pennock "Betting on Permutations" explores the auctioneer's order matching problem in auctions involving the outcomes/ordering of candidates, e.g. participants in a race. Since allowing all n! different possible bets would make the auctioneer's problem computationally hard, the authors provide two betting languages to restrict the betting space. The first is subset betting, which allows bets of the form "some candidate in subset A will finish in position b" and bets of the form "candidate a will finish in some position in subset B". Even though the number of such bets in exponential in the number of candidates, the authors prove that the auctioneer can solve the order matching problem in polynomial time complexity. The second betting language is pair betting, which allows bets of the form "candidate a will finish ahead of candidate b". Even though the number of such bets is only polynomial in the number of candidates, the auctioneer's corresponding matching problem is NP-hard. "Pricing Combinatorial Markets for Tournaments" focuses on the market-maker's pricing problem in a single-elimination tournament setting. The authors show that the general problem (i.e. all possible bets are accepted) is #P complete. However, the authors give polynomial algorithms for computing restricted bets of the following types: "team i wins game k", "team i wins game k given that they make it to game k", and "team i beats team j given that they face-off". Several interesting questions for further research come to mind. First, what makes an expressive betting language easy to compute? Or perhaps, how can you construct a betting language so that it is easy to compute? Second, what is the underlying difference between the market maker's problem and the auctioneer's problem? |
Peter Blair |
1. Computation in a Distributed Information Market, Feigenbaum et al 2. Betting on Permutations, Chen et al 3. Pricing Combinatorial Markets for Tournaments, Chen et al In [1] the authors introduce a prediction market model in which players are assumed to bet truthfully. The focus of this paper is to study the conditions under which the market price converges to the correct probability of observing the event underlying the security. It turns out that we obtain convergence of the market price to the true value of underlying if and only if the payoff is giving by threshold functions of the information available to investors. A major strength of this paper is it's use of helpful simplifications e.g. uniform distribution of information bits over a binary outcome space (0,1); and it's explanation using clever examples, e.g. Example #1 where it is shown that two agents can have common priors but different posteriors owing to their different information (Example #3 is also an excellent example!). One weakness of the paper is that it does not specify why investors trade, which is an important piece of information when it comes to defining the investors motivations and possibly can speak to whether not the investor will play truthfully or bluff or be reticent. 1. Why is price defined p = (sum b_i)/(sum q_i)? 2. How do we understand the weights specifically if w < 0? Also what does the condition sum x_i *w_i >= 1 correspond to physically? 3. A priori, why would we expect C(x,y) -- the carry bit to be an interesting physical quantity? 4. In the paper the authors claim that the result of the paper would hold even if there were "super investors" who knew all the information. Is this trivially evident? Additionally what would happen if we did not require that each investor have a unique bit of information, e.g. two players share the same bit of information (and know, or maybe are unaware) that other players possess that same bit of information. It seems like having this bit of information diminishes in value given that another player also hold this information, is this intuition correct? 5. "The effect of price accuracy and precision": the authors make an interesting point here for future research -- investigating the stability of these threshold influence of convergence equilibria, by introducing noise arising from irrational traders. It could be however that this model incorporates irrational traders already, given that the motivation of players is not specified in the paper: "We sidestep this issue by simply assuming that the informed agents will trade (for unspecified reasons)" (pg. 6) 6."Incremental updates": this recommendation for future research would investigate the effect of changing the information bits as the system updates. This is quite an interesting proposal. A similar proposal would be to increase the total amount of information in the market over time and study it's effects on the results presented in the paper e.g. during the course of a primary election and a general election we learn more about candidates as voters and this undoubtedly affects how we would vote and also how we believe our neighbors would vote in the election as well, e.g. Barack Obama's relationship with Ayers, Rezko & Rev Wright being revealed at various points in the primary and general election campaign; analogously for John McCain, his selection of Sarah Palin as his VP in addition to the Keating 5 scandal. In [2] the authors study markets governed by games in which investors bet on the ordering of n outcomes. This general game subdivides into "pair betting" in which a player can wager on one candidate beating another; and in "subset betting," which centers on a candidate having a given ranking in a range that is a subset of the total outcome space. The central result of this paper is that the case of pair betting is an NP-hard problem, where as the game involing subset betting can be solved in polynomial time. This is a surprising result given that we would naively expect the pair betting game to be more tractable than the subset betting, which has a potentially larger outcome space. One of my hope for this classes presentation is that the speaker elucidates why this is the case. Indeed this is explained in the paper, nonetheless, I'm hoping for a intuitive reasoning. I would also appreciate a demonstration of how we could construct graphs of the type in Figure 3 to understand the games studied in this paper. In [3] the authors consider pricing of contracts governing "combinatorial markets for single-elimination tournaments," e.g. NCAA tournament. As a member of the Duke class of 2006, I especially appreciated the reference to Duke in the beginning! The case for studying these kinds of markets is made convincingly give that combinatorial markets provide a means of connecting securities/bets that are related to each other, e.g. the probability that Duke wins the tournament, and the probability that Duke beats UConn for the title. A surprising result in the paper relates to how the authors condition on the outcome of a game acausally. While, I don't fully understand how this arises (Figure #1), it reminds me of a Bayesian game in perfect information, where we know the full game tree and work using backwards induction to find the BSPE. There is something unnaturally about this concept and I would be curious to see how this type of acausal condition compares to causal condition when it comes to predicting the outcome of NCAA tournaments. If one were to run an experiment testing these two types of condition using NCAA data (from previous brackets filled out: "bracketology"), what would be the best way to structure an experiment of this nature? Aside: Can we have a section on Computational concepts e.g. NP Hard etc. where we also touch on some of the terminology of the stock market and perhaps go over some of the logarithmic scoring rules that are so important to the previous 5 papers that we have read. This would help people encountering these concepts for the first time to feel more at home with them? |
Avner May |
Pricing Combinatorial Markets for Tournaments, by Yiling Chen, Sharad Goel, and David Pennock Betting on Permutations, by Yiling Chen, Lance Fortnow, Evdokia Nikolova, David M. Pennock These papers focus on the difficulty of aggregating information about complicated probability spaces, like single-elimination tournaments or permutations. What these probability spaces have in common is that their outcome space grows exponentially with the number of parties involved (like teams in a tournament, or horses in a race), and thus lend themselves to much more difficult algorithmic questions regarding matching interested betters. I thought that one of the main contributions of these papers was in describing the tradeoff between the expressiveness of a certain betting language, and the computational complexity of solving its respective matching problem. I think it’s an insightful realization that people could have opinions about a wide range of aspects of the eventual outcome, and that the market would ideally be able to gather as many of these opinions as possible. The applications of these papers are extremely diverse – to any scenario where people are betting on single-elimination tournaments (eg, NCAA), or any competition in which the contestants are ranked. One assumption which was made throughout both of these papers was that the auctioneer would be unwilling to take on any risk whatsoever. Often the optimization problem was the auctioneer attempting to maximize worst-case profit. This seemed like a very strong assumption; it would seem more reasonable to me to assume that the auctioneer was attempting to maximize expected profit, or something more along those lines. |
Xiaolu Yu |
General pricing problem for tournaments is #P-hard in general, but can be simplified using the polynomial-time pricing algorithm when asset types are appropriately restricted to a natural betting language. As an example to the tractable market-maker driven combinatorial market, single-elimination tournaments is investigated, and the side effect of the expressiveness of this natural betting language – introducing inexisting dependencies between games (sequential games in a game tree) – is recognized, with a possible solution as further restricting the language (however, many real-world pricing problems may not be simply restricted). In the case of "match-up winners", the global distribution is constructed from the independent market via the Bayesian network, and each trade updates only a single parameter of the Bayesian network. These significantly alleviate the computational complexity, and the execution time for traders is linear in the number of teams. "Betting on Permutation" discloses some fairly interesting characteristics of two expressive betting languages -- subset betting and pair betting as well as shows that the exponentially big linear subset betting program has a corresponding separation problem that reduces to maximum weighted bipartite matching and consequently can be solved in time polynomial in the number of orders, because for example the existence of a match and the optimal match problems can be solved as long as the corresponding separation problem can be reduced to the maximum weighted bipartite matching problem. If orders are divisible, an auctioneer can find the optimal set and quantity of orders such that his worst-case profit is maximized in polynomial time. In contrast, pair betting suffers NP-hard complexity. By reducing to the minimum feedback arc set problem, pair betting still leave an auctioneer with an NP-hard optimal matching problem. A sufficient condition for the existence of a match for pair betting is proved, but necessary conditions are not pretty much identified. Given the greedy algorithm gives poor approximation for indivisible pair betting, better approximation for the pair betting markets are still an open question. Other possible expressive betting languages for permutation betting scenario, which can overcome the side effects and weak points of these mentioned ones, need to be explored as well. |
Alice Gao |
Paper A (Betting on Permutations) considers the auctioneer problem of risklessly matching up bets when the market uses a call market mechanism. The authors are really using this scenario to examine the issue of designing market mechanisms. Some design goals are: (1) The language should be expressive. It should be able to reveal as much information as possible for each order. (2) It should be intuitive and easy for traders to use. For instance, high-level description of things might appeal to the natural way of thinking used by most people. (3) It should be computationally feasible for the auctioneer to match up bets. The central concern for this paper is in a way how to balance the trade off between expressiveness of the language and computational complexity of the matching problem. Paper B (Pricing Combinatorial Prediction Market for Tournaments) considers a market run by the logarithmic market maker mechanism. In particular, this paper considers the problem of pricing combinatorial markets for single-elimination tournaments. The main result is a polynomial-time algorithm for the pricing problem when only certain types of bets are allowed. This is the first example of a tractable market-maker driven combinatorial market. In general, these two papers leave me with the impression that designing combinatorial markets is a very hard problem. These two papers only examined a tiny fraction of the possible questions that could be asked, and we have already had to use so much theoretical arguments. Moreover, It seems to me that different types of combinatorial markets might have different special structures within them. Paper B would be an example of taking advantage of some special structure of the combinatorial market to find a tractable market mechanism. So a natural direction of future work would be regarding a classification of combinatorial markets and their similarities and differences, with special regard to the topic of market mechanism design. Paper B also mentioned an important point: The large number of bet types typical in the combinatorial setting would require liquidity added by a market-maker in order to aggregate information effectively. Indeed, paper A was mainly concerned with finding a betting language that will be computationally feasible for the market maker and they have not tackled the question of the effects of different betting languages on information aggregation. In particular, the usefulness of a betting language should in part be quantified by how their usages would help with the information aggregation. |
Travis May |
One of the central problems of prediction market design is to create markets that both reveal substantial amounts of information and do not overwhelm traders with possibilities. This creates a key tradeoff, as gathering additional information requires additional markets, which makes it more difficult to establish markets between willing buyers and sellers. The two papers we examine for this week propose optimal solutions for this tradeoff. In “Betting on Permutations,” the authors propose that there are two key types of information that market-makers are seeking to elicit in combinatorial markets – markets in which there are many candidates competing. The first is called “pair betting,” and is a bet on whether candidate A beats candidate B; the second is called subset betting, and allows betting on the subset of candidates that can finish at a particular position, or a subset of positions that a particular candidate could end at. The paper shows that subset betting probabilities can be computed in polynomial time, and pair betting is NP-hard. Similarly, in “Pricing Combinatorial Markets for Tournaments,” the authors propose that the key questions that market-makers seek to elicit in tournaments are: “will team i win game k,” “will team i win game k given they make it to game k,” and “will team i beat team j at game k given that they face off.” The paper shows that this set of questions can be answered in polynomial time by representing market distributions as a Bayesian network. Determining an entire joint probability distribution is inefficient and computationally costly, typically provides much more information than what market designers seek to elicit. The papers each show that there are much more efficient methods than computing a full joint distribution, allowing markets to more effectively reveal relevant information. |
Victor Chan |
Betting On Permutations The main contribution of this paper was to determine how an auctioneer in a permutation betting setting can match up bets. The paper deals with whether a match exists, and the computational complexity of finding the optimal match. The types of permutation bets include subset betting, where bets are placed for whether a certain outcome will fall within a range, or pair betting, where bets are placed on the relative order of two candidates in the permutation. The matching problem is examined in the divisible and indivisible cases. The results of the paper showed that subset betting can be represented as a weighted bipartite graph, and solving for the optimal solution in the worst case can be achieved in polynomial time. (Which is the solution to the maximum wighted bipartite matching problem). The pair betting solutions were not as easily solved and both the optimal indivisible and divisible matching problems were both NP-hard. The paper shows how bets normally placed in different categories, ie the winner category, 2nd place category, can all be aggregated together to give more information regarding the outcome of the permutations. Since this paper shows that the matching problem in subset betting can be solved in polynomial time, it is possible for auctioneers/book makers, to combine certain betting categories. Any betting market, where the out come of the events are a permutation of candidates, can use the results from this article to create a single subset betting scheme. Furthermore, since the matching problem that is solved is also giving the auctioneer the optimal profit (no risk), they would have incentive to implement such a scheme. I was unclear on the benefit of finding whether a match existed in pair betting. If one cannot find the actual optimal solution, then is the notion of the existence of match of any use? Perhaps it can be used by the auctioneer to determine whether to accept a bet, since it will know if a match exists? Pricing Combinatorial Markets For Tournaments The main contribution of this paper is to show how to price a sale of a combinatorial outcome, and how the sale affects the overall distribution of the combinatorial market. First the paper shows that a tournament outcome type game can be expressed as a Bayesian network. Then a purchase of team j winning game g, will only cause the distribution of game g and it's ancestors in the network to change. The paper further shows that having this type of distribution calculation was a result of the language being used, such as "team i wins game k", which leads to an unexpected dependencies in the distribution. To reintroduce the Independence, a language for betting only allowed bets such as "team i beats team j, given they face off". This change maintains the over structure of the Bayesian network, however the arrows are now pointing to the final game. In this manner, when betting occurs, only the game at stake's distribution changes. The paper also shows that pricing the combinatorial market is #P-hard, and presents an approximation method in the appendix. This approximation can be used to help implement a market, where combinatorial betting can occur as described in the paper. Further more the paper also shows how to place bets that are conditional events, ie to bet A|B, is to buy AB, and short ~AB. This can be used by investors to make conditional bets in a market where it is not supported. Finally one project that can be done for this paper, is to implement the approximation tool to see its effectiveness in pricing bets. This will be interesting, since we can see if the resulting distributions are close to the actual distributions of event outcomes. |
Michael Aubourg |
Betting on Permutations In markets, bilateral agreements do sometimes not exist. However, multilateral agreements may be possible, so it's not a problem. In our case, traders are going to bet on the outcome of a competition among n candidates. In order to build on exchange, we need to find two or more that together form an agreeable match. The problem is the following one : It's too heavy to compute the n! possible outcomes for the final order. And the probability of convergence between two traders, would be very small. So the solution is to try to divide each order. We have of course to care about the heaviness, I mean, the time we need to compute it. Here, it is a polynomial time. 2 different kinds of bet : 1) the subset betting 2) the pair betting The drawback with the indivisible bet is that it is an all or nothing order. If we compare with financial markets : These markets are not exactly complete. Indeed there are more states than securities. However, the dynamic trading allow us to consider these markets as complete. This is very convenient, because if this hypothesis is true, it means that the "law of one price" can be apply. This is the most basic concept in Finance. The goal here of the researcher is to find a language that : - reveal as much information as possible - maintain computational feasibility The concrete goal is to define the optimal matching problem for the auctioneer. The existence of a match is a decision problem. It only returns wheter trade can occur without any risks to the auctioneer. In a subset betting market an auctioneer can find the optimal set and quantity of orders to accept such that his worst-case profit is maximized in polynomial time IF orders are divisible. In a pair betting market, the optimal matching problem is NP-Hard with both indivisible and divisible orders. Pricing Combinatorial Markets for tournaments Goal of combinatorial markets : work to more efficiently aggregate information by estimating the entire joint distribution of dependent observations. In the betting language, agents can buy and sell assets of the form "team i wins game k", or conditional assets. Unfortunately, it is impossible to preserve independence in an aggregate information. Thus, the usual relationships are restored if we only permit bets of the form "team i beats team j given that they face off". The main argument is that the general pricing problem for tournament is #P-Hard. Here is the framework of the paper : the general pricing problem for tournament is #P-Hard ---> This is a difficulty----> We need to derive a polynomial-time pricing algorithm by restricting to a natural language. ----> As a consequence, dependencies are introduced between games which we would think they are independent This is due to the fact that it is impossible to preserve independence in an aggregate distribution. ---> If we restrict the language, we can however restore the usual independence. This is the strongest point. |
Hao-Yuh Su |
1. Betting on Permutations 2. Pricing Combinatorial Markets for Tournaments Both of these two papers address the relations between the expressiveness of the bidding language and the computational complexity. In the first paper, the authors proposed two types of betting: subset betting and pair betting and discussed about two problems: matching and optimal matching problems, correspondently. Each problem is investigated in both divisible and indivisible ordered. It is showed that the corresponding separation problem, which can be reduced to the maximum weighted bipartite matching problem, of the subset betting can be efficiently solved in polynomial time by ellipsoid algorithm. The statement is true for the two problems. As for the pair betting type, it can be illustrated as a directed graph problem. By graph algorithm, it is showed that finding the optimal match in either indivisible or divisible pair betting is NP-hard. In the final part of the paper, the sufficient condition for the existence of a match for pair betting is stated, and the calculation is showed to be in polynomial time. In the second paper, it is showed that the general pricing problem for tournaments is NP-hard, and a polynomial-time algorithm is derived when applying appropriate restricted betting types. The whole deriving process is based on the Bayesian networks. There are several terminologies I want to make clear in these two papers. The first is the "minimum feedback arc set problem." The second is the "greedy algorithm." Can I have the descriptions of them? In addition, on page 4 in the first paper, it says: "since in the worst-case state the auctioneer needs to pay $1 to every order in the cycle except one." Why is that? How come there is an exception of one? One more question is about the Theorem 3.1 in the second paper. What does it mean? I don't fully understand the statement of the theorem. I think the idea to apply graph algorithms on betting market is very interesting, and surprisingly, each type in the paper can find its corresponding graph model to solve the matching problem. I am wondering if there is any other application of graph algorithms in market. Since graph algorithms are well developed in the field of computer science, if we can find other explicitly/implicitly related market problems, we can solve these problems efficiently. |
Brian Young |
Betting on Permutations (Chen, Fortnow, Nikolova, Pennock) Pricing Combinatorial Markets for Tournaments (Chen, Goel, Pennock) We have discussed general prediction markets in previous classes; problems we have seen include betting on conditional outcomes, which tends to cause a great deal of blowup in the amount of space required, as well as in complexity of finding a match. Both papers for Monday focus on specific areas in which conditional betting might be more feasibly computed, specifically, permutations and tournaments. Permutation betting is appropriate for situations in which we might wish to determine the relative orders of candidates. The paper discusses two possible questions we might ask: in which positions could a given candidate wind up in (subset betting), or which of two candidates will finish better in relation to one another (pair betting). Surprisingly, attempting to match traders in subset betting, which intuitively seems to involve more options, actually reduces to a polynomial-time-solvable problem, while doing the same for pair betting proves less computationally tractable; finding the optimal pair-betting match is NP-hard, while a greedy algorithm achieves only a O(n^0.5) approximation. Tournament betting is based on the situation of competitors playing single-elimination, head-to-head matches until only one remains. The paper models this using market scoring rules and attempts to figure out how difficult it is to determine the price of any particular security. Although doing this for general betting is #P-hard, we can restrict the permissible bets such that the new problem is tractable. I found the conclusions reached by both papers fascinating; my questions mostly relate to how reasonable it is to examine these particular types of scenarios. How generalizable are these results? Might there be some way to convert more general cases into instances of permutation- or tournament-style scenarios? Most examples that we have seen of the use of prediction markets involve a set of outcomes, only one of which can actually occur; for many of these, it seems unnatural to try to assign places or ordering to the remaining cases. (Reasoning from the suggestion of elections, some fairly natural cases suggest themselves: the relative popularity of various options, for example, or the percentage of something, be it votes or money or otherwise, assigned to the options.) As for tournaments, outside of sports and similar competitions, I can think of very few situations that involve this kind of progressive elimination. However, can we generalize these results beyond single-elimination tournaments to more complicated formats? |