The Relative Accuracy of LMSR and CDA Prediction Markets

Abstract

Prediction markets are a highly successful forecasting method, and they have outperformed other methods (polls, regressions, etc.) in a variety of settings. Most markets are run with one of two mechanisms: the continuous double-auction (CDA), which resembles a stock market; and the logarithmic market scoring rule (LMSR), which has a market maker and a cost function to determine price. While a good deal is known about various benefits and drawbacks to each mechanism, relatively little is known about how they compare in terms of accuracy. We use trader belief data from a set of CDA markets to simulate the corresponding LMSR markets on those same events, and find that the two mechanisms generally have statistically similar performance on average. This holds true for a variety of parameter settings for the LMSR. However, which mechanism is superior in a given market is heavily dependent on liquidity. Leveraging this fact, we propose a hybrid algorithm that combines the predictions of the two mechanisms in a liquidity-dependent way. This algorithm is able to obtain better average accuracy in many cases.