Valerio Restocchi, Frank McGroarty, Enrico Gerding

Prediction markets constitute an excellent laboratory to test theories on decision making and financial markets, as they have characteristics which facilitate analysis. Importantly, they have a definite end-point at which all uncertainty is resolved. However, whereas in financial markets there has been extensive data-driven analysis of stylized facts, i.e. empirical regularities observed in most financial time series, there is no such work for prediction markets, mainly due to historical insufficiency of data. Consequently, many of the models of prediction markets lack robust empirical validation.

Using a dataset from PredictIt, consisting of 3385 prediction markets on political events, we compile a set of stylized facts for prediction markets. Our analysis consists of three main points. First, we show that percentage and logarithmic returns are inadequate to describe prediction markets, and advocate the use of raw returns instead. Second, we characterize the statistical properties of time series by analyzing changes in prices and volumes. We examine several properties, including the well-known heavy tails, volatility clustering, long memory properties, and autocorrelation of returns. Last, we characterize the temporal evolution of the favorite-longshot bias, an empirical regularity whereby state-contingent claim contracts on likely outcomes are underpriced and those on unlikely outcomes are overpriced. This important anomaly, although extensively studied in sports betting, lacks strong evidence and significant characterization in prediction markets.

Overall, we find that prediction markets behave similarly to financial markets, but with some significant differences. To account for these differences, we suggest that new models, especially agent-based, should be designed to study prediction markets. More importantly, we advocate the use of this set of stylized facts for a more robust validation of quantitative prediction markets models.