Real-money vs. play-money forecasting accuracy in online prediction markets - empirical insights from Ipredict

Prediction markets are online trading platforms where contracts on future events are traded with payoffs being exclusively linked to event occurrence. Scientific research has shown that market prices of such contracts imply high forecasting accuracy through effective information aggregation of dispersed knowledge. This phenomenon is related to incentives for truthful aggregation in the form of real-money or play-money rewards. The question whether real- or play-money incentives enhance higher relative forecast accuracy has been addressed by previous works with diverse findings. The current state of empirical research in his field is subject to two inherent deficiencies. First, inter-market studies suffer from market disparities and differences in the definition of underlying events. Comparisons between two different platforms (one for play-money contracts, one for real-money contracts) are potentially biased by different trading behaviour. Second, the majority of studies are based upon identical datasets of market platforms (IOWA stock exchange, Tradesports/Intrade, NewsFutures). This paper contributes new insights by analysing 44,169 trading observations on ipredict, where real-money and play-money contracts are traded on a variety of events. Forecasting accuracy is analysed on overall trading activity as well as comparison of equal contracts under different monetary incentive schemes. Statistical models are built to analyse the influence of order volumes and days to expiry under both incentive schemes. Ignoring different events in underlying trading activity, play-money contracts imply statistically insignificant excess accuracy. In direct comparison of equal events, real-money contracts, however, real-money contracts predict at significantly higher accuracy. This paper finds a relationship between order volumes and forecasting accuracy whereas the influence of days to expiry and aggregated volumes showed lower R² than was expected by formed hypotheses.