International Symposium on Nonlinear Theory and its Applications
Ensembling via Prediction Market Mechanisms for Time-series Forecasting
Christian Merkwirth, Maciej J. Ogorza?ek,
PDF download (100.5KB)
In this paper we propose the use of prediction market mechanisms for time series forecasting. Prediction markets give predictions much closer to real outcomes than polls. F. Hayek proposed in 1945 that the market acts as an information aggregation mechanism. While prediction markets were recently applied by economists to predict the outcome of elections or the probability of a single event, our proposed mechanism is intended to improve the aggregation of a series of forecasted values to an ensemble forecast. Prediction market approaches assume that all information about a security is reflected in its price. It is expected that in a similar manner the output of the participating models in an ensemble could be aggregated via prediction market mechanisms instead of simply pooling results. Having to bet a certain amount of credit points for each forecasted value makes the creator of a model more aware of the specific strengthes and weaknesses of the participating model. Market mechanisms can thus be used for building useful ensembles of models which might outperform classical aggregation methods.