Towards a methodology for measuring the true degree of efficiency in a speculative market

Abstract

Betting markets have drawn much attention in the economics, finance and operational research literature because they provide a valuable window on the manner in which individuals use information in wider financial markets. One question that has received particular attention is to what extent individuals discount information in market prices. The predominant approach to explore this issue involves predictive modeling to forecast market outcomes and examining empirically whether abnormal returns can be made by employing these forecasts. It is argued here that present practices to assess such forecasting models, including the use of point estimates and information, which would not be available in practice (at the forecasting stage) and failing to update forecasting models with information from the recent past, may give rise to misleading conclusions regarding a market's informational efficiency. Hypotheses are developed to conceptualize these views and are tested by means of extensive empirical experimentation using real-world data from the Hong Kong horserace betting market. Our study identifies several sources of bias and confirms that current practices may not be relied upon. A more appropriate modeling procedure for assessing the true degree of market efficiency is then proposed.

Keywords:

• forecasting accuracy
• market efficiency
• point estimates
• betting markets

Notes

1 Measuring the true degree of market efficiency requires that practical considerations for exploiting information are considered. Consequently, a day-wise rather than a race-wise jackknife is employed because the latter would imply that individual forecasting models could be built for each race. This is practically impossible given that informative odds, which form an important part of these models, are only available at most 5 min before the start of a race.

2 Equivalent to the homogeneity of variance assumption in ordinary (ie, between-subjects) ANOVA.

3 This follows from the fact that the results observed for a training-set size of 100 racing days (ie, ∼700 races) does not differ significantly from those observed for any larger setting.