Are betting returns a useful measure of accuracy in (sports) forecasting?

In an economic context, forecasting models are judged in terms not only of accuracy, but also of profitability. The present paper analyses the counterintuitive relationship between accuracy and profitability in probabilistic (sports) forecasts in relation to betting markets. By making use of theoretical considerations, a simulation model, and real-world datasets from three different sports, we demonstrate the possibility of systematically or randomly generating positive betting returns in the absence of a superior model accuracy. The results have methodological implications for sports forecasting and other domains related to betting markets. Betting returns should not be treated as a valid measure of model accuracy, even though they can be regarded as an adequate measure of profitability. Hence, an improved predictive performance might be achieved by carefully considering the roles of both accuracy and profitability when designing models, or, more specifically, when assessing the in-sample fit of data and evaluating out-of-sample forecasting performances.