Does the consideration of market prices in model selection increase model profitability? Evidence from theory, artificial data and real-world data
研究了在模型选择中考虑市场价格（而非仅预测精度）能否提高盈利能力，通过理论证明和体育博彩数据验证，发现盈利优化模型能隐式检测并应对市场偏差，从而提升样本外收益。
It is well-acknowledged in the literature that besides accuracy, profitability (i.e. the economic value) of a forecast is crucial in domains including economics, energy, weather or sports forecasting. The present study argues that selection of profitable forecasting models benefits from consideration of market prices in-sample (i.e. when optimizing model parameters on training data). Such models can be specifically optimised for profitability instead of accuracy, as most commonly done so far. Previous literature suggests that, under specific circumstances, accuracy- and profitability-optimisation leads to equal model selection. We theoretically show that this result is contingent upon strict assumptions and does not hold anymore if relaxing any of these assumptions. Moreover, we present evidence that out-of-sample profitability (i.e. returns generated on unseen data) are actually increased for profitability-optimised models compared to accuracy-optimised models based on artificial and real-world data from the domain of sports betting. This result is explained by the ability of profitability-optimised models to implicitly detect and counter an existing market bias, provided that the model family allows to capture such biases. The present study intends to stimulate discussion on optimal model selection for profitable forecasting and to encourage researchers to consider profitability-optimisation in forecasting applications to real-world data.