Statistical Evaluation of Predictability Models in the Stock Market

Determining the optimal number of variables to include in a predictive model is one of the niches of the research community that analyzes trends in the stock market. The methodology of statistical analysis is based on the EP test statistics, which assesses the independence between financial revenues and their forecasts. The data used to compare models represent historical ranges of oil and gas sector stocks — LKOH and TATN, as well as the S&P500 index. The choice in favor of shares of Tatneft and Lukoil companies fully creates the applicability of the article in the work of practitioners and researchers of the Russian market. The null hypothesis of this study is that when the complexity of the model increases, the quality of prediction does not improve. Based on the results of the EP test, the proposed hypothesis is not rejected — an increase in the number of predictors does not affect the accuracy of forecasting. Judging by the obtained data, the quality of prediction is affected by an increase in the window — the historical range of data and the transition to a linear model form. Having calculated the profitability from using a trading strategy on models with the highest EP-statistic value, we have come to the conclusion that assessing the quality of prediction models can be complicated by the presence of anomalies that are not predictable by algorithms. The article is organized as follows: the introduction explains the background and motivation of the study, the literature review presents previous articles on this topic, the Data and the Methodology sections present data and methods for testing the model, the Results section answers questions, and the Conclusion section presents the conclusions of our work.