Comparative analysis of traditional time series, machine learning, deep learning and hybrid models for profit forecasting in financial markets

Abstract

This thesis compared the forecasting performance of traditional time series, machine learning, deep learning and hybrid models on daily banking profit data in financial markets area aggregated on three different levels. To evaluate different methods, this thesis used a novel performance metric - corrected mean average scaled error (cMASE), which improves interpretability of MASE by using T one-step naive forecasts instead of T −1, which results in naive method always having a score of cMASE = 1. Despite advancements in computational power, traditional time series method SARIMA still outperformed other models, also showing the most consistent results between average cross-validation cMASE and testing cMASE. For best hybrid models, gradient boosting methods complemented SARIMA by correcting forecasts using long lags, rolling means and standard deviations. While SARIMA models required refitting after every forecast, the machine learning, deep learning and non-linear parts of hybrid models performed best when refit only on average once every two weeks, which reduced the overall computing cost significantly.