Hybrid deep learning and econometric models for forecasting equity returns

Authors:
  • INIfeanyi Martins Nwanegbo
    Ifeanyi Martins Nwanegbo
  • JDJonathan Daniel
    Jonathan Daniel
  • ANAdaeze Pauline Nwaohiri
    Adaeze Pauline Nwaohiri
Open Access
Publisher website
Google Scholar
Abstract
The ability to predict equity returns is considered one of the most common problems in investment strategy and financial economics. The traditional econometric models, such as ARIMA and GARCH, are appreciated in terms of interpretability and the rigor of the statistics but fail to represent the nonlinear market behavior as well as the abrupt shifts in the regimes and intricate interactions among macroeconomic and market-specific factors. On the other hand, deep learning models like LSTM, GRU, and attention-based networks can acquire complex temporal dynamics on high-dimensional sequences, but can be a black box and have no structured comprehension of classical models. This paper suggests a hybrid modeling system that can combine both econometric and deep learning network to maximize the prediction of equity returns. We build both sequential hybrids, in which the econometric residual is operated on by deep learning instantiations, and ensemble hybrids, in which predictions of one family are cosmetically fused with those of the other. On daily equity index returns, enhanced with technical, macroeconomic, and sentiment-based characteristics, our models are tested in terms of statistical (RMSE, MAE, MAPE), statistical (DieboldMariano) and economic (risk-adjusted portfolio returns and Value-at-Risk) performance. As well, we evaluate the model robustness under different market regimes and we perform SHAP-based interpretability analysis to determine the contribution of each input. We find that hybrid models are again accurate and effective in times of high volatility, and they offer not only a higher predictive power but also valuable information. The paper has proposed the possibility of hybrid approaches that can be used to overcome the lack of statistical rigor and computation intelligence and provide applications of importance in algorithmic trading, risk management, and portfolio optimization.