A Comparison Between a Long Short-Term Memory Network Hybrid Model and an ARIMA Hybrid Model for Stock Return Predictability

This thesis explores the applicability of neural networks in stock return forecasts by designing a hybrid LSTM (long short-term memory) network and compares its forecasting ability with both a static LSTM network and an ARIMA hybrid model. The S&P100 stock set is employed as the prediction sample. The hybrid models use the neural network approach and frequentist method respectively to estimate Fama-French risk factors, then predict stock returns based on factor estimations that benefit from the prediction ability and computational power of the LSTM network and the ARIMA model as well as the Fama-French model’s explanatory power of returns. Better factor predictions are made by the LSTM network with a 31% reduction of Mean Squared Error (MSE) and broader ranges of estimation than the ARIMA model. Hybrid models demonstrate a better fit, resulting in more accurate predictions compared to the static LSTM network by an average of 4.6% (LSTM-FF) and 3.1% (ARIMA-FF). However, I find that the slight outperformance of the LSTM-FF hybrid model over the ARIMA-FF hybrid model is not statistically significant.