A Comparative Study of Machine Learning and Neural Network Models in Short-term Market Prediction

Abstract The prediction of the stock market and the prices of other commodities like crude oil, 1 constitutes a challenging task. Recently, the rapid progress in the field of Machine Leaning (ML), 2 led to an increased interest in applying ML techniques to market price predictions. In this study, 3 we conduct a comprehensive comparative analysis of the performance of 15 different ML models in 4 predicting the close prices of crude oil futures. These models include an Auto-Regressive Integrated 5 Moving Average (ARIMA) model, the Meta Prophet Library, a simple Recurrent Neural Network 6 (RNN), a Long-Short Term Memory (LSTM), a Gated Recurrent Unit (GRU), a Bi-directional LSTM 7 (BLSTM), a Bi-directional GRU (BGRU) and a number of hybrid models, including an LSTM-BGRU 8 and a BLSTM-BGRU. In addition, we evaluate the effectiveness of using Denoising Autoencoders 9 (DAE) in enhancing the performance of the networks by evaluating hybrid models with DAE layers, 10 including DAE-LSTM, DAE-BLSTM, DAE-GRU, DAE-BGRU, DAE-LSTM-GRU and DAE-BLSTM- 11 BGRU. In our analysis, we focus on short-term predictions dedicated for day trading. We compare 12 the performance of these models using the Root Mean Square Error (RMSE), Mean Absolute Error 13 (MAE), Mean Absolute Error Percentage (MAPE) and the R2. We find that the BGRU model yields the 14 best performance, with the GRU model not far behind. We also find that hybrid models containing 15 BGRUs or GRUs tend to perform better than other hybrid models. We find mixed evidence vis-a-vis 16 the effectiveness of DAEs in improving the performance of networks, where in some models the 17 performance improves, whereas for others the performance deteriorates. We find that the performance 18 of all models deteriorate when predicting sharp and sudden changes in prices. The key takeaway of 19 this study is that BGRUs, and to a lesser extent GRUs, seem to be the best route to follow in applying 20 AI/Ml to the task of market prediction.