Integrating Singular Spectrum Analysis with Deep Learning Algorithms for Australian Stock Forecast (Data and Code)

Abstract:

Stock price prediction is challenging as with full of nonlinear relationships, hence many people apply deep learning to predict stock prices and construct trading strategies to gain benefits. However, due to the problem of low signal-to-noise ratios of financial time series, deep learning can easily fail. This paper proposes a model combining Singular Spectrum Analysis (SA) and Deep Learning (DL), targeting stocks on ASX 50 in the Australian market. First, stock prices are decomposed into valuable sequences and eliminating noise, and then multiple DLs are trained using the denoised series to make forecasts and construct appropriate trading strategies. The experimental results show that the SSA-DL models can uncover valid information in stock prices, construct denoised stock price prediction data, and obtain better prediction results as well as investment returns. The best model obtained in this study was SSA-CNN-LSTM, which was able to generate a Sharpe Ratio of up to 1.88 and 67% ROI.

The submitted file with data and code is for review purpose.