DFT-Triangle: A Novel Frequency Spectrogram Feature for Stock Price Prediction

Discrete Fourier Transform(DFT) was popular for extracting frequency feature from price series. However, it has been observed that two similar price series exhibit distinct features in the frequency domain, leading to confusion in time-frequency analysis and learning. Secondly, as common knowledge suggests, it is believed that more recent prices have a greater impact on future price predictions in the stock market. Inspired by these points, this study proposed a novel frequency-domain feature of price series called DFT-triangle, which functions as a novel spectrogram for transfer learning and stock forecasting. By stacking price series into 2-dimensional matrices with transferred frequency features, DFT-triangle is formed within a scale-frequency coordinate system, where scale refers to the size of sub-sequences generated from the current price backward to each of the previous ones, hence the name DFT-triangle. Experiments on entire 470 stocks were then carried out for evaluating the power of DFT-triangle in stock prediction, based on neural networks. It was noted that DFT-triangle with a single-layer CNN produced promising results in stock prediction, compared to advanced sequential networks like Attention-LSTM and CNN-LSTM, as well as traditional econometric predictors such as AR and ARIMA.