International Journal of Data Science and Analytics Информационные системы
Wavelet transform in conjunction with temporal convolutional networks for prediction of stock continuous trend
Аннотация
Stock-trend prediction requires a clear definition of the target and an appropriate representation of price dynamics. This study presents DCTTCN, a framework combining discrete wavelet transform (DWT), continuous-trend labeling, and a temporal convolutional network (TCN). DWT preprocessing is followed by labeling based on turning points and a relative-price-change threshold. The labels describe trend intervals that can include local price movements in the opposite direction. The TCN learns temporal patterns from historical closing-price windows to predict these interval labels. The framework was evaluated on a public dataset comprising 400 stocks. DCTTCN achieved an average test accuracy of 72.85%, compared with 71.89% for DAELM, corresponding to an absolute improvement of 0.96 percentage points. The comparisons demonstrate competitive prediction performance under the adopted continuous-trend definition. A trading-strategy evaluation further illustrates the application of the predicted labels. The framework integrates wavelet-based representation construction with temporal modeling for stock-trend classification.