Wavelet and Multi-Task Attention Models for Stock Selection

Summary

This article introduces Multitask-Stockformer, a stock-selection framework that combines wavelet decomposition with temporal, convolutional, and graph attention components. A discrete wavelet transform separates historical asset data into low-frequency features associated with longer-term movement and high-frequency features associated with shorter fluctuations and sharp events. The model processes these streams through a dual-frequency spatiotemporal encoder and decoder, then produces both return forecasts and trend-classification probabilities. The article also describes an MQL5 implementation, beginning with a convolution-based layer that uses fixed Legendre wavelet filters to extract both frequency components.

The discussion explains the architecture and implementation choices, including disabling weight updates for static wavelet filters. It does not present training results or a trading performance evaluation: the author says implementation and testing on historical data will continue in a later article. The stated motivation is that wavelets retain signal structure as well as frequency information, but the excerpt does not establish that this model predicts returns accurately or produces profitable portfolios.

Key ideas

• The framework decomposes market data into low- and high-frequency components with a discrete wavelet transform.
• Separate temporal and convolutional processing paths encode the two frequency bands before graph attention models relationships across assets and time.
• The decoder combines both streams and generates return forecasts alongside trend probabilities.
• The implementation uses static Legendre wavelet filters and disables optimization of those filter parameters.
• The article describes model design and implementation but supplies no empirical forecast or portfolio results.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.