Multitask-Stockformer: Wavelet Decomposition and Dual-Frequency Attention

Summary

This article describes a practical MQL5 adaptation of Multitask-Stockformer, a neural architecture for financial time series. It first outlines the framework’s use of discrete wavelets to split signals into low-frequency trends and high-frequency fluctuations, then processes the streams with temporal self-attention and causal convolutions. Graph representations are used to model temporal structure and relationships between assets, before a fusion decoder combines the streams for tasks such as return forecasting and trend-change estimation.

The implementation simplifies the original graph attention design, using trainable positional embeddings for asset relationships and parameter-free Node-Adaptive Feature Smoothing modules. Self-attention and cross-attention fuse long- and short-term information. The article reports training on historical data and testing in MetaTrader 5, but the excerpt omits most experimental details and quantitative results. Its conclusion describes the results as promising potential rather than proof of trading effectiveness, and recommends more representative data and comprehensive testing before live use.

Key ideas

• Wavelet decomposition separates long-term structure from short-term fluctuations for distinct processing.
• Self-attention and causal convolutions analyze low- and high-frequency components, respectively.
• Graph-based representations are intended to capture temporal links and relationships among assets.
• The implementation simplifies graph attention with positional embeddings and parameter-free feature smoothing.
• The authors recommend broader data and testing before considering real trading.

Tags

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