Hidformer’s Dual-Tower Transformer for Financial Time-Series Forecasting

Summary

The article introduces Hidformer, a Transformer-style architecture designed for time-series forecasting, and describes its dual-tower encoder. One encoder analyzes temporal patterns, while another works in the frequency domain to identify dependencies and filter noise. The sequence is divided into subsequences that are merged through processing, and the proposed attention mechanisms include recursive attention for temporal data and linear attention for frequency analysis. A multilayer perceptron decoder predicts a full sequence in one step, which the article says avoids error accumulation from sequential forecasts.

The practical section discusses an MQL5 interpretation of parts of the approach, including attention with different context windows for separate heads, and implementation work using OpenCL. The text explains the intended computational and modeling benefits but supplies no quantitative trading results or forecasting benchmarks in the material shown. It also states that the implementation is incomplete and will continue in a later article. Accordingly, the architecture’s claimed advantages should be treated as design rationale rather than demonstrated investment performance.

Key ideas

• Hidformer uses separate encoder towers to process temporal sequences and frequency characteristics.
• Its temporal attention uses context windows that vary across attention heads.
• The proposed decoder forecasts a sequence in one step using a multilayer perceptron.
• The article describes an MQL5 and OpenCL implementation of selected model components.
• The presented material does not report quantitative forecasting or trading performance, and the implementation is unfinished.

Tags

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