Hybrid Linear and U-Transformer Forecasting on Model Residuals

Summary

The article describes a two-stage forecasting system for financial time series. A linear autoregressive model uses a 25-feature input set, including price lags, transformed values, indicators, and cyclical time features. A U-Transformer then models the linear model’s residuals, combining an encoder-decoder structure with skip connections and self-attention. Its correction is adaptively weighted alongside the linear forecast, with the linear model intended to dominate when the neural component is unreliable.

The article reports experimental advantages over purely linear methods and describes real-time trading logic, periodic re-optimization, and position management. It does not provide enough detail here to assess the experiments’ scope, data, or statistical significance. The implementation also has stated limits: simplified backpropagation, no modern regularization, a fixed network structure, and static arrays. The approach is a proposed modeling framework, not evidence that the forecasts will be profitable out of sample.

Key ideas

• A linear autoregressive model first captures baseline patterns from an extended feature set.
• The U-Transformer is trained to estimate residual dependencies left by the linear model.
• An adaptive weighting scheme adjusts how much the neural correction contributes to the combined forecast.
• The architecture adapts U-Net skip connections and Transformer attention to time-series inputs.
• The article acknowledges limited regularization, simplified training, and fixed network capacity.

Tags

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