DUET Channel Clustering with Frequency-Domain Distances

Summary

This article presents a practical implementation of DUET’s channel clustering stage for multivariate time-series forecasting. It transforms each channel with a Fourier method, compares frequency-amplitude vectors, normalizes the pairwise distances, and uses the resulting relationships to create a channel mask. The article also explains how the module fits into DUET alongside temporal clustering, feature fusion, and a forecasting network.

The implementation simplifies the proposed approach: it uses vector distances instead of a learnable Mahalanobis metric, and its masking calculation has no trainable parameters or backpropagation. The article reports an out-of-sample trading test with 53 trades, more than 56% profitable, and a profit factor of 2.44. These results are an example from the author’s implementation, not evidence of broad or repeatable performance. The conclusion calls for training on more representative data and comprehensive testing before live use.

Key ideas

• DUET combines temporal clustering and channel clustering to represent multivariate time-series patterns.
• Frequency-domain amplitude vectors are used to compare relationships between channels.
• The described implementation normalizes pairwise distances and uses them to create a channel mask.
• The implementation substitutes simple vector distances for a learnable Mahalanobis metric.
• The reported trading results come from one out-of-sample test and need broader validation.

Tags

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