DUET Dual Clustering for Multivariate Time Series Forecasting

Summary

The document introduces DUET, a neural forecasting framework for multivariate time series that clusters information along both time and variable channels. Its Temporal Clustering Module groups temporal patterns and uses a mixture of encoders with noisy gating to select relevant experts. Its Channel Clustering Module uses frequency characteristics to identify related channels and reduce redundant signals. A masked attention fusion module combines the two representations before the prediction stage.

The article also outlines an MQL5 interpretation, beginning with a simple parallel encoder implementation for the temporal module. The authors of the framework are reported to have found DUET more accurate than existing methods, but this account gives no detailed benchmark figures or market-specific performance results. The practical implementation is incomplete in this installment: channel clustering, full integration, and testing on historical data are deferred to a later article. Forecasting gains therefore should not be read as evidence of profitable trading performance.

Key ideas

• DUET combines clustering across temporal patterns and input channels to forecast multivariate series.
• The temporal module uses gated selection among multiple encoders for different sequence segments.
• The channel module uses frequency-domain characteristics to find related variables and filter noise.
• A masked attention fusion stage combines temporal and channel representations before prediction.
• The MQL5 implementation presented here begins with the temporal module and defers full testing.

Tags

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