Temporal Query Networks for Multivariate Market Forecasting

Summary

The article presents Temporal Query Network (TQNet), a neural architecture for forecasting multiple related time series. Its attention layer forms queries from trainable vectors that are shifted periodically through time, while keys and values come from the current input sequence. This design aims to combine stable, learned cross-variable relationships with dependencies specific to the current observation. A shallow feed-forward network and residual connections further process the representation before producing forecasts across a chosen horizon.

The article describes the temporal query mechanism and reports a comparison of conventional data-derived attention, TQNet’s hybrid approach, and an approach using only global trainable representations. On large multivariate datasets averaged across four forecast horizons, the hybrid method performed best, the conventional method followed, and the global-only version ranked last. These findings are attributed to the framework’s authors and do not establish trading profitability or reliable performance on live financial data. The article’s practical section focuses on implementing and organizing model components; historical market testing is deferred to subsequent work.

Key ideas

• TQNet targets forecasts of several related variables over a chosen horizon.
• Periodic shifts of trainable query vectors encode recurring, learned relationships across variables.
• Using current input data for keys and values preserves information about the current context.
• The reported benchmark comparison favored combining global learned queries with local input data.
• Forecast accuracy comparisons alone do not demonstrate profitable trading or robustness in live markets.

Tags

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