DA-CG-LSTM Dual Attention for Multivariate Market Forecasting

Summary

This article explains DA-CG-LSTM, a neural architecture intended to forecast multivariate time series such as prices, volume, indicators, and news-related features. It outlines two attention stages: input attention weights features and time positions before recurrent processing, while a later temporal attention layer scores the CG-LSTM hidden states and builds a context for the forecast.

The Conversion-Gated LSTM modifies input and forget gate activations to address saturation and respond to short-lived movements while retaining longer-term information. The article describes the architecture and begins an MQL5 implementation of the CG-LSTM block, but the practical portion is incomplete and continues in a later installment. It gives no forecasting results, comparison with baseline models, or trading performance evidence, so claims of improved responsiveness and prediction should be treated as design rationale rather than demonstrated outcomes.

Key ideas

• DA-CG-LSTM applies attention to both input features and temporal positions before recurrent processing.
• A modified CG-LSTM block is designed to preserve longer-term context while reacting to short-lived signals.
• A second temporal attention layer scores hidden states and forms a context vector for forecasting.
• The article presents architectural explanations but no empirical forecasting or trading results.
• The MQL5 implementation is partial and intended to continue in a later article.

Tags

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