DA-CG-LSTM Attention and Reinforcement Learning for Trading Decisions

Summary

This article presents an MQL5 implementation of a trading model built around DA-CG-LSTM, which combines feature attention, temporal attention, and a modified recurrent block. The attention stages weight input variables and historical intervals; the recurrent component then aggregates information to produce a forecast. The author also describes integrating this encoder with global and local market skills, a probabilistic trend model, and an Actor-Director-Critic reinforcement learning setup that selects trading actions while considering account state and risk.

The reported evaluation uses data outside the training period. The profitable-trade share was slightly above 40%, while average gains relative to average losses were said to yield a positive overall outcome. The article characterizes the models as exploratory: it does not provide enough detail here to establish robustness, and it calls for broader datasets and testing across varied market conditions before live use.

Key ideas

• DA-CG-LSTM uses one attention stage to weight features and another to weight historical intervals.
• A modified recurrent block is intended to combine information across features and time while filtering noise.
• The proposed system joins forecasting components with an Actor-Director-Critic framework for trading decisions.
• An out-of-sample test is reported as positive overall despite a profitable-trade share slightly above 40%.
• The author describes the system as exploratory and recommends broader evaluation before live deployment.

Tags

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