ACEFormer for Trading: Adaptive Decomposition and Probabilistic Attention

Summary

The article describes a trading-oriented implementation of ACEFormer, combining adaptive empirical mode decomposition with a Transformer forecasting architecture. ACEEMD is used to reduce high-frequency noise while retaining turning points; the resulting modal representation feeds a distillation stage with probabilistic attention, followed by self-attention for broader sequence context and a prediction head. The article focuses on integrating these mechanisms into the main program and building a modular attention component, including query and key selection and gradient handling.

The implementation is integrated into an Actor–Director–Critic training setup. The authors report a positive return on an out-of-sample test period, but also note low trading activity and suggest that probabilistic attention or limited training-data diversity may contribute. The excerpt offers no detailed performance statistics or broad comparisons, so the reported result is limited evidence of viability rather than proof of general forecasting or trading performance. The model’s practical value depends on the data, training design, and evaluation setting.

Key ideas

• ACEEMD decomposes a financial series into adaptive modes and targets high-frequency noise while preserving turning points.
• Probabilistic attention selects informative queries using sampled keys, while a separate self-attention block captures broader sequence context.
• The article presents a modular MQL5 implementation integrated into an Actor–Director–Critic training architecture.
• The authors report positive out-of-sample returns but also describe trading activity as low.
• The reported experiment does not establish that the architecture will generalize across datasets or market conditions.

Tags

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