Adaptive Bottlenecks and Dual Decoders for Market Anomaly Detection

Summary

The article implements a version of the DADA time-series anomaly detection framework, focusing on its adaptive bottleneck module. The design uses multiple autoencoder experts with different latent compression sizes. A gating mechanism selects the most relevant experts for each input, while patching and masking are described as ways to emphasize local patterns and train the model to reconstruct hidden data. The implementation adapts an existing mixture-of-experts structure in MQL5 and uses convolutional components to form and reconstruct the latent representations. The broader DADA architecture also includes separate decoders for typical and anomalous behavior.

The article reports that models trained on real historical data were profitable in testing, but their equity curve did not rise consistently. It therefore treats the outcome as preliminary and says more refinement is required. The implementation includes modifications to the original framework, so results apply to this version rather than necessarily to DADA in general. The text offers an architectural account and testing summary, but the available material does not establish robust performance across markets or regimes.

Key ideas

• The adaptive bottleneck uses autoencoder experts with different latent dimensions to vary compression.
• A top-k gating mechanism selects which experts process an input segment.
• The implementation reuses mixture-of-experts infrastructure and adds convolutional layers for encoding and reconstruction.
• Patching and random masking are presented as tools for learning local structure and latent dependencies.
• The reported tests showed profitability but an unstable equity curve, and the implementation differs from the original framework.

Tags

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