Mantis Feature Extraction with an Actor–Director–Critic Trading Agent

Summary

This article describes using the Mantis time-series architecture as a market representation module within a trading agent. It processes separate market-data channels and their first differences, extracts local features with convolutions, pools them into temporal patches, and applies attention to connect information across segments. The resulting embedding feeds an Actor–Director–Critic design: the Actor proposes actions, the Director filters or corrects them, and the Critic evaluates their strategic value.

The article also discusses training the encoder with self-supervised contrastive learning and augmenting examples with controlled noise. It reports a test over January through March 2025 and says the implementation was profitable, while noting that some parameters need improvement. The provided account does not give enough detail here to assess robustness, costs, or out-of-sample reliability, so the reported result is an initial experiment rather than proof of durable performance.

Key ideas

• The Mantis encoder combines per-channel inputs, first differences, convolutional features, temporal patches, and attention.
• Its embedding serves as the perception input to an Actor–Director–Critic trading architecture.
• The Director filters proposed actions, while the Critic assesses their value in context.
• Contrastive learning with augmented examples is used to train the encoder representation.
• The reported test was profitable, but the article acknowledges further tuning and provides limited evidence of generalization.

Tags

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