RefMask3D Object Clustering for Neural Trading Models

Summary

This article completes a trading-oriented implementation of RefMask3D, a language-guided 3D point-cloud segmentation framework. It focuses on the Object Cluster Module, which combines two self-attention blocks, an intervening cross-attention block, and a feed-forward network to aggregate object information. The described inputs include point-cloud-informed linguistic primitives and embeddings of the target description. The article explains the module’s structure and implementation in MQL5, building on earlier work on geometric attention and linguistic primitives.

The author reports a 1.57% performance improvement from adding the Object Cluster Module and describes a trading test in which profitable trades outnumber losing trades in both directions, with winning trades larger on average. The text gives limited detail about the test setup and acknowledges that the small number of trades prevents firm conclusions about long-term effectiveness. It also states that the programs are demonstrations and are not ready for live trading.

Key ideas

• The Object Cluster Module combines self-attention, cross-attention, and a feed-forward stage to aggregate object embeddings.
• Its cross-attention uses the target description embeddings as queries and decoder output as context.
• The article reports a performance gain from adding the module, but does not establish long-term trading effectiveness.
• The implementation is presented as a research demonstration rather than a live trading system.

Tags

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