Multi-Future Transformers for Joint Multimodal Trajectory Forecasting

Summary

The article explains the Multi-future Transformer, a neural architecture for predicting multiple plausible futures for several interacting agents. It frames future uncertainty as a mixture of unimodal scene outcomes, separating distinct interaction modes instead of asking multiple output heads to decode alternatives from one shared feature representation. Shared encoders process observed motion and scene context; parallel interaction blocks model mode-specific agent-to-agent and agent-to-context relationships; prediction heads produce trajectories and confidence estimates for agents and whole scenes.

The authors’ approach uses attention mechanisms and scene-level winner-take-all training to reduce interference between modes while maintaining joint scene predictions. The article reports implementation and testing in MetaTrader 5, with 13 trades, six profitable closures, and a profit factor of 1.63 in the stated test period. That small reported result is not enough to establish robust trading performance, and the method originates in multi-agent behavior prediction rather than being a trading-specific forecasting model. The supplied text omits much of the practical section.

Key ideas

• The method represents a multimodal future as several distinct unimodal scene modes.
• Separate parallel interaction blocks let each mode learn its own agent interaction pattern.
• Self-attention models agent-to-agent interaction, while cross-attention connects agents to scene context.
• Prediction heads estimate both individual trajectories and confidence at agent and scene levels.
• The reported trading test is limited, with 13 trades and no evidence of broader robustness.

Tags

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