UniTraj: Generating Complete Multi-Agent Trajectories from Masked Data
Summary
The article explains UniTraj, a framework designed to handle trajectory prediction and reconstruction by treating input as a sequence with visible and missing observations. A mask marks which agent states are known, and the model aims to reconstruct observed regions while generating missing ones. Its architecture combines spatial Transformer attention with Ghost Spatial Masking to represent incomplete interaction patterns, and a bidirectional temporal Mamba encoder with a Bidirectional Temporal Scaled module to model time gaps and dependencies. A latent-variable decoder then generates trajectories.
The article adapts these ideas in MQL5 for a trading model, modifying how masks and future data are supplied so training can use future trajectories while real-time operation uses historical inputs. It reports a one-month test with 65 trades, 33 profitable, and a profit factor of 1.51, while explicitly noting that this sample is too small to establish long-term stability. The excerpt does not provide enough market, benchmark, or validation detail to judge generalization, and the model’s broader trajectory-generation claims do not by themselves demonstrate a durable trading edge.
Key ideas
- UniTraj represents varied trajectory tasks as completion of masked sequences.
- Spatial attention and Ghost Spatial Masking model interactions and missing-data patterns.
- Bidirectional temporal processing uses time-gap information to represent missing observations.
- The MQL5 adaptation uses future trajectories during training and historical inputs for real-time prediction.
- The reported one-month trading test is limited in size and cannot establish lasting performance.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.