Reducing Trajectory Model Cost with Social Attention and Map Priors

Summary

The article adapts lightweight motion-prediction methods from autonomous driving to market trajectory forecasting, seeking lower inference cost without sacrificing forecast quality. It describes encoding agents’ relative historical movements with an LSTM, modeling interactions with a graph neural network and multi-head self-attention, and decoding future paths autoregressively. A simplified map prior is generated through preprocessing rather than supplied as a high-resolution scene representation. The method estimates current velocity and acceleration by fitting recent observations, then filters and perturbs plausible centerlines to form candidate paths.

The article reports experiments on a trading model trained using seven months of historical data: it produced profits over at least three subsequent months but completed only 11 trades, which the author considers too few. The small sample and limited trading frequency make the result weak evidence of robustness or general profitability. Its main contribution is a model-efficiency design and an initial application to trading, while the reported performance remains preliminary and dependent on the chosen data and setup.

Key ideas

• Relative agent trajectories are encoded before graph and attention layers model interactions.
• A simple preprocessed map prior can replace high-resolution map input and reduce model demands.
• Least-squares filtering estimates recent velocity and acceleration to select plausible future paths.
• The reported trading evaluation spans seven months of training data and at least three profitable months afterward, with only 11 trades.
• The sparse trading sample limits conclusions about performance stability.

Tags

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