Hierarchical Vector Transformers for Market Forecasting and Trading

Summary

This article completes an implementation of the Hierarchical Vector Transformer (HiVT), adapting a model originally designed for multi-agent motion prediction to financial time-series forecasting. The architecture builds local representations for individual series, models interactions across those local regions, and decodes the combined representation into multiple possible future trajectories with associated probabilities. The implementation combines embedding, attention, temporal encoding, and trajectory decoding components in an MQL5 neural network.

The model is integrated into an environmental state encoder and trained for an agent-based trading setup. The reported test had 39 trades, a win rate slightly above 43%, and a profit factor of 1.22; average and maximum gains per trade exceeded corresponding losses, yielding a small net profit. The author cautions that the balance curve lacked a clear trend and the trade count was limited. These results therefore offer preliminary evidence only, not proof that the model will generalize or support a durable profitable strategy.

Key ideas

• HiVT separates forecasting into local feature extraction, global interaction modeling, and trajectory decoding.
• The implementation combines attention and temporal components to represent interactions among time-series variables.
• The model generates multiple candidate future trajectories and associated probabilities.
• A trading test reported a small net gain and a profit factor of 1.22 across 39 trades.
• The limited trade sample and unclear balance trend make the reported result inconclusive.

Tags

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