A Live Crypto Trading System Combining LightGBM and Transformers

Summary

The article outlines a real-time ETH/USDT trading framework that combines LightGBM for tabular inputs with a Transformer for sequential data. It describes WebSocket market-data ingestion, minute-bar construction, technical, statistical, volume, and order-book features, and a three-class prediction target for rising, falling, or sideways conditions. It also discusses time-series cross-validation and Bayesian optimization for LightGBM tuning, plus drift monitoring, retraining, model hot-swapping, state persistence, and computational optimizations.

The examples emphasize avoiding future information in feature calculations and treating operational reliability as part of system design. However, the article gives no verified predictive or trading results; its performance claims are not supported by a detailed evaluation. It explicitly says the presented system lacks an order-submission interface, so it is not a complete deployable trading strategy. The author advises testing before live use and notes that past performance does not ensure future results.

Key ideas

• The proposed system fuses LightGBM tabular predictions with Transformer sequence features for three-way market classification.
• Its data pipeline builds minute bars from live exchange feeds and derives price, volume, technical, statistical, and order-book features.
• The article uses time-series cross-validation and Bayesian optimization in LightGBM model tuning.
• Feature drift monitoring can trigger retraining, while hot-swapping and persistence support operational continuity.
• The system description provides no verified trading results and omits order submission.

Tags

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