A LightGBM Workflow for Three-Class Futures Trading Signals

Summary

This project description outlines a LightGBM system for classifying the direction of futures prices over a configurable forward horizon. It processes one-minute bars across multiple contracts, builds technical and statistical features, and labels future returns as rising, falling, or sideways using static or rolling thresholds. The proposed workflow includes time-ordered train, validation, and test splits, time-series cross-validation, Optuna hyperparameter search, class weighting, and checks for overlap or leakage between training and validation data.

Predictions are turned into simulated trades with configurable position sizing, transaction costs, slippage, stop-losses, take-profits, and a minimum confidence threshold. The description lists classification measures and trading outcomes such as return, Sharpe ratio, drawdown, win rate, and profit-loss ratio, but supplies no actual results. It is a project guide rather than evidence that the model is profitable. The author advises checking data quality, tailoring settings by contract, monitoring offline before live use, managing risk, and retraining as market conditions change.

Key ideas

• The system frames futures direction over a future horizon as a three-class prediction problem.
• It combines technical indicators, statistical summaries, and lagged values as model features.
• Time-ordered validation, cross-validation, and leakage checks are included to assess model generalization.
• The backtest accounts for trading costs, slippage, stops, targets, and position sizing.
• The document describes an implementation workflow but reports no measured predictive or trading performance.

Tags

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