LLM-Trading

The problem

Most retail strategy research overfits a single train/test split, or leaks future information into “causal” features without anyone checking. This is a research pipeline for XAUUSD minute data built to make both mistakes hard to make by accident: a sealed test split, an explicit leakage checker, and — once a strategy graduates to reinforcement learning — walk-forward validation instead of one backtest window.

What I built

A pipeline from raw MT4/MT5-style M1 ticks to a trade-ready policy:

• Data handling — loads and validates raw M1 data, parses broker/EET time safely, and resamples M1 execution data into H1 decision bars while keeping M1 for intrabar TP/SL simulation.
• Feature set — 25 causal, ATR-normalized features. Four redundant ones were dropped after a collinearity audit, each correlated |r| ≥ 0.96 with a feature already retained.
• Baseline — a simple trend-following policy tuned on train/validation splits with the test split sealed until the end.
• RL agent (optional) — a PPO agent with a MultiDiscrete([direction, SL bucket, TP/R bucket]) action space, trained with sliding walk-forward validation: each fold trains on 5 years, selects its checkpoint on the next 6 months, and is judged out-of-sample on the 6 months after that — simulating “retrain every 6 months, trade the next 6 live.” Roughly 35 folds over the ~23-year dataset, stitched into one continuous out-of-sample equity curve.
• Leakage checks — a future-append stability test the pipeline runs before a feature set is trusted.
• Holdout reveal — a separate final_holdout_eval.py entry point whose only job is revealing the sealed test split once, after the model is frozen.
• Dashboard — a Next.js dashboard (replacing an earlier Streamlit one) for viewing pipeline output, deployed on Railway.

Pipeline flow

Each stage below is a real file in the repo, in run order:

Stack & approach

Python for the pipeline (config.py, data_loader.py, features.py, a Gymnasium-compatible bracket-trading environment in env_bracket.py, train_ppo.py), a guided Jupyter notebook for walkthroughs, and a TypeScript/Next.js dashboard for results. When TP and SL land in the same M1 candle, the simulator assumes SL first — deliberately pessimistic. Position size is fixed-fractional and risk-based; the RL agent controls direction and bracket shape, not size.

Split boundaries carry a 200-bar embargo on each side — sized to EMA-200, the longest lookback in the feature set, which still retains about 37% of a bar’s weight after 200 steps, so nothing upstream of a boundary can leak into validation or test. And getting a model into production isn’t automatic: config.py defines a deployment gate that requires a minimum number of walk-forward folds to clear a profit-factor floor, sets a hard floor under the worst individual fold, and checks that mean out-of-sample Sharpe is positive. Fail any of it and the run is quarantined under models/walk_forward/ instead of promoted — the gate can only block a bad deploy, it can’t manufacture a good one.

Status / results

Actively developed research pipeline. The holdout hasn’t been revealed on a frozen model yet, so no Sharpe ratio, win rate, or walk-forward return numbers are published here — none exist in the repo to report. This section will be updated with real results once a holdout run is finalized and worth standing behind.