PatchTST for Trading: Patching Price and Volume Sequences
Summary
The article presents a MetaTrader 5 trading system based on PatchTST, a transformer-style model that divides time series into patches to represent local patterns while modeling relationships across a longer history. Its described input uses normalized price changes and log-scaled tick volume as separate channels. The implementation also addresses class imbalance by weighting rare market states more heavily and maps model predictions into trading decisions. The article argues that this patch-based design can reduce attention's computational burden compared with processing every bar as an independent token.
The author reports a historical test on EUR/USD M15 over July and August 2025, including profitability, drawdown, Sharpe ratio, win rate, and other trade statistics. These are results from a short backtest, not independent or forward validation; the article also says longer testing is computationally expensive. It provides no basis for assuming the reported performance will persist across market regimes or instruments. The author notes resource demands, dependence on good historical data, exposure to abrupt market changes, and possible decay as conditions evolve.
Key ideas
- PatchTST groups consecutive bars into patches to represent local movement patterns and longer temporal relationships.
- The described model combines normalized price change and tick-volume inputs.
- Class weighting is used to give uncommon market states greater influence during training.
- The article reports a short EUR/USD historical backtest, which does not establish durable performance.
- Compute requirements, data quality, market shifts, and strategy decay are stated limitations.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.