Why Directional Accuracy Fails to Translate into LLM Trading Profit
Summary
The article diagnoses a gap between a language model’s ability to classify future price direction and its ability to make profitable trades. It describes balancing upward and downward examples, building a feature set from common technical indicators, and fine-tuning an LLM with structured prompts. The author stresses that class balance and predictive accuracy do not define an economic edge: labels based only on future direction omit trade costs, position decisions, and whether a forecast is worth acting on.
The reported prototype has strong backtest figures, but the article acknowledges that performance can deteriorate after spread, swaps, slippage, and market-regime changes. It does not provide convincing live evidence; demo or micro-account validation is presented as a next step. Proposed changes include labels tied to expected PnL, explicit cost modeling, an option to abstain from trading, and evaluation with trading measures such as risk-adjusted returns and drawdown rather than accuracy alone.
Key ideas
- Balanced UP and DOWN samples can address class imbalance but do not by themselves create a tradable edge.
- The feature set combines volatility, momentum, trend, volume, and price-channel indicators.
- A direction label does not capture transaction costs, trade selection, or the magnitude of potential gains and losses.
- The article reports promising backtest figures but says live or demo validation with realistic costs is still needed.
- Training targets and evaluation should focus on expected trading outcomes and allow the model to choose no trade.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.