LLM model node

Large language model reasoning over factors and news.

Where it sits in the pipeline

This node accepts one upstream connection and passes its output downstream. Connect it between the data source and the trading logic block.

Output

The model writes a score between -1 and 1 into the output column. Use it in conditions such as llm_score > 0.45.

Training

• Target: usually label_ret from the processing node (forward return; the sign is the class).
• Walk-forward: at each bar the model is fitted only on rows whose label is already known, then predicts the next `retrain every` bars. Nothing after the current bar enters a fit.
• Features: every upstream numeric column except raw OHLCV and the label (or the ones you tick). Mostly-NaN and constant columns are dropped.
• Backend: the Python sidecar uses LightGBM → scikit-learn → NumPy logistic regression, whichever is installed; without Python an in-process logistic regression runs. The report states which one was used.
• Diagnostics: out-of-sample sign accuracy and IC (rank correlation) are reported per symbol. Accuracy ≈ 50% and |IC| < 0.02 means no edge.