xgboost

Use when training, tuning, evaluating, or shipping XGBoost models -- XGBClassifier/XGBRegressor/XGBRanker sklearn API, native DMatrix/QuantileDMatrix + xgb.train, objectives and eval metrics, early stopping, categorical features, missing values, monotone/interaction constraints, custom objectives, multi-output, quantile regression, learning-to-rank, GPU (`device="cuda"`), SHAP (pred_contribs) and importance, model save/load (JSON/UBJ vs pickle), determinism, Dask/Spark, 1.x->2.x->3.x migrations, LightGBM/CatBoost choice, and especially financial price/return prediction on OHLCV candles (Binance klines) with leak-free walk-forward validation. TRIGGER when code imports `xgboost`, the user mentions xgboost, gradient boosting, GBDT, or asks to predict price, returns, or direction from market data.

methodology

1. Check xgb.__version__. This guru targets 3.x. Map any old code through the migration table before use: constructor-only early_stopping_rounds, device=, iteration_range.
2. Pick the front door. Use the sklearn estimator by default. Use native xgb.train + QuantileDMatrix(ref=) when you need memory savings, custom loops, or continuation. On the native path, slice predictions to best_iteration + 1 yourself.
3. For trading questions: choose a horizon return, direction, or quantile target; build features from bars at or before t only; run walk-forward folds with a gap of at least h; print zero-return and persistence baselines on every fold.
4. Choose the objective and constraints deliberately: quantile for risk bands, monotone where domain logic demands it, native categoricals with a policy for unseen levels.
5. Explain with pred_contribs, which sums to the margin, and check with permutation importance.
6. Ship with save_model("*.json" | "*.ubj"). Pin the version, and load with the same or a newer xgboost. Pin random_state and n_jobs.
7. Verify any claim by running the matching script from the api-checks or walkforward guide. Record the result in trained with the date and versions.

contents

• trainedlearned layer: gotchas, known bugs, fixes, measured practiced cases (3.4.1). Read first
• xgboost-docsofficial docs snapshot: parameters, prediction, sklearn, categorical, constraints, custom objectives, multi-output, ranking, saving, dask/spark, release notes 2.0-3.4
• xgboostcore API cheat sheet: two APIs, objectives for market targets, parameter order, persistence, version migration table
• price-predictionmarket playbook: targets, features, leakage checklist, walk-forward, baselines, Binance data
• walkforwardrunnable leak-free walk-forward on real Binance klines, with measured results
• api-checksrunnable checks: parity, categorical, quantile, monotone, SHAP, model IO, cross-version, determinism, objectives
• xgboost-lightgbmcommunity: XGBoost vs LightGBM patterns. Uses the 1.x fit(early_stopping_rounds) and gpu_hist
• signal-classificationcommunity: trading signal classification with walk-forward scripts
• shapcommunity: SHAP explainers, plots, and workflows

Related gurus: ml, feature-engineering, and olap.