scripts/train_metalabeler_v3.py

#!/usr/bin/env python3
"""
train_metalabeler_v3.py — Binary Meta-Labeler for Jinjing V3

Trains a binary LightGBM classifier using REAL ranker predictions
(from ranzer_predictions_v3.parquet or similar).

Strategy
--------
- Load predictions parquet with columns: date, symbol, label, pred_rank, +features
- For each date, select top-K candidates by pred_rank (default top-200)
- Use 'label' column (0/1) directly as the meta-label
- Time-based train/test split (80/20 by date)
- Train binary LightGBM with 5-fold time-aware CV
- Evaluate AUC, precision, recall, confusion matrix
- Upload model to HF dataset

Usage:
    export HF_TOKEN=hf_...
    python train_metalabeler_v3.py \
        --predictions /tmp/ranker_predictions_v3.parquet \
        --output /tmp/meta_labeler_v3.txt

Output:
    - /tmp/meta_labeler_v3.txt  (LightGBM Booster text dump)
    - Uploaded to cedwyh/jinjing-shared-data/models/meta_labeler_v3.txt
    - Console: AUC, Precision@K, best threshold, confusion matrix
"""

from __future__ import annotations

import argparse
import json
import os
import sys
import warnings
from pathlib import Path

import numpy as np
import pandas as pd
from sklearn.metrics import (
    confusion_matrix,
    precision_score,
    recall_score,
    roc_auc_score,
    roc_curve,
)
from sklearn.model_selection import TimeSeriesSplit

# ---------------------------------------------------------------------------
# LightGBM may trigger many informational messages; keep output clean.
# ---------------------------------------------------------------------------
warnings.filterwarnings("ignore", category=UserWarning, module="lightgbm")
try:
    import lightgbm as lgb
except ImportError as exc:
    print(f"[FATAL] lightgbm is required — install with: pip install lightgbm")
    sys.exit(1)


# ===========================================================================
#  Constants / defaults
# ===========================================================================
HF_REPO = "cedwyh/jinjing-shared-data"
DEFAULT_TOP_K = 200          # candidates per date
DEFAULT_CV_FOLDS = 5
DEFAULT_OUTPUT = "/tmp/meta_labeler_v3.txt"
DEFAULT_PREDICTIONS = "/tmp/ranker_predictions_v3.parquet"
RANDOM_STATE = 42
TRAIN_SPLIT = 0.8            # fraction of dates for training
SCRIPT_DIR = Path(__file__).parent.resolve()
RESULTS_DIR = SCRIPT_DIR / "results"


# ===========================================================================
#  1.  Load data
# ===========================================================================
def load_predictions(path: str) -> pd.DataFrame:
    """Load the predictions parquet file. If not found locally, download from HF dataset."""
    import os
    if not os.path.isfile(path):
        print(f"  Local file not found: {path}", flush=True)
        print(f"  Downloading from HF dataset cedwyh/jinjing-shared-data/models/ranker_predictions.parquet ...", flush=True)
        from huggingface_hub import hf_hub_download
        token = os.environ.get("HF_TOKEN")
        path = hf_hub_download(
            repo_id="cedwyh/jinjing-shared-data",
            filename="models/ranker_predictions.parquet",
            repo_type="dataset",
            token=token,
        )
        print(f"  Downloaded to cache: {path}", flush=True)
    print(f"[1/7] Loading predictions from {path} …", flush=True)
    df = pd.read_parquet(path)
    required = {"date", "symbol", "label", "pred_rank"}
    missing = required - set(df.columns)
    if missing:
        print(f"[FATAL] Missing required columns: {missing}", flush=True)
        sys.exit(1)
    print(f"       Loaded {len(df):,} rows × {len(df.columns)} cols", flush=True)
    print(f"       Date range: {df['date'].min()} to {df['date'].max()}", flush=True)
    print(f"       Symbols: {df['symbol'].nunique():,}", flush=True)
    return df


# ===========================================================================
#  2.  Build meta-labeler training set from real predictions
# ===========================================================================
def build_meta_dataset(
    df: pd.DataFrame,
    top_k: int = DEFAULT_TOP_K,
) -> pd.DataFrame:
    """
    For each date, select top-K stocks by pred_rank, use label directly as meta_label.

    Parameters
    ----------
    df : DataFrame with columns [date, symbol, label, pred_rank, ...features]
    top_k : number of top candidates per date to keep

    Returns
    -------
    DataFrame with meta_label (0/1) and feature columns.
    """
    print(f"[2/7] Selecting top-{top_k} candidates per date by pred_rank …", flush=True)

    # Sort within each date by pred_rank descending, keep top-K
    df = df.copy()
    df["_rank_in_date"] = df.groupby("date")["pred_rank"].rank(
        method="dense", ascending=False
    )
    candidates = df[df["_rank_in_date"] <= top_k].copy()
    candidates = candidates.drop(columns=["_rank_in_date"])

    # Use 'label' column directly as the meta-label
    candidates["_meta_label"] = candidates["label"].astype(int)

    print(f"       Candidates selected: {len(candidates):,} rows", flush=True)

    # Class balance
    pos = (candidates["_meta_label"] == 1).sum()
    neg = (candidates["_meta_label"] == 0).sum()
    print(f"       Positives (label=1): {pos:,}  "
          f"Negatives (label=0): {neg:,}  "
          f"Ratio: {pos/neg:.3f}", flush=True)

    # Date range
    print(f"       Date range: {candidates['date'].min()} to {candidates['date'].max()}", flush=True)
    print(f"       Unique dates: {candidates['date'].nunique()}", flush=True)

    return candidates


# ===========================================================================
#  3.  Feature engineering
# ===========================================================================
def get_feature_columns(df: pd.DataFrame) -> list[str]:
    """Identify feature columns (exclude metadata / id / label columns)."""
    exclude = {
        "date", "symbol", "label", "pred_rank", "query_group", "trend_type",
        "_meta_label", "_pred_score", "_rank_in_date", "index",
    }
    return [
        c for c in df.columns
        if c.lower() not in exclude and not c.startswith("__")
    ]


def prepare_xy(
    df: pd.DataFrame,
    feature_cols: list[str] | None = None,
) -> tuple[pd.DataFrame, pd.Series, np.ndarray | None]:
    """Extract feature matrix X, target y, and optional sample weights."""
    if feature_cols is None:
        feature_cols = get_feature_columns(df)

    X = df[feature_cols].copy()
    y = df["_meta_label"].values

    # Basic sanity
    assert X.shape[1] > 0, "No feature columns found!"
    assert y.ndim == 1

    # Handle any remaining NaN in features
    nan_count = X.isna().sum().sum()
    if nan_count > 0:
        print(f"       Warning: {nan_count} NaN values in features — filling with 0.",
              flush=True)
        X = X.fillna(0)

    # Ensure all numeric (no categoricals leaked)
    for col in X.columns:
        if X[col].dtype == "object":
            print(f"       Warning: dropping non-numeric feature '{col}'", flush=True)
            X = X.drop(columns=[col])

    return X, y, None


# ===========================================================================
#  4.  Time-based train/test split (80/20 by date)
# ===========================================================================
def time_split(
    df: pd.DataFrame,
    train_frac: float = TRAIN_SPLIT,
) -> tuple[pd.DataFrame, pd.DataFrame]:
    """Split by date: oldest train_frac fraction of dates for training, rest for test."""
    dates = sorted(df["date"].unique())
    split_idx = int(len(dates) * train_frac)
    train_dates = set(dates[:split_idx])
    test_dates = set(dates[split_idx:])

    train_df = df[df["date"].isin(train_dates)].copy()
    test_df = df[df["date"].isin(test_dates)].copy()

    print(f"[3/7] Time-based split: {len(train_dates)} train dates "
          f"({df[df['date'].isin(train_dates)]['_meta_label'].sum():,} positives), "
          f"{len(test_dates)} test dates "
          f"({df[df['date'].isin(test_dates)]['_meta_label'].sum():,} positives)",
          flush=True)
    print(f"       Train: {len(train_df):,} rows, Test: {len(test_df):,} rows",
          flush=True)

    return train_df, test_df


# ===========================================================================
#  5.  Train with cross-validation (on training split)
# ===========================================================================
def train_cv(
    X: pd.DataFrame,
    y: np.ndarray,
    n_folds: int = DEFAULT_CV_FOLDS,
    random_state: int = RANDOM_STATE,
) -> tuple[lgb.Booster, float, np.ndarray]:
    """
    Time-series cross-validation using expanding window.

    Splits are sequential — train on past, validate on future — matching
    how the model will be used in production.

    Returns
    -------
    (trained_booster, mean_val_auc, oof_preds)
    """
    print(f"[4/7] Training LightGBM with {n_folds}-fold time-series CV …", flush=True)

    tscv = TimeSeriesSplit(n_splits=n_folds)

    params = {
        "objective": "binary",
        "metric": "auc",
        "boosting_type": "gbdt",
        "num_leaves": 63,
        "learning_rate": 0.05,
        "feature_fraction": 0.8,
        "bagging_fraction": 0.8,
        "bagging_freq": 5,
        "reg_alpha": 0.1,
        "reg_lambda": 1.0,
        "verbose": -1,
        "seed": random_state,
        "first_metric_only": True,
    }

    fold_models: list[lgb.Booster] = []
    fold_aucs: list[float] = []
    oof_preds = np.zeros(len(y), dtype=np.float64)

    for fold, (train_idx, val_idx) in enumerate(tscv.split(X)):
        X_tr, X_val = X.iloc[train_idx], X.iloc[val_idx]
        y_tr, y_val = y[train_idx], y[val_idx]

        train_data = lgb.Dataset(X_tr, label=y_tr)
        val_data = lgb.Dataset(X_val, label=y_val, reference=train_data)

        model = lgb.train(
            params=params,
            train_set=train_data,
            valid_sets=[val_data],
            valid_names=["valid"],
            callbacks=[
                lgb.early_stopping(stopping_rounds=50, verbose=False),
                lgb.log_evaluation(period=10),
            ],
            num_boost_round=1000,
        )

        fold_pred = model.predict(X_val, num_iteration=model.best_iteration)
        fold_auc = roc_auc_score(y_val, fold_pred)
        oof_preds[val_idx] = fold_pred

        fold_models.append(model)
        fold_aucs.append(fold_auc)
        print(f"       Fold {fold + 1}/{n_folds}  —  AUC = {fold_auc:.5f}  "
              f"(best_iter={model.best_iteration})", flush=True)

    mean_val_auc = float(np.mean(fold_aucs))
    oof_auc = roc_auc_score(y, oof_preds)

    print(f"       Mean CV AUC       = {mean_val_auc:.5f}", flush=True)
    print(f"       Out-of-fold AUC   = {oof_auc:.5f}", flush=True)

    # Retrain on full training data
    print("       Retraining on full training set …", flush=True)
    full_data = lgb.Dataset(X, label=y)
    best_rounds = int(np.median([m.best_iteration for m in fold_models]))
    final_model = lgb.train(
        params=params,
        train_set=full_data,
        num_boost_round=best_rounds,
        valid_sets=[full_data],
        valid_names=["train"],
        callbacks=[lgb.log_evaluation(period=0)],
    )

    # Save CV metrics to JSON
    RESULTS_DIR.mkdir(parents=True, exist_ok=True)
    metrics = {
        "fold_aucs": [float(a) for a in fold_aucs],
        "mean_cv_auc": mean_val_auc,
        "oof_auc": oof_auc,
        "best_iterations": [m.best_iteration for m in fold_models],
        "median_best_iteration": best_rounds,
        "n_folds": n_folds,
        "n_train": len(X),
    }
    metrics_path = RESULTS_DIR / "train_metalabeler_v3_metrics.json"
    try:
        with open(metrics_path, "w") as fh:
            json.dump(metrics, fh, indent=2)
        print(f"       Metrics saved to {metrics_path}", flush=True)
    except Exception as e:
        print(f"       WARNING: Could not save metrics: {e}", flush=True)

    return final_model, oof_auc, oof_preds


# ===========================================================================
#  6.  Evaluation on test set
# ===========================================================================
def evaluate(
    y_true: np.ndarray,
    y_pred: np.ndarray,
    label: str = "Test",
) -> tuple[float, list[tuple[int, float, float]]]:
    """
    Evaluate binary classifier on a hold-out set.

    Returns
    -------
    (best_threshold, precision_at_k_list)
    """
    print(f"[5/7] Evaluating on {label} set …", flush=True)

    auc = roc_auc_score(y_true, y_pred)
    fpr, tpr, roc_thresholds = roc_curve(y_true, y_pred)

    # Youden index -> best threshold
    youden = tpr - fpr
    best_idx = int(np.argmax(youden))
    best_thr = float(roc_thresholds[best_idx])

    print(f"       AUC             = {auc:.5f}", flush=True)
    print(f"       Best threshold  = {best_thr:.5f}  (Youden index)", flush=True)

    # --- Precision@K ---
    order = np.argsort(y_pred)[::-1]
    y_sorted = y_true[order]
    precisions: list[tuple[int, float, float]] = []
    for k in (10, 20, 50, 100, 200):
        if k <= len(y_sorted):
            prec_k = float(precision_score(y_sorted[:k], np.ones(k), zero_division=0.0))
        else:
            prec_k = float(precision_score(y_sorted, np.ones(len(y_sorted)), zero_division=0.0))
        precisions.append((k, prec_k, best_thr))
        print(f"       Precision@{k:<3d} = {prec_k:.5f}", flush=True)

    # --- Confusion matrix ---
    y_pred_bin = (y_pred >= best_thr).astype(int)
    tn, fp, fn, tp = confusion_matrix(y_true, y_pred_bin).ravel()
    print(f"\n       Confusion Matrix @ threshold = {best_thr:.5f}", flush=True)
    print(f"                    Pred Neg    Pred Pos", flush=True)
    print(f"       Actual Neg   {tn:>8d}   {fp:>8d}", flush=True)
    print(f"       Actual Pos   {fn:>8d}   {tp:>8d}", flush=True)
    total = tn + fp + fn + tp
    acc = (tn + tp) / total
    prec = tp / (tp + fp) if (tp + fp) > 0 else 0.0
    rec = tp / (tp + fn) if (tp + fn) > 0 else 0.0
    f1 = 2 * prec * rec / (prec + rec) if (prec + rec) > 0 else 0.0
    print(f"       Accuracy  = {acc:.5f}", flush=True)
    print(f"       Precision = {prec:.5f}", flush=True)
    print(f"       Recall    = {rec:.5f}", flush=True)
    print(f"       F1        = {f1:.5f}", flush=True)

    return best_thr, precisions


# ===========================================================================
#  7.  Save model
# ===========================================================================
def save_model(model: lgb.Booster, path: str) -> None:
    """Save LightGBM booster in text format and upload to HF dataset."""
    import shutil

    # Always save to /tmp first (writable)
    tmp_path = "/tmp/meta_labeler_v3.txt"
    model.save_model(tmp_path)
    print(f"       Model saved to {tmp_path}", flush=True)

    # Copy to requested output path if different
    if path and path != tmp_path:
        try:
            shutil.copy(tmp_path, path)
            print(f"       Copied to {path}", flush=True)
        except Exception as e:
            print(f"       WARNING: Could not copy to {path}: {e}", flush=True)

    # Upload to HF dataset
    try:
        from huggingface_hub import HfApi
        api = HfApi(token=os.environ.get("HF_TOKEN"))
        api.upload_file(
            path_or_fileobj=tmp_path,
            path_in_repo="models/meta_labeler_v3.txt",
            repo_id="cedwyh/jinjing-shared-data",
            repo_type="dataset",
        )
        print(f"       Uploaded: cedwyh/jinjing-shared-data/models/meta_labeler_v3.txt",
              flush=True)
    except Exception as e:
        print(f"       WARNING: HF upload failed: {e}", flush=True)

    fsize = os.path.getsize(tmp_path) / 1024
    print(f"       Model saved ({fsize:.1f} KB).", flush=True)


# ===========================================================================
#  8.  Summary
# ===========================================================================
def print_summary(
    oof_auc: float,
    test_auc: float,
    test_threshold: float,
    test_precisions: list[tuple[int, float, float]],
    feature_cols: list[str],
) -> None:
    """Final summary table."""
    print(f"[7/7] Final Summary", flush=True)
    print("=" * 60, flush=True)
    print(f"  {'Metric':<25s}  {'Value':>12s}", flush=True)
    print("-" * 60, flush=True)
    print(f"  {'CV OOF AUC':<25s}  {oof_auc:>12.5f}", flush=True)
    print(f"  {'Test AUC':<25s}  {test_auc:>12.5f}", flush=True)
    print(f"  {'Test best threshold':<25s}  {test_threshold:>12.5f}", flush=True)
    for k, p, _ in test_precisions:
        print(f"  {'Test Precision@' + str(k):<25s}  {p:>12.5f}", flush=True)
    print(f"  {'Features':<25s}  {len(feature_cols):>12d}", flush=True)
    print("=" * 60, flush=True)


# ===========================================================================
#  CLI
# ===========================================================================
def parse_args(argv: list[str] | None = None) -> argparse.Namespace:
    parser = argparse.ArgumentParser(
        description="Train a binary meta-labeler for Jinjing V3 using real ranker predictions.",
        formatter_class=argparse.ArgumentDefaultsHelpFormatter,
    )
    parser.add_argument(
        "--predictions",
        type=str,
        default=DEFAULT_PREDICTIONS,
        help="Path to predictions parquet file.",
    )
    parser.add_argument(
        "--top-k",
        type=int,
        default=DEFAULT_TOP_K,
        help="Number of top candidates per date by pred_rank.",
    )
    parser.add_argument(
        "--cv-folds",
        type=int,
        default=DEFAULT_CV_FOLDS,
        help="Number of cross-validation folds.",
    )
    parser.add_argument(
        "--output",
        type=str,
        default=DEFAULT_OUTPUT,
        help="Path to save the trained LightGBM model (text format).",
    )
    parser.add_argument(
        "--train-split",
        type=float,
        default=TRAIN_SPLIT,
        help="Fraction of dates (oldest) for training.",
    )
    parser.add_argument(
        "--hf-token",
        type=str,
        default=None,
        help="Hugging Face token (default: $HF_TOKEN).",
    )
    return parser.parse_args(argv)


# ===========================================================================
#  Main
# ===========================================================================
def main() -> None:
    args = parse_args()
    hf_token = args.hf_token or os.environ.get("HF_TOKEN")
    if not hf_token:
        print(
            "[FATAL] HF_TOKEN not set. Provide via --hf-token or "
            "the HF_TOKEN environment variable.",
            flush=True,
        )
        sys.exit(1)

    # Validate --train-split range
    if not 0 < args.train_split < 1:
        print(f"[FATAL] --train-split must be between 0 and 1 (got {args.train_split})",
              flush=True)
        sys.exit(1)

    # ---- 1. Load predictions ----
    df = load_predictions(args.predictions)

    # ---- 2. Build meta dataset (top-K by pred_rank) ----
    meta_df = build_meta_dataset(df, top_k=args.top_k)

    # ---- 3. Feature columns ----
    feat_cols = get_feature_columns(meta_df)
    print(f"[3/7] Using {len(feat_cols)} feature columns: "
          f"{', '.join(feat_cols)}", flush=True)

    # ---- 4. Time-based split ----
    train_df, test_df = time_split(meta_df, train_frac=args.train_split)

    # Sort training data chronologically so TimeSeriesSplit splits correctly
    train_df = train_df.sort_values("date").reset_index(drop=True)

    # ---- 5. Prepare X, y ----
    X_tr, y_tr, _ = prepare_xy(train_df, feat_cols)
    X_te, y_te, _ = prepare_xy(test_df, feat_cols)

    # ---- 6. Train with CV ----
    model, oof_auc, oof_preds = train_cv(X_tr, y_tr, n_folds=args.cv_folds)

    # ---- 7. Evaluate on test set ----
    test_preds = model.predict(X_te, num_iteration=model.best_iteration)
    test_thr, test_precisions = evaluate(y_te, test_preds, label="Test")

    # ---- 8. Save model ----
    save_model(model, args.output)

    # ---- 9. Summary ----
    print_summary(oof_auc, roc_auc_score(y_te, test_preds), test_thr, test_precisions, feat_cols)


if __name__ == "__main__":
    main()