evaluation/evaluate_frozen_forecasts.py

#!/usr/bin/env python3
"""Evaluate RestockIQ's frozen production forecast artifacts on June 2024.

This runner deliberately imports the production snapshot builder, feature
engineering, artifact loader, and inference code from the pinned Git checkout.
Oracle latent demand is loaded separately and used only after predictions have
been produced.
"""

from __future__ import annotations

import argparse
import hashlib
import json
import math
import os
import platform
import shutil
import subprocess
import sys
import tempfile
from datetime import date, timedelta
from pathlib import Path
from typing import Any

import numpy as np
import pandas as pd
import lightgbm as lgb
import sqlalchemy
from sqlalchemy import create_engine
from sqlalchemy.orm import Session


DEFAULT_EXPECTED_GIT_SHA = "06ae958730f87cc44b3a5dbb7094fb8a3c88f7a3"
DEFAULT_EXPECTED_WORKBOOK_SHA256 = (
    "9d13eb22b2d2a2acecc46661d4984420e8390861a3bf9c0590cae57f9f5858e0"
)
DATASET_ID = "demo-retail-v1"
HORIZONS = (1, 7, 14)
EVALUATION_START = date(2024, 6, 1)
EVALUATION_END = date(2024, 6, 30)
ORACLE_LABEL = "units_demanded_est"
ORACLE_FORBIDDEN_FEATURES = {
    "avg_daily_demand_per_store",
    "cash_locked_in_stock_rp",
    "demand_profile",
    "units_demanded_est",
}


def sha256_file(path: Path) -> str:
    digest = hashlib.sha256()
    with path.open("rb") as handle:
        for block in iter(lambda: handle.read(1024 * 1024), b""):
            digest.update(block)
    return digest.hexdigest()


def git_output(repo_root: Path, *args: str) -> str:
    return subprocess.check_output(
        ["git", "-C", str(repo_root), *args], text=True
    ).strip()


def croston_daily(history: np.ndarray, alpha: float = 0.1) -> float:
    """Causal Croston estimate after consuming all supplied observations."""
    level = 0.0
    interval = 1.0
    elapsed = 0
    initialized = False
    for value in np.asarray(history, dtype=float):
        elapsed += 1
        if value > 0:
            if not initialized:
                level = float(value)
                interval = float(max(elapsed, 1))
                initialized = True
            else:
                level = alpha * float(value) + (1.0 - alpha) * level
                interval = alpha * float(elapsed) + (1.0 - alpha) * interval
            elapsed = 0
    return max(0.0, level / interval) if initialized and interval > 0 else 0.0


def causal_baselines(history: pd.Series, horizon: int) -> dict[str, float]:
    values = pd.to_numeric(history, errors="coerce").fillna(0).clip(lower=0)
    if values.empty:
        return {
            "seasonal_naive_7d": 0.0,
            "moving_average_28d": 0.0,
            "ewma_28d": 0.0,
            "croston_01": 0.0,
        }

    last_week = values.tail(7).to_numpy(dtype=float)
    seasonal = float(
        sum(last_week[index % len(last_week)] for index in range(horizon))
    )
    moving_average = float(values.tail(28).mean() * horizon)
    ewma = float(
        values.ewm(span=28, adjust=False, min_periods=1).mean().iloc[-1]
        * horizon
    )
    croston = float(croston_daily(values.to_numpy(dtype=float)) * horizon)
    return {
        "seasonal_naive_7d": seasonal,
        "moving_average_28d": moving_average,
        "ewma_28d": ewma,
        "croston_01": croston,
    }


def regression_metrics(actual: np.ndarray, prediction: np.ndarray) -> dict[str, Any]:
    y = np.asarray(actual, dtype=float)
    p = np.asarray(prediction, dtype=float)
    error = p - y
    nonzero = y != 0
    return {
        "n": int(len(y)),
        "actual_sum": float(y.sum()),
        "prediction_sum": float(p.sum()),
        "mae": float(np.mean(np.abs(error))),
        "rmse": float(np.sqrt(np.mean(np.square(error)))),
        "wmape": float(np.abs(error).sum() / max(np.abs(y).sum(), 1e-12)),
        "wmape_percent": float(
            100.0 * np.abs(error).sum() / max(np.abs(y).sum(), 1e-12)
        ),
        "bias": float(np.mean(error)),
        "mape_nonzero": (
            float(np.mean(np.abs(error[nonzero] / y[nonzero])))
            if bool(nonzero.any())
            else None
        ),
        "mape_nonzero_percent": (
            float(100.0 * np.mean(np.abs(error[nonzero] / y[nonzero])))
            if bool(nonzero.any())
            else None
        ),
        "zero_actual_rows": int((~nonzero).sum()),
    }


def pinball(actual: np.ndarray, prediction: np.ndarray, alpha: float) -> float:
    residual = np.asarray(actual, dtype=float) - np.asarray(prediction, dtype=float)
    return float(np.mean(np.maximum(alpha * residual, (alpha - 1.0) * residual)))


def atomic_json(path: Path, value: Any) -> None:
    temp = path.with_suffix(path.suffix + ".tmp")
    temp.write_text(json.dumps(value, indent=2, sort_keys=True) + "\n")
    temp.replace(path)


def parse_args() -> argparse.Namespace:
    parser = argparse.ArgumentParser(
        description="Exact-artifact RestockIQ forecast evaluation"
    )
    parser.add_argument("--repo-root", type=Path, required=True)
    parser.add_argument("--output-dir", type=Path, required=True)
    parser.add_argument("--workbook", type=Path)
    parser.add_argument(
        "--expected-git-sha", default=DEFAULT_EXPECTED_GIT_SHA
    )
    parser.add_argument(
        "--expected-workbook-sha256", default=DEFAULT_EXPECTED_WORKBOOK_SHA256
    )
    return parser.parse_args()


def main() -> None:
    args = parse_args()
    repo_root = args.repo_root.expanduser().resolve()
    output_dir = args.output_dir.expanduser().resolve()
    workbook = (
        args.workbook.expanduser().resolve()
        if args.workbook
        else repo_root
        / "backend"
        / "data"
        / "synthetic"
        / "RestockIQ_Dataset_Sintetis.xlsx"
    )
    artifact_dir = repo_root / "backend" / "artifacts" / "restockiq-demand-v1"

    if not (repo_root / ".git").exists():
        raise SystemExit(f"Not a Git checkout: {repo_root}")
    if not workbook.is_file():
        raise SystemExit(f"Workbook not found: {workbook}")
    if not (artifact_dir / "manifest.json").is_file():
        raise SystemExit(f"Frozen artifact manifest not found: {artifact_dir}")

    git_sha = git_output(repo_root, "rev-parse", "HEAD")
    if git_sha != args.expected_git_sha:
        raise SystemExit(
            f"Git SHA mismatch: expected {args.expected_git_sha}, got {git_sha}. "
            "Do not silently evaluate another release."
        )
    workbook_sha = sha256_file(workbook)
    if workbook_sha != args.expected_workbook_sha256:
        raise SystemExit(
            "Workbook SHA-256 mismatch: "
            f"expected {args.expected_workbook_sha256}, got {workbook_sha}"
        )

    backend_root = repo_root / "backend"
    sys.path.insert(0, str(backend_root))
    os.environ.setdefault("DATABASE_URL", "sqlite:///:memory:")

    from app.db.seed import seed
    from app.ml.artifact_store import load_model_artifacts
    from app.ml.demand_engine import generate_demand_forecasts
    from app.services.retail_snapshot_service import build_retail_snapshot

    artifacts = load_model_artifacts(artifact_dir, force_reload=True)
    manifest = artifacts.manifest
    used_oracle = set(manifest.get("oracle_fields_used_as_features", []))
    if used_oracle:
        raise SystemExit(f"Artifact manifest declares Oracle features: {used_oracle}")
    declared_features = set(artifacts.reconstruction_feature_columns)
    for artifact in artifacts.forecasts.values():
        declared_features.update(artifact.feature_columns)
    leaked_features = declared_features & ORACLE_FORBIDDEN_FEATURES
    if leaked_features:
        raise SystemExit(f"Oracle feature leak detected: {sorted(leaked_features)}")
    if manifest.get("training_cutoff") != "2024-05-31":
        raise SystemExit(
            f"Unexpected training cutoff: {manifest.get('training_cutoff')}"
        )

    sales = pd.read_excel(workbook, sheet_name="Fact_Daily_Sales")
    stores = pd.read_excel(workbook, sheet_name="Dim_Stores")
    sales["date"] = pd.to_datetime(sales["date"]).dt.date
    required_columns = {
        "date",
        "store_id",
        "sku_id",
        "units_sold",
        ORACLE_LABEL,
    }
    missing_columns = required_columns - set(sales.columns)
    if missing_columns:
        raise SystemExit(f"Workbook missing evaluation columns: {missing_columns}")

    origin_end = EVALUATION_END - timedelta(days=min(HORIZONS))
    origins = list(
        pd.date_range(EVALUATION_START, origin_end, freq="D").date
    )
    prediction_rows: list[dict[str, Any]] = []

    with tempfile.TemporaryDirectory(prefix="restockiq-eval-") as temp_dir:
        database_path = Path(temp_dir) / "evaluation.sqlite3"
        database_url = f"sqlite:///{database_path}"
        seed(database_url=database_url, workbook_path=workbook)
        engine = create_engine(database_url)

        with Session(engine) as db:
            for origin_index, origin in enumerate(origins, start=1):
                eligible_horizons = [
                    horizon
                    for horizon in HORIZONS
                    if origin + timedelta(days=horizon) <= EVALUATION_END
                ]
                if not eligible_horizons:
                    continue
                print(
                    f"[{origin_index:02d}/{len(origins):02d}] origin={origin} "
                    f"horizons={eligible_horizons}",
                    flush=True,
                )

                for store_id in stores["store_id"].astype(str).sort_values():
                    snapshot = build_retail_snapshot(
                        db,
                        dataset_id=DATASET_ID,
                        store_id=store_id,
                        decision_date=origin,
                        horizon_days=max(eligible_horizons),
                        lookback_days=182,
                    )
                    result = generate_demand_forecasts(
                        snapshot,
                        artifacts,
                        horizon_days=max(eligible_horizons),
                    )

                    for forecast in result.forecasts:
                        sku_id = forecast.sku_id
                        sku_history = sales.loc[
                            (sales["store_id"].astype(str) == store_id)
                            & (sales["sku_id"].astype(str) == sku_id)
                            & (sales["date"] <= origin)
                        ].sort_values("date")

                        for horizon in eligible_horizons:
                            target_start = origin + timedelta(days=1)
                            target_end = origin + timedelta(days=horizon)
                            target_rows = sales.loc[
                                (sales["store_id"].astype(str) == store_id)
                                & (sales["sku_id"].astype(str) == sku_id)
                                & (sales["date"] >= target_start)
                                & (sales["date"] <= target_end)
                            ]
                            if len(target_rows) != horizon:
                                raise SystemExit(
                                    "Incomplete Oracle target window for "
                                    f"{origin}/{store_id}/{sku_id}/H{horizon}: "
                                    f"expected {horizon} rows, got {len(target_rows)}"
                                )
                            actual = float(target_rows[ORACLE_LABEL].sum())
                            quantiles = forecast.forecasts[horizon]
                            baselines = causal_baselines(
                                sku_history["units_sold"], horizon
                            )
                            prediction_rows.append(
                                {
                                    "decision_date": origin.isoformat(),
                                    "target_start": target_start.isoformat(),
                                    "target_end": target_end.isoformat(),
                                    "store_id": store_id,
                                    "sku_id": sku_id,
                                    "horizon_days": horizon,
                                    "actual_oracle_demand": actual,
                                    "frozen_product_q10": quantiles.q10,
                                    "frozen_product_q50": quantiles.q50,
                                    "frozen_product_q90": quantiles.q90,
                                    **baselines,
                                    "model_version": result.model_version,
                                    "training_cutoff": manifest["training_cutoff"],
                                    "git_sha": git_sha,
                                }
                            )

    predictions = pd.DataFrame(prediction_rows).sort_values(
        ["horizon_days", "decision_date", "store_id", "sku_id"],
        kind="stable",
    )
    if predictions.empty:
        raise SystemExit("Evaluation produced no prediction rows")

    expected_rows = sum(
        (
            EVALUATION_END
            - (EVALUATION_START + timedelta(days=horizon))
        ).days
        + 1
        for horizon in HORIZONS
    ) * len(stores) * sales["sku_id"].nunique()
    if len(predictions) != expected_rows:
        raise SystemExit(
            f"Unexpected prediction count: expected {expected_rows}, "
            f"got {len(predictions)}"
        )

    point_models = [
        "frozen_product_q50",
        "seasonal_naive_7d",
        "moving_average_28d",
        "ewma_28d",
        "croston_01",
    ]
    metric_rows: list[dict[str, Any]] = []
    store_metric_rows: list[dict[str, Any]] = []
    quantile_rows: list[dict[str, Any]] = []

    for horizon, frame in predictions.groupby("horizon_days", sort=True):
        y = frame["actual_oracle_demand"].to_numpy(dtype=float)
        for model_name in point_models:
            metrics = regression_metrics(
                y, frame[model_name].to_numpy(dtype=float)
            )
            metric_rows.append(
                {"horizon_days": int(horizon), "model": model_name, **metrics}
            )

        q10 = frame["frozen_product_q10"].to_numpy(dtype=float)
        q50 = frame["frozen_product_q50"].to_numpy(dtype=float)
        q90 = frame["frozen_product_q90"].to_numpy(dtype=float)
        crossing_rows = int(((q10 > q50) | (q50 > q90)).sum())
        quantile_rows.append(
            {
                "horizon_days": int(horizon),
                "n": int(len(frame)),
                "coverage_80": float(np.mean((y >= q10) & (y <= q90))),
                "coverage_80_percent": float(
                    100.0 * np.mean((y >= q10) & (y <= q90))
                ),
                "mean_interval_width": float(np.mean(q90 - q10)),
                "pinball_q10": pinball(y, q10, 0.10),
                "pinball_q50": pinball(y, q50, 0.50),
                "pinball_q90": pinball(y, q90, 0.90),
                "quantile_crossing_rows": crossing_rows,
            }
        )

        for store_id, store_frame in frame.groupby("store_id", sort=True):
            store_y = store_frame["actual_oracle_demand"].to_numpy(dtype=float)
            for model_name in point_models:
                metrics = regression_metrics(
                    store_y, store_frame[model_name].to_numpy(dtype=float)
                )
                store_metric_rows.append(
                    {
                        "horizon_days": int(horizon),
                        "store_id": str(store_id),
                        "model": model_name,
                        **metrics,
                    }
                )

    metrics = pd.DataFrame(metric_rows).sort_values(
        ["horizon_days", "wmape", "model"], kind="stable"
    )
    store_metrics = pd.DataFrame(store_metric_rows).sort_values(
        ["horizon_days", "store_id", "wmape", "model"], kind="stable"
    )
    quantile_metrics = pd.DataFrame(quantile_rows).sort_values("horizon_days")

    comparisons: list[dict[str, Any]] = []
    for horizon, frame in metrics.groupby("horizon_days", sort=True):
        product = frame.loc[frame["model"] == "frozen_product_q50"].iloc[0]
        baselines = frame.loc[frame["model"] != "frozen_product_q50"]
        best = baselines.sort_values("wmape", kind="stable").iloc[0]
        comparisons.append(
            {
                "horizon_days": int(horizon),
                "product_wmape_percent": float(product["wmape_percent"]),
                "best_baseline": str(best["model"]),
                "best_baseline_wmape_percent": float(best["wmape_percent"]),
                "relative_wmape_improvement_vs_best_baseline_percent": float(
                    100.0 * (best["wmape"] - product["wmape"]) / best["wmape"]
                ),
                "product_beats_best_baseline": bool(product["wmape"] < best["wmape"]),
            }
        )
    comparison_frame = pd.DataFrame(comparisons)

    output_dir.mkdir(parents=True, exist_ok=True)
    evaluator_copy = output_dir / "evaluate_frozen_forecasts.py"
    shutil.copy2(Path(__file__).resolve(), evaluator_copy)
    predictions_path = output_dir / "predictions.csv"
    metrics_path = output_dir / "metrics.csv"
    store_metrics_path = output_dir / "metrics_by_store.csv"
    quantile_path = output_dir / "quantile_metrics.csv"
    comparison_path = output_dir / "baseline_comparison.csv"
    predictions.to_csv(predictions_path, index=False)
    metrics.to_csv(metrics_path, index=False)
    store_metrics.to_csv(store_metrics_path, index=False)
    quantile_metrics.to_csv(quantile_path, index=False)
    comparison_frame.to_csv(comparison_path, index=False)

    manifest_out = {
        "evaluation_name": "RestockIQ exact frozen-artifact June 2024 evaluation",
        "status": "completed_synthetic_controlled_evaluation",
        "git_sha": git_sha,
        "git_remote": git_output(repo_root, "remote", "get-url", "origin"),
        "git_status_porcelain": git_output(repo_root, "status", "--porcelain=v1"),
        "evaluator_sha256": sha256_file(evaluator_copy),
        "artifact_manifest_sha256": sha256_file(artifact_dir / "manifest.json"),
        "workbook": workbook.name,
        "workbook_sha256": workbook_sha,
        "dataset_id": DATASET_ID,
        "evaluation_period": [
            EVALUATION_START.isoformat(),
            EVALUATION_END.isoformat(),
        ],
        "origin_rule": "decision date t; target is Oracle t+1 through t+H",
        "label": ORACLE_LABEL,
        "label_role": "synthetic Oracle evaluation-only; never a model feature",
        "model_version": artifacts.version,
        "training_cutoff": manifest["training_cutoff"],
        "training_data_hash": manifest["training_data_hash"],
        "oracle_fields_used_as_features": sorted(used_oracle),
        "horizons": list(HORIZONS),
        "stores": int(len(stores)),
        "skus": int(sales["sku_id"].nunique()),
        "prediction_rows": int(len(predictions)),
        "runtime_versions": {
            "python": platform.python_version(),
            "numpy": np.__version__,
            "pandas": pd.__version__,
            "scipy": __import__("scipy").__version__,
            "lightgbm": lgb.__version__,
            "sqlalchemy": sqlalchemy.__version__,
        },
        "baseline_definitions": {
            "seasonal_naive_7d": "repeat the last seven observed-sales days causally",
            "moving_average_28d": "28-day observed-sales mean times H",
            "ewma_28d": "causal span-28 observed-sales EWMA times H",
            "croston_01": "causal Croston alpha=0.1 daily rate times H",
        },
        "limitations": [
            "Synthetic controlled evaluation; not evidence of real merchant impact.",
            "Oracle latent demand is used only as an evaluation label.",
            "Metrics apply to the pinned frozen artifact and workbook only.",
            "No stockout reduction, savings, revenue, ROI, or service-level claim is produced.",
        ],
    }
    atomic_json(output_dir / "run_manifest.json", manifest_out)

    metrics_payload = {
        "point_metrics": metrics.to_dict(orient="records"),
        "quantile_metrics": quantile_metrics.to_dict(orient="records"),
        "baseline_comparison": comparison_frame.to_dict(orient="records"),
    }
    atomic_json(output_dir / "metrics.json", metrics_payload)

    summary_lines = [
        "# RestockIQ Frozen-Artifact Forecast Evaluation",
        "",
        "**Status:** controlled synthetic evaluation completed.",
        "",
        f"- Git SHA: `{git_sha}`",
        f"- Model: `{artifacts.version}`",
        f"- Training cutoff: `{manifest['training_cutoff']}`",
        f"- Workbook SHA-256: `{workbook_sha}`",
        f"- Prediction rows: `{len(predictions):,}`",
        "- Target: synthetic Oracle demand over `t+1..t+H`.",
        "- Oracle fields used as model features: `none`.",
        "",
        "## Point metrics",
        "",
        metrics.to_markdown(index=False, floatfmt=".4f"),
        "",
        "## Quantile diagnostics",
        "",
        quantile_metrics.to_markdown(index=False, floatfmt=".4f"),
        "",
        "## Comparison with the strongest tested baseline",
        "",
        comparison_frame.to_markdown(index=False, floatfmt=".4f"),
        "",
        "## Claim boundary",
        "",
        "These results validate the pinned forecasting artifact only on the controlled synthetic June 2024 window. They do not demonstrate realized stockout reduction, savings, revenue uplift, service-level improvement, ROI, or generalization to real merchants.",
        "",
    ]
    (output_dir / "EVALUATION_SUMMARY.md").write_text("\n".join(summary_lines))

    checksum_files = sorted(
        path
        for path in output_dir.iterdir()
        if path.is_file() and path.name != "SHA256SUMS"
    )
    (output_dir / "SHA256SUMS").write_text(
        "\n".join(f"{sha256_file(path)}  {path.name}" for path in checksum_files)
        + "\n"
    )

    print("\nEvaluation completed.")
    print(metrics.to_string(index=False))
    print("\nQuantile diagnostics:")
    print(quantile_metrics.to_string(index=False))
    print(f"\nOutputs: {output_dir}")


if __name__ == "__main__":
    main()