Publish frozen RestockIQ demand artifact and evaluation

README.md

@@ -0,0 +1,104 @@
+---
+library_name: lightgbm
+tags:
+- lightgbm
+- retail
+- demand-forecasting
+- quantile-regression
+- inventory
+---
+
+# RestockIQ Demand v1
+
+Frozen LightGBM artifact used by the RestockIQ semifinal product for stockout-censored demand reconstruction and direct cumulative quantile forecasting.
+
+## Artifact identity
+
+- Version: `restockiq-demand-v1-a067286b7c1e`
+- Production Git SHA evaluated: `06ae958730f87cc44b3a5dbb7094fb8a3c88f7a3`
+- Training dataset ID: `demo-retail-v1`
+- Training cutoff: `2024-05-31`
+- Training data hash: `a067286b7c1e966de43f3db9e1d24a7f8a440d236d94c24365c6d717ca44399c`
+- Seed: `20260723`
+- Artifact files covered by the supplied checksum manifest: 11
+- Oracle fields used as model features: none
+
+## Components
+
+The artifact contains:
+
+1. One causal stockout-censor reconstruction model trained with uncensored observed sales as its label.
+2. Nine direct cumulative quantile models: H1, H7, and H14 × Q10, Q50, and Q90.
+3. Frozen feature medians and split-conformal interval adjustments in `manifest.json`.
+4. Inference-time clipping and quantile-order repair in the RestockIQ application code.
+
+The cumulative target contract is demand from `t+1` through `t+H`; the decision-origin day is not part of its own label.
+
+## Controlled synthetic evaluation
+
+The supplied `evaluation/` directory was produced by the standalone evaluator against the pinned production Git SHA and frozen workbook. The evaluator imports the production snapshot builder, feature engineering, artifact loader, and inference path. Predictions are generated before the synthetic Oracle label is joined.
+
+Evaluation window: June 2024. Label: synthetic latent demand `units_demanded_est`, used for evaluation only. The benchmark contains 10,540 store–SKU–origin–horizon prediction rows.
+
+| Horizon | Q50 MAE | Q50 RMSE | Q50 WMAPE | Best tested baseline WMAPE | Relative result vs best baseline | Q10–Q90 coverage |
+|---:|---:|---:|---:|---:|---:|---:|
+| H1 | 2.384 | 3.590 | 31.33% | 32.68% (Croston) | +4.14% | 78.64% |
+| H7 | 9.563 | 14.722 | 18.19% | 17.48% (28-day moving average) | −4.10% | 81.07% |
+| H14 | 14.259 | 20.992 | 13.49% | 14.42% (28-day moving average) | +6.46% | 83.15% |
+
+There were zero quantile-crossing rows after the production inference repair. The frozen Q50 model beats the strongest tested causal baseline at H1 and H14, while the 28-day moving-average baseline remains stronger at H7. This mixed result is retained intentionally.
+
+## Intended use
+
+- Frozen controlled RestockIQ demo inference.
+- Reproduction of the supplied June 2024 synthetic evaluation.
+- Inspection of the model and Oracle-safe feature contract.
+
+The artifact is not a general pretrained retail model and does not perform live retraining.
+
+## Out-of-scope use and claims
+
+Do not use this artifact or its evaluation to claim:
+
+- proven generalization to real merchants or arbitrary POS systems;
+- realized stockout reduction;
+- inventory-cost reduction;
+- service-level improvement;
+- savings, revenue uplift, or ROI;
+- causal promotional effects.
+
+Supplier risk, LMAR, WCAR, and the exact sparse-DP allocator are downstream RestockIQ components. They are not outputs or submodels of these LightGBM files.
+
+## Oracle and leakage boundary
+
+The frozen manifest declares these fields as forbidden model features:
+
+- `units_demanded_est`
+- `demand_profile`
+- `avg_daily_demand_per_store`
+- `cash_locked_in_stock_rp`
+
+Realized future supplier outcomes are likewise excluded from decision-time inputs. Oracle data is permitted only after prediction for controlled evaluation and value decomposition.
+
+## Files
+
+- `manifest.json`
+- `SHA256SUMS`
+- `reconstruction_censor_no_inventory.txt`
+- `forecast_h1_q10.txt`, `forecast_h1_q50.txt`, `forecast_h1_q90.txt`
+- `forecast_h7_q10.txt`, `forecast_h7_q50.txt`, `forecast_h7_q90.txt`
+- `forecast_h14_q10.txt`, `forecast_h14_q50.txt`, `forecast_h14_q90.txt`
+- `evaluation/`: predictions, point metrics, quantile diagnostics, baseline comparison, run manifest, summary, and checksums
+
+## Verification
+
+From the model repository root:
+
+```bash
+sha256sum -c SHA256SUMS
+cd evaluation && sha256sum -c SHA256SUMS
+```
+
+The evaluation manifest records the exact Git SHA, workbook checksum, training cutoff, model version, target timing, baseline definitions, and claim limitations.
+
+No license is asserted in this card because the project team has not frozen a licensing decision.

SHA256SUMS

@@ -0,0 +1,11 @@
+b62a311a1273f3918d4127cbfae3d0891ee52bc2c3e5f749ea0d96db8c4c3869  forecast_h14_q10.txt
+57388956ae19e285081191702e3a63241a3a963659f7418221bdfc430838aa71  forecast_h14_q50.txt
+c4ca1f4dc09d4ec752b8edae02418586975620d77f35890c34988e5388313226  forecast_h14_q90.txt
+5c55e2b88f28c9515166bcd494f9c8244defa5e7f48817fa40da5a28990b0fdf  forecast_h1_q10.txt
+0b76a0775d09bde3f08315573b954bb32e1a104a992fcf87ccdf582f31b3acb1  forecast_h1_q50.txt
+3fe56e903d41ddfc604c3ec7ed0f43a5ae84c3ebce4fb505ce08252a077090fc  forecast_h1_q90.txt
+fd88d57346167401ca7d87f6b328a2ca2c940c2511fa86b351f032e6cf1822d0  forecast_h7_q10.txt
+1394cea95589cf09ce3dae5224046c4b2a7aa4a5286643bb8ccfb1605799b985  forecast_h7_q50.txt
+e03ccea57647d48b7b397cba42f30b9a0c8dffcd32b4333deaaa5617c88c771f  forecast_h7_q90.txt
+c3689bdf67ecb25edb5927584641446374a4feffad686fcd7d44ff6bf73561ac  manifest.json
+762dbcf51d11aa2c2372d768ddea947a5496a8f1b805dc94e94daeb4285a907c  reconstruction_censor_no_inventory.txt

evaluation/EVALUATION_SUMMARY.md

@@ -0,0 +1,51 @@
+# RestockIQ Frozen-Artifact Forecast Evaluation
+
+**Status:** controlled synthetic evaluation completed.
+
+- Git SHA: `06ae958730f87cc44b3a5dbb7094fb8a3c88f7a3`
+- Model: `restockiq-demand-v1-a067286b7c1e`
+- Training cutoff: `2024-05-31`
+- Workbook SHA-256: `9d13eb22b2d2a2acecc46661d4984420e8390861a3bf9c0590cae57f9f5858e0`
+- Prediction rows: `10,540`
+- Target: synthetic Oracle demand over `t+1..t+H`.
+- Oracle fields used as model features: `none`.
+
+## Point metrics
+
+|   horizon_days | model              |    n |   actual_sum |   prediction_sum |     mae |    rmse |   wmape |   wmape_percent |    bias |   mape_nonzero |   mape_nonzero_percent |   zero_actual_rows |
+|---------------:|:-------------------|-----:|-------------:|-----------------:|--------:|--------:|--------:|----------------:|--------:|---------------:|-----------------------:|-------------------:|
+|              1 | frozen_product_q50 | 4495 |   34205.0000 |       32949.6510 |  2.3840 |  3.5900 |  0.3133 |         31.3292 | -0.2793 |         0.4535 |                45.3474 |                490 |
+|              1 | croston_01         | 4495 |   34205.0000 |       33740.3413 |  2.4870 |  3.7331 |  0.3268 |         32.6819 | -0.1034 |         0.4583 |                45.8291 |                490 |
+|              1 | ewma_28d           | 4495 |   34205.0000 |       33446.0336 |  2.4879 |  3.7463 |  0.3269 |         32.6946 | -0.1688 |         0.4678 |                46.7847 |                490 |
+|              1 | moving_average_28d | 4495 |   34205.0000 |       33481.3929 |  2.4919 |  3.7661 |  0.3275 |         32.7474 | -0.1610 |         0.4687 |                46.8670 |                490 |
+|              1 | seasonal_naive_7d  | 4495 |   34205.0000 |       33266.0000 |  3.5989 |  5.6172 |  0.4729 |         47.2943 | -0.2089 |         0.6664 |                66.6439 |                490 |
+|              7 | moving_average_28d | 3565 |  187365.0000 |      187087.5000 |  9.1864 | 13.8617 |  0.1748 |         17.4790 | -0.0778 |         0.2736 |                27.3642 |                 20 |
+|              7 | ewma_28d           | 3565 |  187365.0000 |      187022.9146 |  9.2788 | 13.9711 |  0.1765 |         17.6549 | -0.0960 |         0.2757 |                27.5737 |                 20 |
+|              7 | croston_01         | 3565 |  187365.0000 |      188970.3744 |  9.3552 | 14.0816 |  0.1780 |         17.8002 |  0.4503 |         0.2617 |                26.1677 |                 20 |
+|              7 | frozen_product_q50 | 3565 |  187365.0000 |      192259.3795 |  9.5626 | 14.7218 |  0.1819 |         18.1949 |  1.3729 |         0.2503 |                25.0326 |                 20 |
+|              7 | seasonal_naive_7d  | 3565 |  187365.0000 |      184417.0000 | 12.1930 | 18.8103 |  0.2320 |         23.1996 | -0.8269 |         0.3479 |                34.7891 |                 20 |
+|             14 | frozen_product_q50 | 2480 |  262170.0000 |      275222.3106 | 14.2593 | 20.9916 |  0.1349 |         13.4886 |  5.2630 |         0.1859 |                18.5941 |                  1 |
+|             14 | moving_average_28d | 2480 |  262170.0000 |      260609.5000 | 15.2433 | 23.5362 |  0.1442 |         14.4195 | -0.6292 |         0.2200 |                22.0016 |                  1 |
+|             14 | ewma_28d           | 2480 |  262170.0000 |      262470.1210 | 15.9431 | 24.7648 |  0.1508 |         15.0814 |  0.1210 |         0.2254 |                22.5446 |                  1 |
+|             14 | croston_01         | 2480 |  262170.0000 |      265900.6234 | 16.2052 | 24.9653 |  0.1533 |         15.3294 |  1.5043 |         0.2013 |                20.1350 |                  1 |
+|             14 | seasonal_naive_7d  | 2480 |  262170.0000 |      256412.0000 | 23.4339 | 36.7922 |  0.2217 |         22.1673 | -2.3218 |         0.3118 |                31.1787 |                  1 |
+
+## Quantile diagnostics
+
+|   horizon_days |         n |   coverage_80 |   coverage_80_percent |   mean_interval_width |   pinball_q10 |   pinball_q50 |   pinball_q90 |   quantile_crossing_rows |
+|---------------:|----------:|--------------:|----------------------:|----------------------:|--------------:|--------------:|--------------:|-------------------------:|
+|         1.0000 | 4495.0000 |        0.7864 |               78.6429 |                7.0365 |        0.4401 |        1.1920 |        0.6113 |                   0.0000 |
+|         7.0000 | 3565.0000 |        0.8107 |               81.0659 |               27.9154 |        1.6744 |        4.7813 |        2.1715 |                   0.0000 |
+|        14.0000 | 2480.0000 |        0.8315 |               83.1452 |               47.4906 |        2.6250 |        7.1297 |        3.4248 |                   0.0000 |
+
+## Comparison with the strongest tested baseline
+
+|   horizon_days |   product_wmape_percent | best_baseline      |   best_baseline_wmape_percent |   relative_wmape_improvement_vs_best_baseline_percent | product_beats_best_baseline   |
+|---------------:|------------------------:|:-------------------|------------------------------:|------------------------------------------------------:|:------------------------------|
+|              1 |                 31.3292 | croston_01         |                       32.6819 |                                                4.1389 | True                          |
+|              7 |                 18.1949 | moving_average_28d |                       17.4790 |                                               -4.0958 | False                         |
+|             14 |                 13.4886 | moving_average_28d |                       14.4195 |                                                6.4555 | True                          |
+
+## 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.

evaluation/SHA256SUMS

@@ -0,0 +1,9 @@
+55b21bbb44f3b65869f297b2e91e4c2fcb28ea76b9e2eb25b2d1b531b0dc2ee0  EVALUATION_SUMMARY.md
+8313a3a9e96f8eca646b596f8c858ff9860bf0f0a2684044cb5cb6cc0eb1cc64  baseline_comparison.csv
+2af9dc873572ce23fd6caa70166b083870326130ca5bf250676840d4478050bf  evaluate_frozen_forecasts.py
+81d39e9671ce1b55c8018ff125cc3823143cac953c7dc7f455eb4fc6ac982e53  metrics.csv
+442e8689dae314af374e4bf14b887ea25cc0a9003ef7b22992f957a5b54266c6  metrics.json
+6d4f1b6a32315fee435f918fc4d7eda9d13975781daa009b4b5dbee65e045710  metrics_by_store.csv
+7c4d0dc495f1ded845d37763cb3478687b7d8ba8525a5456678743c1be4808a1  predictions.csv
+4a800bd78d0a114e0b38805619c4a90bcf5c3d70b549271331477fbfe22a328f  quantile_metrics.csv
+590c19bbe3167e29cf95821ef0c8c6b507b049b76db86a8a1a2b94290a8cf808  run_manifest.json

evaluation/baseline_comparison.csv

@@ -0,0 +1,4 @@
+horizon_days,product_wmape_percent,best_baseline,best_baseline_wmape_percent,relative_wmape_improvement_vs_best_baseline_percent,product_beats_best_baseline
+1,31.32924441023557,croston_01,32.681909132204524,4.138879147167244,True
+7,18.194885275997084,moving_average_28d,17.478984869105755,-4.095777942783671,False
+14,13.488613887038298,moving_average_28d,14.419460655299996,6.455489497934565,True

evaluation/evaluate_frozen_forecasts.py

@@ -0,0 +1,557 @@
+#!/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()

evaluation/metrics.csv

@@ -0,0 +1,16 @@
+horizon_days,model,n,actual_sum,prediction_sum,mae,rmse,wmape,wmape_percent,bias,mape_nonzero,mape_nonzero_percent,zero_actual_rows
+1,frozen_product_q50,4495,34205.0,32949.650997892764,2.384019588547514,3.5900039276845876,0.3132924441023557,31.32924441023557,-0.2792767524154025,0.4534735875060879,45.34735875060879,490
+1,croston_01,4495,34205.0,33740.34127127988,2.4869515058221485,3.733116825548437,0.3268190913220453,32.681909132204524,-0.10337235344162793,0.4582907852729936,45.82907852729936,490
+1,ewma_28d,4495,34205.0,33446.03356122295,2.4879196310025415,3.746266449425722,0.32694631607532304,32.6946316075323,-0.1688468161906683,0.4678467136548258,46.78467136548258,490
+1,moving_average_28d,4495,34205.0,33481.392857142855,2.4919354838709675,3.7660556155675997,0.32747405350095016,32.747405350095015,-0.1609804544732242,0.4686702849658086,46.86702849658086,490
+1,seasonal_naive_7d,4495,34205.0,33266.0,3.59888765294772,5.617210958687039,0.47294255225844173,47.294255225844175,-0.2088987764182425,0.666438714609271,66.6438714609271,490
+7,moving_average_28d,3565,187365.0,187087.5,9.186395511921459,13.86165996952744,0.17478984869105757,17.478984869105755,-0.07784011220196353,0.27364156081096036,27.364156081096034,20
+7,ewma_28d,3565,187365.0,187022.9146295055,9.278825931204487,13.97110566879584,0.17654852531019133,17.65485253101913,-0.09595662566465389,0.275737212515037,27.573721251503702,20
+7,croston_01,3565,187365.0,188970.37438391318,9.355235873067242,14.081569695165259,0.17800237978002678,17.800237978002677,0.4503153952070647,0.26167706853457445,26.167706853457446,20
+7,frozen_product_q50,3565,187365.0,192259.37951985394,9.562649873035607,14.72181875410442,0.18194885275997086,18.194885275997084,1.3728974810249464,0.2503262474408005,25.03262474408005,20
+7,seasonal_naive_7d,3565,187365.0,184417.0,12.192987377279103,18.81026546746541,0.23199637072025192,23.19963707202519,-0.8269284712482469,0.34789095468512293,34.78909546851229,20
+14,frozen_product_q50,2480,262170.0,275222.310598772,14.259314124051736,20.991553718900008,0.13488613887038298,13.488613887038298,5.263028467246779,0.18594104321190352,18.59410432119035,1
+14,moving_average_28d,2480,262170.0,260609.5,15.243346774193549,23.536208806857573,0.14419460655299995,14.419460655299996,-0.629233870967742,0.22001623227049832,22.001623227049834,1
+14,ewma_28d,2480,262170.0,262470.12103139434,15.943116721579381,24.764835084808933,0.1508140880707818,15.08140880707818,0.12101654491706937,0.22544594877913723,22.544594877913724,1
+14,croston_01,2480,262170.0,265900.62338449893,16.2052416115614,24.965328108267574,0.15329366135207032,15.32936613520703,1.5042836227818297,0.20134957293100236,20.134957293100236,1
+14,seasonal_naive_7d,2480,262170.0,256412.0,23.433870967741935,36.79224585206385,0.2216729602929397,22.16729602929397,-2.321774193548387,0.3117871657569443,31.17871657569443,1

evaluation/metrics.json

@@ -0,0 +1,290 @@
+{
+  "baseline_comparison": [
+    {
+      "best_baseline": "croston_01",
+      "best_baseline_wmape_percent": 32.681909132204524,
+      "horizon_days": 1,
+      "product_beats_best_baseline": true,
+      "product_wmape_percent": 31.32924441023557,
+      "relative_wmape_improvement_vs_best_baseline_percent": 4.138879147167244
+    },
+    {
+      "best_baseline": "moving_average_28d",
+      "best_baseline_wmape_percent": 17.478984869105755,
+      "horizon_days": 7,
+      "product_beats_best_baseline": false,
+      "product_wmape_percent": 18.194885275997084,
+      "relative_wmape_improvement_vs_best_baseline_percent": -4.095777942783671
+    },
+    {
+      "best_baseline": "moving_average_28d",
+      "best_baseline_wmape_percent": 14.419460655299996,
+      "horizon_days": 14,
+      "product_beats_best_baseline": true,
+      "product_wmape_percent": 13.488613887038298,
+      "relative_wmape_improvement_vs_best_baseline_percent": 6.455489497934565
+    }
+  ],
+  "point_metrics": [
+    {
+      "actual_sum": 34205.0,
+      "bias": -0.2792767524154025,
+      "horizon_days": 1,
+      "mae": 2.384019588547514,
+      "mape_nonzero": 0.4534735875060879,
+      "mape_nonzero_percent": 45.34735875060879,
+      "model": "frozen_product_q50",
+      "n": 4495,
+      "prediction_sum": 32949.650997892764,
+      "rmse": 3.5900039276845876,
+      "wmape": 0.3132924441023557,
+      "wmape_percent": 31.32924441023557,
+      "zero_actual_rows": 490
+    },
+    {
+      "actual_sum": 34205.0,
+      "bias": -0.10337235344162793,
+      "horizon_days": 1,
+      "mae": 2.4869515058221485,
+      "mape_nonzero": 0.4582907852729936,
+      "mape_nonzero_percent": 45.82907852729936,
+      "model": "croston_01",
+      "n": 4495,
+      "prediction_sum": 33740.34127127988,
+      "rmse": 3.733116825548437,
+      "wmape": 0.3268190913220453,
+      "wmape_percent": 32.681909132204524,
+      "zero_actual_rows": 490
+    },
+    {
+      "actual_sum": 34205.0,
+      "bias": -0.1688468161906683,
+      "horizon_days": 1,
+      "mae": 2.4879196310025415,
+      "mape_nonzero": 0.4678467136548258,
+      "mape_nonzero_percent": 46.78467136548258,
+      "model": "ewma_28d",
+      "n": 4495,
+      "prediction_sum": 33446.03356122295,
+      "rmse": 3.746266449425722,
+      "wmape": 0.32694631607532304,
+      "wmape_percent": 32.6946316075323,
+      "zero_actual_rows": 490
+    },
+    {
+      "actual_sum": 34205.0,
+      "bias": -0.1609804544732242,
+      "horizon_days": 1,
+      "mae": 2.4919354838709675,
+      "mape_nonzero": 0.4686702849658086,
+      "mape_nonzero_percent": 46.86702849658086,
+      "model": "moving_average_28d",
+      "n": 4495,
+      "prediction_sum": 33481.392857142855,
+      "rmse": 3.7660556155675997,
+      "wmape": 0.32747405350095016,
+      "wmape_percent": 32.747405350095015,
+      "zero_actual_rows": 490
+    },
+    {
+      "actual_sum": 34205.0,
+      "bias": -0.2088987764182425,
+      "horizon_days": 1,
+      "mae": 3.59888765294772,
+      "mape_nonzero": 0.666438714609271,
+      "mape_nonzero_percent": 66.6438714609271,
+      "model": "seasonal_naive_7d",
+      "n": 4495,
+      "prediction_sum": 33266.0,
+      "rmse": 5.617210958687039,
+      "wmape": 0.47294255225844173,
+      "wmape_percent": 47.294255225844175,
+      "zero_actual_rows": 490
+    },
+    {
+      "actual_sum": 187365.0,
+      "bias": -0.07784011220196353,
+      "horizon_days": 7,
+      "mae": 9.186395511921459,
+      "mape_nonzero": 0.27364156081096036,
+      "mape_nonzero_percent": 27.364156081096034,
+      "model": "moving_average_28d",
+      "n": 3565,
+      "prediction_sum": 187087.5,
+      "rmse": 13.86165996952744,
+      "wmape": 0.17478984869105757,
+      "wmape_percent": 17.478984869105755,
+      "zero_actual_rows": 20
+    },
+    {
+      "actual_sum": 187365.0,
+      "bias": -0.09595662566465389,
+      "horizon_days": 7,
+      "mae": 9.278825931204487,
+      "mape_nonzero": 0.275737212515037,
+      "mape_nonzero_percent": 27.573721251503702,
+      "model": "ewma_28d",
+      "n": 3565,
+      "prediction_sum": 187022.9146295055,
+      "rmse": 13.97110566879584,
+      "wmape": 0.17654852531019133,
+      "wmape_percent": 17.65485253101913,
+      "zero_actual_rows": 20
+    },
+    {
+      "actual_sum": 187365.0,
+      "bias": 0.4503153952070647,
+      "horizon_days": 7,
+      "mae": 9.355235873067242,
+      "mape_nonzero": 0.26167706853457445,
+      "mape_nonzero_percent": 26.167706853457446,
+      "model": "croston_01",
+      "n": 3565,
+      "prediction_sum": 188970.37438391318,
+      "rmse": 14.081569695165259,
+      "wmape": 0.17800237978002678,
+      "wmape_percent": 17.800237978002677,
+      "zero_actual_rows": 20
+    },
+    {
+      "actual_sum": 187365.0,
+      "bias": 1.3728974810249464,
+      "horizon_days": 7,
+      "mae": 9.562649873035607,
+      "mape_nonzero": 0.2503262474408005,
+      "mape_nonzero_percent": 25.03262474408005,
+      "model": "frozen_product_q50",
+      "n": 3565,
+      "prediction_sum": 192259.37951985394,
+      "rmse": 14.72181875410442,
+      "wmape": 0.18194885275997086,
+      "wmape_percent": 18.194885275997084,
+      "zero_actual_rows": 20
+    },
+    {
+      "actual_sum": 187365.0,
+      "bias": -0.8269284712482469,
+      "horizon_days": 7,
+      "mae": 12.192987377279103,
+      "mape_nonzero": 0.34789095468512293,
+      "mape_nonzero_percent": 34.78909546851229,
+      "model": "seasonal_naive_7d",
+      "n": 3565,
+      "prediction_sum": 184417.0,
+      "rmse": 18.81026546746541,
+      "wmape": 0.23199637072025192,
+      "wmape_percent": 23.19963707202519,
+      "zero_actual_rows": 20
+    },
+    {
+      "actual_sum": 262170.0,
+      "bias": 5.263028467246779,
+      "horizon_days": 14,
+      "mae": 14.259314124051736,
+      "mape_nonzero": 0.18594104321190352,
+      "mape_nonzero_percent": 18.59410432119035,
+      "model": "frozen_product_q50",
+      "n": 2480,
+      "prediction_sum": 275222.310598772,
+      "rmse": 20.991553718900008,
+      "wmape": 0.13488613887038298,
+      "wmape_percent": 13.488613887038298,
+      "zero_actual_rows": 1
+    },
+    {
+      "actual_sum": 262170.0,
+      "bias": -0.629233870967742,
+      "horizon_days": 14,
+      "mae": 15.243346774193549,
+      "mape_nonzero": 0.22001623227049832,
+      "mape_nonzero_percent": 22.001623227049834,
+      "model": "moving_average_28d",
+      "n": 2480,
+      "prediction_sum": 260609.5,
+      "rmse": 23.536208806857573,
+      "wmape": 0.14419460655299995,
+      "wmape_percent": 14.419460655299996,
+      "zero_actual_rows": 1
+    },
+    {
+      "actual_sum": 262170.0,
+      "bias": 0.12101654491706937,
+      "horizon_days": 14,
+      "mae": 15.943116721579381,
+      "mape_nonzero": 0.22544594877913723,
+      "mape_nonzero_percent": 22.544594877913724,
+      "model": "ewma_28d",
+      "n": 2480,
+      "prediction_sum": 262470.12103139434,
+      "rmse": 24.764835084808933,
+      "wmape": 0.1508140880707818,
+      "wmape_percent": 15.08140880707818,
+      "zero_actual_rows": 1
+    },
+    {
+      "actual_sum": 262170.0,
+      "bias": 1.5042836227818297,
+      "horizon_days": 14,
+      "mae": 16.2052416115614,
+      "mape_nonzero": 0.20134957293100236,
+      "mape_nonzero_percent": 20.134957293100236,
+      "model": "croston_01",
+      "n": 2480,
+      "prediction_sum": 265900.62338449893,
+      "rmse": 24.965328108267574,
+      "wmape": 0.15329366135207032,
+      "wmape_percent": 15.32936613520703,
+      "zero_actual_rows": 1
+    },
+    {
+      "actual_sum": 262170.0,
+      "bias": -2.321774193548387,
+      "horizon_days": 14,
+      "mae": 23.433870967741935,
+      "mape_nonzero": 0.3117871657569443,
+      "mape_nonzero_percent": 31.17871657569443,
+      "model": "seasonal_naive_7d",
+      "n": 2480,
+      "prediction_sum": 256412.0,
+      "rmse": 36.79224585206385,
+      "wmape": 0.2216729602929397,
+      "wmape_percent": 22.16729602929397,
+      "zero_actual_rows": 1
+    }
+  ],
+  "quantile_metrics": [
+    {
+      "coverage_80": 0.7864293659621802,
+      "coverage_80_percent": 78.64293659621802,
+      "horizon_days": 1,
+      "mean_interval_width": 7.03646395143773,
+      "n": 4495,
+      "pinball_q10": 0.44014761984234607,
+      "pinball_q50": 1.192009794273757,
+      "pinball_q90": 0.6113398602007618,
+      "quantile_crossing_rows": 0
+    },
+    {
+      "coverage_80": 0.8106591865357644,
+      "coverage_80_percent": 81.06591865357645,
+      "horizon_days": 7,
+      "mean_interval_width": 27.915441224950428,
+      "n": 3565,
+      "pinball_q10": 1.6743619017525346,
+      "pinball_q50": 4.781324936517803,
+      "pinball_q90": 2.1715224487756313,
+      "quantile_crossing_rows": 0
+    },
+    {
+      "coverage_80": 0.8314516129032258,
+      "coverage_80_percent": 83.14516129032258,
+      "horizon_days": 14,
+      "mean_interval_width": 47.49057793632384,
+      "n": 2480,
+      "pinball_q10": 2.6250437142517526,
+      "pinball_q50": 7.129657062025868,
+      "pinball_q90": 3.4247838153423285,
+      "quantile_crossing_rows": 0
+    }
+  ]
+}

evaluation/metrics_by_store.csv

@@ -0,0 +1,76 @@
+horizon_days,store_id,model,n,actual_sum,prediction_sum,mae,rmse,wmape,wmape_percent,bias,mape_nonzero,mape_nonzero_percent,zero_actual_rows
+1,S01,frozen_product_q50,899,7547.0,7008.038742326617,2.5935169972314185,4.097968219941704,0.3089402120724851,30.894021207248514,-0.5995119662662783,0.4651960112731215,46.51960112731215,96
+1,S01,moving_average_28d,899,7547.0,7342.928571428571,2.605434609883998,4.151300048569791,0.3103598402392625,31.03598402392625,-0.2269982520260607,0.4771929584471576,47.71929584471576,96
+1,S01,ewma_28d,899,7547.0,7362.376898207124,2.6298176935207134,4.170678840704454,0.31326435755599863,31.32643575559986,-0.20536496306215377,0.4811680586963763,48.11680586963763,96
+1,S01,croston_01,899,7547.0,7396.6709013561,2.650207908828989,4.181784577878802,0.31569324367791984,31.56932436779198,-0.1672181297484979,0.4743040774308254,47.43040774308254,96
+1,S01,seasonal_naive_7d,899,7547.0,7327.0,3.7508342602892104,5.9195571490555245,0.4468000530011925,44.680005300119255,-0.2447163515016685,0.6705668920662955,67.05668920662956,96
+1,S02,frozen_product_q50,899,6169.0,6103.558551832481,2.274512112540314,3.3969334573158525,0.33146156413904077,33.14615641390408,-0.07279360196609436,0.4635122679287246,46.35122679287246,113
+1,S02,moving_average_28d,899,6169.0,6006.714285714286,2.2812648975051646,3.499202132455104,0.3324456383298984,33.24456383298984,-0.18051803591291904,0.4518102445684703,45.18102445684703,113
+1,S02,ewma_28d,899,6169.0,6002.653776014198,2.293449404968332,3.4823627409463374,0.33422127007076197,33.4221270070762,-0.18503473190856704,0.4535604213156051,45.35604213156051,113
+1,S02,croston_01,899,6169.0,6076.74820783173,2.3225944847590476,3.495677079445841,0.33846854300508733,33.84685430050874,-0.10261600908595044,0.4492188728839176,44.92188728839176,113
+1,S02,seasonal_naive_7d,899,6169.0,6034.0,3.460511679644049,5.358660201451128,0.5042956719079267,50.429567190792675,-0.15016685205784205,0.6690980777354153,66.90980777354153,113
+1,S03,frozen_product_q50,899,6257.0,6229.845008838718,2.3843499402369037,3.598903799876326,0.3425812044546869,34.2581204454687,-0.03020577437295042,0.4735864467733925,47.35864467733925,114
+1,S03,moving_average_28d,899,6257.0,6219.464285714286,2.4517320832671223,3.6887882107886094,0.3522626087353593,35.22626087353593,-0.04175274114095024,0.479831986762188,47.983198676218805,114
+1,S03,ewma_28d,899,6257.0,6190.993647285974,2.4544359508713565,3.684497263279992,0.352651097943639,35.2651097943639,-0.0734219718732212,0.4809204149858744,48.09204149858744,114
+1,S03,croston_01,899,6257.0,6301.400825248424,2.4544722616158445,3.706680413873509,0.35265631503798056,35.265631503798055,0.049389127083898025,0.46934171807236413,46.93417180723641,114
+1,S03,seasonal_naive_7d,899,6257.0,6228.0,3.589543937708565,5.759066943337441,0.5157423685472271,51.57423685472271,-0.03225806451612903,0.6900051406557649,69.0005140655765,114
+1,S04,frozen_product_q50,899,7372.0,6984.017864335636,2.3921623266663943,3.4669573903120314,0.2917191985449116,29.171919854491158,-0.4315707849436755,0.44372248665884567,44.37224866588457,70
+1,S04,ewma_28d,899,7372.0,7355.33357793934,2.5086676679494566,3.549284418204959,0.30592678153642994,30.592678153642993,-0.01853884545123516,0.48228832694506796,48.22883269450679,70
+1,S04,croston_01,899,7372.0,7345.213618088715,2.5155683380088227,3.5656880149124572,0.306768303834771,30.6768303834771,-0.029795752960273716,0.46645502345834206,46.6455023458342,70
+1,S04,moving_average_28d,899,7372.0,7372.5,2.5306689972985854,3.5964020388789333,0.30860979769010155,30.860979769010154,0.0005561735261400925,0.48596514076779673,48.59651407677967,70
+1,S04,seasonal_naive_7d,899,7372.0,7218.0,3.597330367074527,5.445043464608839,0.43868692349430277,43.868692349430276,-0.17130144605116795,0.6764235142502553,67.64235142502552,70
+1,S05,frozen_product_q50,899,6860.0,6624.190830559316,2.2755565660625425,3.337100305934235,0.29821069284114077,29.821069284114074,-0.2623016345280138,0.4222909927185685,42.22909927185685,97
+1,S05,croston_01,899,6860.0,6620.307718754913,2.4919145358980392,3.676970894110038,0.32656431017089466,32.65643101708947,-0.2666210024973157,0.43189267284396016,43.18926728439602,97
+1,S05,ewma_28d,899,6860.0,6534.675661776311,2.553227437702849,3.8050242225513657,0.33459933913919265,33.459933913919265,-0.3618735686581641,0.4407856572306093,44.07856572306093,97
+1,S05,moving_average_28d,899,6860.0,6539.785714285714,2.5905768313999684,3.859650340562184,0.339493960849646,33.9493960849646,-0.3561894168123312,0.4478584565079267,44.78584565079267,97
+1,S05,seasonal_naive_7d,899,6860.0,6459.0,3.596218020022247,5.585118475654458,0.47128279883381924,47.12827988338192,-0.4460511679644049,0.626311246699029,62.6311246699029,97
+7,S01,moving_average_28d,713,41207.0,40925.75,8.798387096774194,12.830979259267515,0.1522374839226345,15.22374839226345,-0.3944600280504909,0.2354115586193961,23.54115586193961,9
+7,S01,ewma_28d,713,41207.0,41047.79068578307,9.005937797512788,13.22935571749057,0.15582870991886374,15.582870991886372,-0.22329497085124014,0.2392031871060224,23.92031871060224,9
+7,S01,croston_01,713,41207.0,41192.757039714874,9.227868897234726,13.891407339083585,0.15966875831116945,15.966875831116944,-0.019976101381663,0.21748309657291126,21.748309657291127,9
+7,S01,frozen_product_q50,713,41207.0,40642.93643623597,10.35880972334837,15.758798281771412,0.1792372978558834,17.923729785588343,-0.7911129926564245,0.22757371540125257,22.757371540125256,9
+7,S01,seasonal_naive_7d,713,41207.0,40685.0,11.708274894810659,17.41531302129633,0.20258693911228676,20.258693911228676,-0.7321178120617111,0.3160898098378841,31.60898098378841,9
+7,S02,moving_average_28d,713,33852.0,33491.25,8.024894810659186,12.03261186305136,0.1690225097483162,16.90225097483162,-0.5059607293127629,0.28903693390177115,28.903693390177114,8
+7,S02,ewma_28d,713,33852.0,33523.37178970822,8.277459624942532,12.269187668017189,0.17434209832754416,17.434209832754416,-0.4609091308440112,0.2901937001159493,29.01937001159493,8
+7,S02,croston_01,713,33852.0,34043.1404597861,8.537787539855438,12.794603097903096,0.17982519543651565,17.982519543651566,0.26807918623576565,0.3055641396760597,30.556413967605973,8
+7,S02,frozen_product_q50,713,33852.0,35431.11257229469,8.861727075258273,14.116776528571721,0.1866480977389563,18.66480977389563,2.214744140665759,0.2706420937291821,27.064209372918206,8
+7,S02,seasonal_naive_7d,713,33852.0,33158.0,11.800841514726507,18.045135867353512,0.24855252274607112,24.855252274607114,-0.9733520336605891,0.38526601738669763,38.52660173866976,8
+7,S03,moving_average_28d,713,34504.0,34844.0,9.326086956521738,13.729638755796032,0.19271678645954093,19.271678645954093,0.47685834502103785,0.3012077032491755,30.12077032491755,3
+7,S03,ewma_28d,713,34504.0,34859.729017562124,9.728361599425838,14.342939309868816,0.20102949862017802,20.1029498620178,0.49891867820774904,0.31154310335915864,31.154310335915863,3
+7,S03,croston_01,713,34504.0,35498.68199457943,10.065145216156397,15.193984851812544,0.2079888864803939,20.798888648039387,1.3950659110510883,0.298083921360115,29.808392136011502,3
+7,S03,frozen_product_q50,713,34504.0,36429.038828024975,10.279213709269445,16.12097617899276,0.21241245579379536,21.241245579379534,2.6999142048036195,0.2802818377205175,28.02818377205175,3
+7,S03,seasonal_naive_7d,713,34504.0,34819.0,12.730715287517532,19.714120924222485,0.26307094829584976,26.307094829584976,0.4417952314165498,0.38096309226709213,38.09630922670921,3
+7,S04,ewma_28d,713,40236.0,41130.257515187724,9.540718099803591,13.339377167000292,0.16906581183914804,16.906581183914806,1.2542181138677722,0.2691191327734908,26.911913277349083,0
+7,S04,croston_01,713,40236.0,41084.83547149017,9.549775147975652,13.715767715145837,0.16922630680253106,16.922630680253107,1.190512582735173,0.2392290342631686,23.92290342631686,0
+7,S04,moving_average_28d,713,40236.0,41316.25,9.60483870967742,13.349083074656527,0.17020205785863407,17.020205785863407,1.5150771388499298,0.2671855799376947,26.71855799376947,0
+7,S04,frozen_product_q50,713,40236.0,41138.4585675644,10.261090916277801,14.840325492899849,0.1818311418457618,18.18311418457618,1.265720291114162,0.2510373588293576,25.103735882935762,0
+7,S04,seasonal_naive_7d,713,40236.0,39926.0,12.746143057503506,18.3856917353974,0.22586738244358287,22.586738244358287,-0.43478260869565216,0.3340551154007905,33.405511540079054,0
+7,S05,frozen_product_q50,713,37566.0,38617.833115733905,8.052407941024144,12.486612855432128,0.1528341282529472,15.283412825294722,1.475221761197618,0.2221630207982935,22.21630207982935,0
+7,S05,croston_01,713,37566.0,37150.9594183426,9.395602564114002,14.690149740993796,0.1783278663742023,17.83278663742023,-0.5821046026050404,0.2481129086967657,24.81129086967657,0
+7,S05,ewma_28d,713,37566.0,36461.76562126438,9.841652534337683,16.332521485070885,0.18679386298734943,18.679386298734943,-1.5487158187035392,0.26847864096622803,26.847864096622803,0
+7,S05,moving_average_28d,713,37566.0,36510.25,10.177769985974754,16.870367786644785,0.19317334824043017,19.31733482404302,-1.4807152875175316,0.27517218738227195,27.517218738227196,0
+7,S05,seasonal_naive_7d,713,37566.0,35829.0,11.978962131837307,20.33544084296428,0.22735984666986103,22.735984666986106,-2.4361851332398317,0.323237830553814,32.3237830553814,0
+14,S01,moving_average_28d,496,57910.0,56841.0,13.433467741935484,20.361179262698464,0.11505784838542565,11.505784838542565,-2.155241935483871,0.20087221940414207,20.087221940414207,0
+14,S01,frozen_product_q50,496,57910.0,58204.98100863425,13.803100668390389,21.031549461720267,0.11822375982596499,11.8223759825965,0.5947197754722764,0.1718111283038845,17.18111283038845,0
+14,S01,ewma_28d,496,57910.0,57229.383817786744,14.89455426021201,22.526844871683696,0.12757207586021682,12.757207586021684,-1.3722100447848053,0.2092230354915216,20.92230354915216,0
+14,S01,croston_01,496,57910.0,57389.357671930295,15.62997130863394,24.325876962672808,0.13387093367436426,13.387093367436426,-1.0496821130437577,0.17939433270515398,17.939433270515398,0
+14,S01,seasonal_naive_7d,496,57910.0,55812.0,22.70967741935484,34.1754250510415,0.19450872042825074,19.450872042825072,-4.229838709677419,0.3039472313949942,30.39472313949942,0
+14,S02,frozen_product_q50,496,47264.0,50485.49647920423,12.6444871299441,18.455588418595106,0.13269434699670518,13.269434699670517,6.4949525790407785,0.19921168285096308,19.921168285096307,1
+14,S02,moving_average_28d,496,47264.0,46641.0,13.088709677419354,19.002387371234477,0.13735612728503724,13.735612728503725,-1.2560483870967742,0.20985079400046414,20.985079400046413,1
+14,S02,ewma_28d,496,47264.0,47115.379630169635,13.328824055888694,19.283110508285432,0.13987594642266402,13.987594642266403,-0.2996378423999212,0.21514165829371168,21.51416582937117,1
+14,S02,croston_01,496,47264.0,47905.14761469116,13.717048009688243,20.51881723666265,0.14395006374418942,14.39500637441894,1.292636319941848,0.2163957448947876,21.63957448947876,1
+14,S02,seasonal_naive_7d,496,47264.0,46480.0,20.181451612903224,31.29155301870747,0.21178909952606634,21.178909952606634,-1.5806451612903225,0.31372179184253124,31.372179184253124,1
+14,S03,moving_average_28d,496,48373.0,48341.0,15.286290322580646,21.628134678640674,0.15674033034957519,15.674033034957517,-0.06451612903225806,0.24517335676390203,24.517335676390204,0
+14,S03,frozen_product_q50,496,48373.0,52072.995370655175,16.645790701140122,24.23823940547198,0.1706801767053005,17.06801767053005,7.459668085998339,0.2040305755742317,20.40305755742317,0
+14,S03,ewma_28d,496,48373.0,48919.650871919985,17.14817225079005,24.645128037230805,0.17583142324006917,17.583142324006918,1.1021186933870668,0.2594107125820982,25.94107125820982,0
+14,S03,croston_01,496,48373.0,50145.475996192894,18.395445616077378,27.338263577528032,0.18862053264371403,18.862053264371404,3.5735403149050344,0.2379882472200417,23.79882472200417,0
+14,S03,seasonal_naive_7d,496,48373.0,48464.0,25.941532258064516,39.725088562397474,0.2659954933537304,26.599549335373037,0.18346774193548387,0.3489111337228377,34.89111337228377,0
+14,S04,frozen_product_q50,496,56240.0,59096.926701870834,15.625830463045663,21.90985715646559,0.13780960010082946,13.780960010082946,5.759932866675063,0.19088058993353935,19.088058993353936,0
+14,S04,ewma_28d,496,56240.0,58022.70828513917,16.649460454644995,23.838215360892065,0.14683734682617208,14.683734682617207,3.594169929716068,0.22685816963001115,22.685816963001116,0
+14,S04,croston_01,496,56240.0,58048.0209546545,16.67752361436658,24.696730741936456,0.14708484553210927,14.708484553210926,3.6452035376098877,0.1837238431670304,18.372384316703037,0
+14,S04,moving_average_28d,496,56240.0,57794.5,16.982862903225808,23.954688829861347,0.14977773826458038,14.977773826458037,3.1340725806451615,0.22591838301476064,22.591838301476063,0
+14,S04,seasonal_naive_7d,496,56240.0,55674.0,24.302419354838708,35.56150963811766,0.21433143669985774,21.433143669985775,-1.1411290322580645,0.3023863168449406,30.23863168449406,0
+14,S05,frozen_product_q50,496,52383.0,55361.91103840753,12.577361657738416,18.781845780866554,0.11909152553764112,11.90915255376411,6.005869029047432,0.16379799471875187,16.379799471875188,0
+14,S05,croston_01,496,52383.0,52412.62114703008,16.60621950904086,27.317736403269404,0.15723965554634645,15.723965554634644,0.05972005449613455,0.18927603169211876,18.927603169211878,0
+14,S05,moving_average_28d,496,52383.0,50992.0,17.425403225806452,30.862500734919443,0.16499627741824638,16.49962774182464,-2.8044354838709675,0.21824591333400092,21.824591333400093,0
+14,S05,ewma_28d,496,52383.0,51182.9984263788,17.694572586361154,31.8048006687276,0.16754496693269064,16.75449669326906,-2.4193580113330615,0.21657539311913923,21.657539311913922,0
+14,S05,seasonal_naive_7d,496,52383.0,49982.0,24.034274193548388,42.17252870553715,0.22757383120478017,22.757383120478018,-4.840725806451613,0.28997325543523567,28.997325543523566,0

evaluation/quantile_metrics.csv

@@ -0,0 +1,4 @@
+horizon_days,n,coverage_80,coverage_80_percent,mean_interval_width,pinball_q10,pinball_q50,pinball_q90,quantile_crossing_rows
+1,4495,0.7864293659621802,78.64293659621802,7.03646395143773,0.44014761984234607,1.192009794273757,0.6113398602007618,0
+7,3565,0.8106591865357644,81.06591865357645,27.915441224950428,1.6743619017525346,4.781324936517803,2.1715224487756313,0
+14,2480,0.8314516129032258,83.14516129032258,47.49057793632384,2.6250437142517526,7.129657062025868,3.4247838153423285,0

evaluation/run_manifest.json

@@ -0,0 +1,51 @@
+{
+  "artifact_manifest_sha256": "c3689bdf67ecb25edb5927584641446374a4feffad686fcd7d44ff6bf73561ac",
+  "baseline_definitions": {
+    "croston_01": "causal Croston alpha=0.1 daily rate times H",
+    "ewma_28d": "causal span-28 observed-sales EWMA times H",
+    "moving_average_28d": "28-day observed-sales mean times H",
+    "seasonal_naive_7d": "repeat the last seven observed-sales days causally"
+  },
+  "dataset_id": "demo-retail-v1",
+  "evaluation_name": "RestockIQ exact frozen-artifact June 2024 evaluation",
+  "evaluation_period": [
+    "2024-06-01",
+    "2024-06-30"
+  ],
+  "evaluator_sha256": "2af9dc873572ce23fd6caa70166b083870326130ca5bf250676840d4478050bf",
+  "git_remote": "https://github.com/Dzakiyaadila/triple-sigma.git",
+  "git_sha": "06ae958730f87cc44b3a5dbb7094fb8a3c88f7a3",
+  "git_status_porcelain": "",
+  "horizons": [
+    1,
+    7,
+    14
+  ],
+  "label": "units_demanded_est",
+  "label_role": "synthetic Oracle evaluation-only; never a model feature",
+  "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."
+  ],
+  "model_version": "restockiq-demand-v1-a067286b7c1e",
+  "oracle_fields_used_as_features": [],
+  "origin_rule": "decision date t; target is Oracle t+1 through t+H",
+  "prediction_rows": 10540,
+  "runtime_versions": {
+    "lightgbm": "4.7.0",
+    "numpy": "2.2.6",
+    "pandas": "2.2.3",
+    "python": "3.12.13",
+    "scipy": "1.14.1",
+    "sqlalchemy": "2.0.35"
+  },
+  "skus": 31,
+  "status": "completed_synthetic_controlled_evaluation",
+  "stores": 5,
+  "training_cutoff": "2024-05-31",
+  "training_data_hash": "a067286b7c1e966de43f3db9e1d24a7f8a440d236d94c24365c6d717ca44399c",
+  "workbook": "RestockIQ_Dataset_Sintetis.xlsx",
+  "workbook_sha256": "9d13eb22b2d2a2acecc46661d4984420e8390861a3bf9c0590cae57f9f5858e0"
+}

manifest.json

@@ -0,0 +1,332 @@
+{
+  "artifact_schema_version": 1,
+  "created_at": "2026-08-08T13:36:18.291254+00:00",
+  "forecast_signal": "rolling_origin_reconstructed_demand",
+  "forecast_target": "t+1_through_t+H_cumulative_reconstructed_demand",
+  "forecasts": {
+    "1": {
+      "calibration_adjustment": 0.03861231249552724,
+      "calibration_date_max": "2024-05-30",
+      "calibration_date_min": "2024-05-10",
+      "calibration_target": "future_cumulative_reconstructed_demand",
+      "feature_columns": [
+        "sig_lag_1",
+        "sig_lag_2",
+        "sig_lag_3",
+        "sig_lag_7",
+        "sig_lag_14",
+        "sig_lag_21",
+        "sig_lag_28",
+        "sig_roll_mean_7",
+        "sig_roll_mean_14",
+        "sig_roll_mean_28",
+        "sig_roll_std_7",
+        "sig_roll_std_14",
+        "sig_roll_std_28",
+        "sig_zero_rate_7",
+        "sig_zero_rate_14",
+        "sig_zero_rate_28",
+        "promo_flag",
+        "is_weekend",
+        "is_holiday",
+        "is_payday_week",
+        "day_of_month",
+        "month",
+        "unit_cost_rp",
+        "unit_price_rp",
+        "margin_rp",
+        "shelf_life_days",
+        "is_perishable",
+        "promised_lead_time_days"
+      ],
+      "medians": {
+        "day_of_month": 16.0,
+        "is_holiday": 0.0,
+        "is_payday_week": 0.0,
+        "is_perishable": 0.0,
+        "is_weekend": 0.0,
+        "margin_rp": 4787.0,
+        "month": 3.0,
+        "promised_lead_time_days": 3.0,
+        "promo_flag": 0.0,
+        "shelf_life_days": 292.0,
+        "sig_lag_1": 5.0,
+        "sig_lag_14": 5.0,
+        "sig_lag_2": 5.0,
+        "sig_lag_21": 5.0,
+        "sig_lag_28": 5.0,
+        "sig_lag_3": 5.0,
+        "sig_lag_7": 5.0,
+        "sig_roll_mean_14": 5.7005760535559915,
+        "sig_roll_mean_28": 5.857142857142857,
+        "sig_roll_mean_7": 5.428571428571429,
+        "sig_roll_std_14": 2.400970831230673,
+        "sig_roll_std_28": 2.514245129542576,
+        "sig_roll_std_7": 2.217355782608345,
+        "sig_zero_rate_14": 0.0,
+        "sig_zero_rate_28": 0.0,
+        "sig_zero_rate_7": 0.0,
+        "unit_cost_rp": 31339.0,
+        "unit_price_rp": 35700.0
+      },
+      "model_files": {
+        "0.10": "forecast_h1_q10.txt",
+        "0.50": "forecast_h1_q50.txt",
+        "0.90": "forecast_h1_q90.txt"
+      },
+      "model_fit_date_max": "2024-05-09",
+      "training_date_max": "2024-05-30",
+      "training_date_min": "2024-01-01",
+      "training_rows": 23405
+    },
+    "14": {
+      "calibration_adjustment": 1.5054051780677558,
+      "calibration_date_max": "2024-05-17",
+      "calibration_date_min": "2024-04-27",
+      "calibration_target": "future_cumulative_reconstructed_demand",
+      "feature_columns": [
+        "sig_lag_1",
+        "sig_lag_2",
+        "sig_lag_3",
+        "sig_lag_7",
+        "sig_lag_14",
+        "sig_lag_21",
+        "sig_lag_28",
+        "sig_roll_mean_7",
+        "sig_roll_mean_14",
+        "sig_roll_mean_28",
+        "sig_roll_std_7",
+        "sig_roll_std_14",
+        "sig_roll_std_28",
+        "sig_zero_rate_7",
+        "sig_zero_rate_14",
+        "sig_zero_rate_28",
+        "promo_flag",
+        "is_weekend",
+        "is_holiday",
+        "is_payday_week",
+        "day_of_month",
+        "month",
+        "unit_cost_rp",
+        "unit_price_rp",
+        "margin_rp",
+        "shelf_life_days",
+        "is_perishable",
+        "promised_lead_time_days"
+      ],
+      "medians": {
+        "day_of_month": 14.0,
+        "is_holiday": 0.0,
+        "is_payday_week": 0.0,
+        "is_perishable": 0.0,
+        "is_weekend": 0.0,
+        "margin_rp": 4787.0,
+        "month": 3.0,
+        "promised_lead_time_days": 3.0,
+        "promo_flag": 0.0,
+        "shelf_life_days": 292.0,
+        "sig_lag_1": 5.0,
+        "sig_lag_14": 5.0,
+        "sig_lag_2": 5.0,
+        "sig_lag_21": 5.0,
+        "sig_lag_28": 5.0,
+        "sig_lag_3": 5.0,
+        "sig_lag_7": 5.0,
+        "sig_roll_mean_14": 5.656655617050996,
+        "sig_roll_mean_28": 5.821428571428571,
+        "sig_roll_mean_7": 5.428571428571429,
+        "sig_roll_std_14": 2.4074094507812367,
+        "sig_roll_std_28": 2.5084924177763783,
+        "sig_roll_std_7": 2.225394561056746,
+        "sig_zero_rate_14": 0.0,
+        "sig_zero_rate_28": 0.0,
+        "sig_zero_rate_7": 0.0,
+        "unit_cost_rp": 31339.0,
+        "unit_price_rp": 35700.0
+      },
+      "model_files": {
+        "0.10": "forecast_h14_q10.txt",
+        "0.50": "forecast_h14_q50.txt",
+        "0.90": "forecast_h14_q90.txt"
+      },
+      "model_fit_date_max": "2024-04-26",
+      "training_date_max": "2024-05-17",
+      "training_date_min": "2024-01-01",
+      "training_rows": 21390
+    },
+    "7": {
+      "calibration_adjustment": 1.219129646667811,
+      "calibration_date_max": "2024-05-24",
+      "calibration_date_min": "2024-05-04",
+      "calibration_target": "future_cumulative_reconstructed_demand",
+      "feature_columns": [
+        "sig_lag_1",
+        "sig_lag_2",
+        "sig_lag_3",
+        "sig_lag_7",
+        "sig_lag_14",
+        "sig_lag_21",
+        "sig_lag_28",
+        "sig_roll_mean_7",
+        "sig_roll_mean_14",
+        "sig_roll_mean_28",
+        "sig_roll_std_7",
+        "sig_roll_std_14",
+        "sig_roll_std_28",
+        "sig_zero_rate_7",
+        "sig_zero_rate_14",
+        "sig_zero_rate_28",
+        "promo_flag",
+        "is_weekend",
+        "is_holiday",
+        "is_payday_week",
+        "day_of_month",
+        "month",
+        "unit_cost_rp",
+        "unit_price_rp",
+        "margin_rp",
+        "shelf_life_days",
+        "is_perishable",
+        "promised_lead_time_days"
+      ],
+      "medians": {
+        "day_of_month": 15.0,
+        "is_holiday": 0.0,
+        "is_payday_week": 0.0,
+        "is_perishable": 0.0,
+        "is_weekend": 0.0,
+        "margin_rp": 4787.0,
+        "month": 3.0,
+        "promised_lead_time_days": 3.0,
+        "promo_flag": 0.0,
+        "shelf_life_days": 292.0,
+        "sig_lag_1": 5.0,
+        "sig_lag_14": 5.0,
+        "sig_lag_2": 5.0,
+        "sig_lag_21": 5.0,
+        "sig_lag_28": 5.0,
+        "sig_lag_3": 5.0,
+        "sig_lag_7": 5.0,
+        "sig_roll_mean_14": 5.661661141858831,
+        "sig_roll_mean_28": 5.850288026777996,
+        "sig_roll_mean_7": 5.428571428571429,
+        "sig_roll_std_14": 2.4053511772118203,
+        "sig_roll_std_28": 2.5058866383772487,
+        "sig_roll_std_7": 2.214669705568283,
+        "sig_zero_rate_14": 0.0,
+        "sig_zero_rate_28": 0.0,
+        "sig_zero_rate_7": 0.0,
+        "unit_cost_rp": 31339.0,
+        "unit_price_rp": 35700.0
+      },
+      "model_files": {
+        "0.10": "forecast_h7_q10.txt",
+        "0.50": "forecast_h7_q50.txt",
+        "0.90": "forecast_h7_q90.txt"
+      },
+      "model_fit_date_max": "2024-05-03",
+      "training_date_max": "2024-05-24",
+      "training_date_min": "2024-01-01",
+      "training_rows": 22475
+    }
+  },
+  "horizons": [
+    1,
+    7,
+    14
+  ],
+  "library_versions": {
+    "lightgbm": "4.7.0",
+    "numpy": "2.2.6",
+    "pandas": "2.2.2"
+  },
+  "oracle_fields_used_as_features": [],
+  "oracle_forbidden_fields": [
+    "avg_daily_demand_per_store",
+    "cash_locked_in_stock_rp",
+    "demand_profile",
+    "units_demanded_est"
+  ],
+  "quantiles": [
+    0.1,
+    0.5,
+    0.9
+  ],
+  "reconstruction": {
+    "feature_columns": [
+      "lag_1",
+      "lag_2",
+      "lag_3",
+      "lag_7",
+      "lag_14",
+      "lag_21",
+      "lag_28",
+      "roll_mean_7",
+      "roll_mean_14",
+      "roll_mean_28",
+      "roll_std_7",
+      "roll_std_14",
+      "roll_std_28",
+      "zero_rate_7",
+      "zero_rate_14",
+      "zero_rate_28",
+      "ewm_7",
+      "ewm_28",
+      "past_stockout_rate_28",
+      "promo_flag",
+      "is_weekend",
+      "is_holiday",
+      "is_payday_week",
+      "day_of_month",
+      "month",
+      "unit_cost_rp",
+      "unit_price_rp",
+      "margin_rp",
+      "shelf_life_days",
+      "is_perishable",
+      "promised_lead_time_days"
+    ],
+    "medians": {
+      "day_of_month": 16.0,
+      "ewm_28": 5.8243961187092514,
+      "ewm_7": 5.13282937654271,
+      "is_holiday": 0.0,
+      "is_payday_week": 0.0,
+      "is_perishable": 0.0,
+      "is_weekend": 0.0,
+      "lag_1": 5.0,
+      "lag_14": 5.0,
+      "lag_2": 5.0,
+      "lag_21": 5.0,
+      "lag_28": 5.0,
+      "lag_3": 5.0,
+      "lag_7": 5.0,
+      "margin_rp": 4787.0,
+      "month": 3.0,
+      "past_stockout_rate_28": 0.0,
+      "promised_lead_time_days": 3.0,
+      "promo_flag": 0.0,
+      "roll_mean_14": 5.214285714285714,
+      "roll_mean_28": 5.464285714285714,
+      "roll_mean_7": 5.0,
+      "roll_std_14": 2.6196635972272015,
+      "roll_std_28": 2.738415182157349,
+      "roll_std_7": 2.370453040886408,
+      "shelf_life_days": 292.0,
+      "unit_cost_rp": 31339.0,
+      "unit_price_rp": 35700.0,
+      "zero_rate_14": 0.07142857142857142,
+      "zero_rate_28": 0.07142857142857142,
+      "zero_rate_7": 0.0
+    },
+    "model_file": "reconstruction_censor_no_inventory.txt",
+    "training_rows": 21558
+  },
+  "reconstruction_label": "uncensored_observed_units_sold",
+  "reconstruction_variant": "censor_no_inventory",
+  "seed": 20260723,
+  "training_cutoff": "2024-05-31",
+  "training_data_hash": "a067286b7c1e966de43f3db9e1d24a7f8a440d236d94c24365c6d717ca44399c",
+  "training_dataset_id": "demo-retail-v1",
+  "version": "restockiq-demand-v1-a067286b7c1e"
+}