Store × SKU Demand Forecasting

A large retail group replenishes 4,000+ SKUs across 33 stores every day. Too little stock loses the sale; too much ties up cash and spoils. One model per SKU-store series means more than 130,000 models to train and monitor, and naive forecasts ("same day last week") ignore promotions, seasonality and the way one product steals demand from another.

• Users: replenishment, production-planning and inventory-allocation teams, and the downstream planning applications that call the forecasts.
• Requirement: a forecast per SKU × store, with a confidence indicator and model metadata, available through an API.
• Not GenAI. This is classical applied ML, where a number has to be right every morning.

Daily sales history per SKU × store, promotions, calendar effects, product and store attributes. For the research rebuild, a product/store graph adds substitutes, complements, store similarity and proximity.

The rebuild runs on simulated data at demo scale — 6 stores × 200 SKUs × 730 days — with a time-based holdout and a 7-day embargo between train and test:

| Split | Rows | |---|---| | Train | 777,600 | | Validation | 8,400 | | Test | 55,200 |

Retraining, validation, experiment tracking and promotion are scheduled with Airflow, tracked in MLflow and packaged in Docker.

• Global and hierarchical forecasting models in Python, SQL, scikit-learn and LightGBM / XGBoost.
• FastAPI endpoints returning SKU-store forecasts, confidence indicators and model metadata.
• The automated ingest → retrain → validate → promote workflow.
• A research rebuild that tests a specific hypothesis: do graph features help, especially for new products?

◇ Illustrative · feature construction pattern, synthetic column names

def add_features(df: pd.DataFrame) -> pd.DataFrame:
    df = df.sort_values(["store", "sku", "date"])
    g = df.groupby(["store", "sku"])["units"]
    for lag in (1, 7, 14, 28):
        df[f"lag_{lag}"] = g.shift(lag)                 # only past values: no leakage
    for w in (7, 28):
        past = g.shift(1)
        df[f"rmean_{w}"] = past.groupby([df.store, df.sku]).transform(lambda s: s.rolling(w).mean())
        df[f"rstd_{w}"] = past.groupby([df.store, df.sku]).transform(lambda s: s.rolling(w).std())
    df["item_age"] = (df["date"] - df.groupby("sku")["date"].transform("min")).dt.days
    return df

• One global model instead of thousands of local ones: series share statistical strength, and new SKUs borrow from similar ones.
• A p90 quantile head alongside the point forecast, because safety stock is a quantile question, not a mean question.
• A three-arm ablation for the graph hypothesis: seasonal-naive, LightGBM without graph features, LightGBM with them. Cold-start subsets by item age isolate where graph context should help most.

Research rebuild, simulated data, test set:

| Arm | Test WMAPE | |---|---| | Seasonal-naive | 0.2273 | | LightGBM (no graph) | 0.1838 | | LightGBM + graph features | +0.4% better than without |

• On ultra-cold items (under 14 days old, n = 1,662) the graph lift rises to +2.1%.
• Graph features account for 0.3% of total LightGBM gain; rolling statistics and item age dominate.
• The p90 head covers 88.5% of outcomes against a 90% target — slightly under-covering, which matters for stock-outs.

• Forecasts are served with model metadata so a planner can see which model version produced a number.
• Promotion is gated: a retrained model that does not beat the validation bar is not promoted.
• Under-coverage of the p90 head is the number to watch in production — it translates directly into stock-outs.

• Production forecasting across 4,000+ SKUs × 33 stores, served by API and refreshed automatically.
• In the rebuild: ~19% lower WMAPE than seasonal-naive (0.2273 → 0.1838).
• Graph features: a small, honest effect — +0.4% overall, +2.1% for brand-new items.
• Earlier, at Noon: daily delivery-demand MAPE cut from 25% to 18% across e-commerce, grocery and quick-commerce.

• The boring features win. Rolling means, volatility and item age carry most of the signal.
• Report the small lift. +0.4% is a real result; dressing it up would make every other number on this page less credible.
• Graph context earns its place at the edges — new products with no history — not in the average.