Grupo Bimbo Demand Forecasting
Weekly demand forecasts at the client × product × route level, trained on 74M+ historical transactions, with prediction intervals and a serving API.
Repo: github.com/aksh-ay06/forecasting · Stack: pandas, LightGBM, MLflow, SHAP, FastAPI, Docker
Business Context
Grupo Bimbo delivers to over a million stores each week across 45,000+ routes. Over-forecasting creates unsold returns (waste), and under-forecasting causes stockouts. The model covers roughly 2,600 products across ~900K clients.
Approach
- Memory engineering: downcasting dtypes shrank the 74M rows from ~6 GB to ~2.4 GB in memory. The data is stored as Parquet partitioned by week for fast access by time.
- No leakage: features for week S use only data from weeks before S, so the model never sees information from the period it predicts.
- ~44 features: lag features (1–5 weeks), rolling mean/std/median, aggregations across 7 entity groups, attributes parsed from product names (brand, weight, pieces, promo), and a hierarchical backoff.
- Cold start: client-product pairs never seen before fall back through a hierarchy (for example, product → route → agency averages).
- Uncertainty: separate quantile-regression models produce 10th and 90th percentile bounds alongside each point forecast.
- Production path: experiments are tracked in MLflow, results are explained with SHAP, the FastAPI service supports single and batch prediction, and a drift-monitoring module is included.
flowchart LR RAW[Raw CSVs] --> PQ[Parquet by week] PQ --> FE[Features: lags, aggs, backoff] FE --> LGB[LightGBM + quantile models] LGB --> MLF[(MLflow)] MLF --> API[FastAPI]