Demand Forecast by ML model of an MNC FMCG company, namely Grupo Bimbo.

Business Questions & Interpretation

Q1. What is purpose of this project or how this project could help the business of Grupo Bimbo?

The Grupo Bimbo Inventory Demand project aims to develop a machine learning model that predicts the weekly demand for bakery products across Grupo Bimbo’s extensive distribution network. Using historical sales, returns, product, customer, and route data, the objective is to accurately forecast future demand so the company can optimize inventory management, reduce product waste caused by overstocking, prevent stock shortages, and improve overall supply chain efficiency.

---

Q2. Which business variables have the largest impact on forecasted demand?

Based on the LightGBM Feature Importance chart, the business variables with the largest impact on forecasted demand are dominated by product and geographic identifiers, with Producto_ID (Product ID) being the single most influential driver by a significant margin, closely followed by product weight as the second most impactful factor. Geographic and logistical variables also play a major role in the model's predictions, as Town, Ruta_SAK (Route), and Agencia_ID (Agency ID) round out the top five most critical features influencing the demand forecast.

---

Q3. How can managers use this forecast?

Managers can use the forecast to improve inventory planning, production scheduling, warehouse allocation, and distribution efficiency.

---

Q4. What actions should be taken when forecast demand is high?

High demand forecasts may require additional production, safety stock adjustments, and proactive logistics planning to avoid stockouts.

---

Q5. Why LightGBM model was choseen over other Machine Learning models?

Based on the RMSLE, R2, and MAE of two models namely XGBoost and LightGBM, I can conclude that LightGBM is slightly better than model for RMSLE, MAE and R2. RMSLE (lower is better) — penalizes percentage-like forecasting errors. MAE (lower is better) — average prediction error in units. R² (higher is better) — explains how much variance the model captures. Thus, LightGBM is selected.

---

Learning outcomes from this project for me:

> Developed a demand forecasting model using LightGBM.
> Performed data preprocessing and feature engineering.
> Gained experience of big data analysis on large-scale FMCG sales data, nearly 75 millions data.
> Conducted hyperparameter optimization to improve model performance.
> Evaluated model performance using RMSLE, MAE, and R².
> Built an interactive web-based prediction dashboard.
> Deployed a machine learning model for real-time forecasting.
> Used Docker for containerized deployment of the application.
> Applied demand forecasting to inventory and supply chain planning.
> Enhanced skills in machine learning, data analytics, and model deployment.
> Translated predictive insights into business decision-making support.
> Polished technical skills like numpy, pandas, matplotlib, scikit learn, shap, pickle, xgboost model, lightGBM model, Catboost model, RandomForest model, and so on.