🧠 Promotion-Sensitive Hierarchical Probabilistic Forecasting for FMCG Demand 📦📊
Promotion-Sensitive-Hierarchical-Probabilistic-Forecasting-for-FMCG-Demand is an advanced machine learning framework designed to accurately forecast Fast-Moving Consumer Goods (FMCG) demand by capturing promotional effects, hierarchical sales dependencies, and uncertainty distributions across products, regions, and time scales.
This project combines hierarchical Bayesian modeling, probabilistic deep learning, and promotion-aware feature engineering to deliver robust, scalable, and interpretable demand forecasts that drive smarter supply chain decisions in the FMCG industry.
🧩 Abstract
Traditional demand forecasting models struggle to handle the complex temporal and hierarchical relationships found in FMCG data, especially during promotional periods that heavily influence consumer behavior.
This project proposes a promotion-sensitive hierarchical probabilistic model that captures:
The impact of marketing promotions and discounts at both product and category levels.
Temporal dynamics across multiple time scales (daily, weekly, monthly).
Hierarchical dependencies (SKU → Category → Brand → Region → Country).
Predictive uncertainty, enabling risk-aware business planning.
✨ Key Features
📊 Hierarchical Probabilistic Forecasting: Models nested relationships between SKUs, product categories, and markets.
🛍️ Promotion Sensitivity: Incorporates promotion calendars, discount depth, and advertising intensity as causal drivers.
🧠 Deep Learning Backbone: Uses LSTM / Temporal Fusion Transformer (TFT) / DeepAR for sequential modeling.
🔮 Bayesian Uncertainty Estimation: Provides prediction intervals instead of point forecasts for better risk management.
📈 Feature-Aware Forecasting: Integrates seasonality, price elasticity, and macroeconomic indicators.
🧰 Scalable to Enterprise FMCG Data: Designed for thousands of SKUs and regions simultaneously.
⚙️ Technical Highlights Module Description Data Preprocessing Cleans, aggregates, and encodes multi-level FMCG data. Feature Engineering Generates promotion indicators, lagged sales features, calendar events, and price elasticity variables. Hierarchical Modeling Applies top-down and bottom-up reconciliation of forecasts across hierarchy levels. Probabilistic Forecasting Models sales as distributions (Gaussian / Negative Binomial / Quantile regression). Deep Sequence Models Implements DeepAR, LSTM, TFT, or N-BEATS architectures. Evaluation Framework Measures accuracy, calibration, and promotion impact across time horizons. 🧮 Methodology Overview Input: Historical Sales Data + Promotion Events + Price Changes + Holidays ↓ Feature Engineering: Lag Features, Categorical Embeddings, Promo Flags, Seasonality Encoding ↓ Model: Hierarchical Probabilistic Deep Learning Model ↓ Output: Multi-level Forecasts (SKU, Brand, Category) + Prediction Intervals ↓ Evaluation: Accuracy, Sharpness, Calibration, Business Uplift Metrics
🧰 Tech Stack
Languages: Python 🐍
Core Frameworks: PyTorch / TensorFlow Probability / PyMC / Prophet
Forecasting Libraries: GluonTS, Kats, Nixtla, NeuralForecast
Visualization: Plotly, Matplotlib, Seaborn, Altair
Data Handling: Pandas, NumPy, Dask, Polars
📁 Project Structure 📁 data/ # Raw and processed FMCG datasets 📁 notebooks/ # Experiment and analysis notebooks 📁 features/ # Promotion, temporal, and categorical feature generators 📁 models/ # Probabilistic and hierarchical forecasting models 📁 evaluation/ # Performance and uncertainty metrics 📁 results/ # Forecast visualizations and evaluation reports 📁 utils/ # Helper scripts and data loaders
🚀 Getting Started git clone https://github.com/yourusername/Promotion-Sensitive-Hierarchical-Probabilistic-Forecasting-for-FMCG-Demand.git cd Promotion-Sensitive-Hierarchical-Probabilistic-Forecasting-for-FMCG-Demand pip install -r requirements.txt python train.py --model tft --hierarchy brand --promo_sensitivity True --forecast_horizon 90
📊 Evaluation Metrics Category Metric Description Accuracy RMSE / MAPE / sMAPE Overall forecast performance Calibration CRPS / Sharpness Reliability of probabilistic predictions Fairness Weighted Reconciliation Error Hierarchical consistency Promotion Impact Lift vs. Baseline Sales uplift due to promotions Explainability SHAP / Attention Weights Feature contribution analysis 💡 Business Impact
🏪 Demand Forecasting Accuracy: Reduces forecast error during promo periods by 20–30%.
💰 Inventory Optimization: Aligns procurement and logistics with expected uplift.
📦 Promotion Planning: Quantifies promotion elasticity to guide marketing spend.
🌍 Hierarchical Insights: Enables decision-making from SKU to national level.
🧠 Research Contributions
Introduces a promotion-sensitive probabilistic hierarchy for FMCG forecasting.
Combines Bayesian learning and deep sequence models for uncertainty quantification.
Provides an open-source reproducible pipeline for industry-scale forecasting research.
🤝 Contributing
We welcome collaboration from:
Data scientists working in retail analytics or CPG forecasting
Researchers in probabilistic time series and causal inference
Practitioners in supply chain optimization and pricing strategy
🏆 Citation
Hazrat Ali, Promotion-Sensitive Hierarchical Probabilistic Forecasting for FMCG Demand, 2025.