Retail Demand Forecasting System

Overview

This project builds an end-to-end retail demand forecasting system to predict weekly units sold for products across different locations and platforms.

The system includes:

• Structured feature engineering for time-series modeling
• Model comparison and optimization
• Leakage-aware evaluation
• Interactive dashboard deployment using Streamlit

The objective is to generate accurate short-term demand forecasts to support inventory and operational planning.

---

Problem Statement

Retail demand is influenced by historical trends, seasonality, pricing, and contextual factors.

Accurate demand forecasting helps businesses:

• Optimize inventory levels
• Reduce stockouts and overstock
• Improve procurement decisions
• Support pricing strategy

This project focuses on weekly demand prediction using supervised machine learning techniques.

---

Dataset

The dataset contains weekly sales records with the following fields:

• Product Name
• Location
• Platform
• Category
• Date
• Units Sold
• Price
• Discount
• Units Returned

The data spans approximately 13 months of historical sales.

---

Feature Engineering

Retail demand is strongly autoregressive. To capture temporal structure and contextual effects, the following features were engineered:

Lag Features

• lag_1
• lag_2
• lag_3
• lag_4
• lag_6
• lag_8

These capture short-term and mid-term demand memory.

Rolling Statistics

• rolling_mean_4
• rolling_mean_8
• rolling_std_4
• rolling_std_8

These capture trend stability and momentum.

Seasonality Encoding

Month was encoded using sine and cosine transformation to represent cyclical patterns:

• month_sin
• month_cos

Pricing Dynamics

• price_lag_1
• price_change

These capture price elasticity effects.

External Context

• Holiday indicator (location-aware)
• Temperature proxy (season-based)
• Fuel index trend proxy

---

Model Development

Two models were evaluated:

1. Random Forest Regressor
2. XGBoost Regressor

A strict time-based train-test split was used to prevent future data leakage.

Revenue was initially included as a feature but removed after identifying target leakage, ensuring realistic performance evaluation.

---

Evaluation Metrics

Model performance was evaluated using:

• MAE (Mean Absolute Error)
• RMSE (Root Mean Squared Error)
• MAPE (Mean Absolute Percentage Error)

Final Model Performance (XGBoost with enhanced features)

• MAE: 7.56
• RMSE: 9.63
• MAPE: ~5.0%

Baseline Random Forest

• MAE: ~9.55
• MAPE: ~6.4%

Feature engineering reduced MAE by approximately 20 percent compared to the baseline model.

---

Key Observations

• Rolling averages were the most influential predictors.
• Demand behavior is trend-driven and strongly autoregressive.
• Seasonality contributed meaningfully through cyclical encoding.
• External contextual features had moderate impact relative to temporal features.

---

Deployment

The system is deployed using Streamlit and includes:

• Sidebar-based forecast controls
• Context snapshot (date, week, holiday status)
• Historical demand visualization
• Compact analytical dashboard layout

To run locally:

streamlit run dashboard/app.py

---

Limitations

• Requires historical sales data (cold-start limitation for new products)
• External shock events are not modeled
• Temperature and fuel features are proxy-based
• Limited historical time span (13 months)

---

Future Improvements

• Multi-step forecasting (predicting multiple future weeks)
• Hierarchical modeling for new product fallback
• Automated retraining pipeline
• Integration with real weather and macroeconomic APIs
• Drift monitoring and performance tracking

---

Tech Stack

• Python
• Pandas
• NumPy
• Scikit-learn
• XGBoost
• Streamlit
• Joblib