Retail Sales Forecasting - Time-Series and ML Benchmark

Overview

Built a weekly sales forecasting pipeline for a Superstore Technology-category dataset, comparing classical time-series methods, additive forecasting, tree-based machine learning, and sequence-based deep learning. The goal was to recommend the best model for retail planning under volatility, seasonality, and limited historical data.

Outcome

• Benchmarked ARIMA, SARIMA, Prophet, Random Forest, XGBoost, and LSTM under a consistent temporal train/test setup.
• Engineered lag, rolling, holiday, seasonal, mutual-information-selected, and multiplicative interaction features to capture demand spikes.
• Selected a seasonal, holiday-aware XGBoost model with nested cross-validation as the strongest balance of accuracy, interpretability, and deployment simplicity.