Abstract: This study presents a feature-based approach for predicting and optimizing stock levels in a retail chain. Utilizing an extensive database of algorithms, interpretable features are extracted from historical stock level time series data. The method, which includes correlation structure, distribution, entropy, stationarity, and scaling properties, reduces dimensionality significantly. Through a forward feature selection process and a linear classifier, the approach autonomously learns distinctions between stock level classes, outperforming conventional classifiers. The selected features not only enhance classification accuracy but also provide crucial insights for optimizing inventory management in diverse retail locations. This research contributes a scalable and interpretable solution to the dynamic challenges of stock prediction in a retail setting.

Keywords: Time Series Analysis, Forecasting, Machine Learning, Big mart sales.

| DOI: 10.17148/IARJSET.2024.11643