Sparse Time Series Demand Forecasting for Intermittent Availability

Abstract

This thesis addresses the challenge of forecasting sales for individual perishable markdown products using historical sales data and other relevant features in the form of time series. The target time series, i.e., daily sales of marked-down units of a certain product, is intermittent, sparse, and highly irregular, and sales can only occur if the marked-down product is available. To solve the problem, various methods were evaluated, ranging from well-established statistical models to newer deep learning-based models. This thesis proposes an interpretable novel method that improves the Temporal Fusion Transformer model with cluster encodings by applying random convolutional kernel transformations to time series. The study found that the compared deep learning models outperformed the baseline statistical models, particularly the RNN and Temporal Fusion Transformer. The novel approach of clustering the markdown series based on markdown features showed no significant change in performance regarding day-to-day prediction. However, it did show a significant improvement in multi-horizon aggregated predictions. Moreover, using clustering resulted in decreased time in training the models. Overall, the results suggest that deep learning models and the Temporal Fusion Transformer with added cluster encodings are promising models for predicting intermittent series with known available inventory. This study has practical implications for retailers and businesses that sell perishable products. Accurately forecasting sales of markdown products can help reduce waste and optimize inventory management, resulting in cost savings and increased profitability.