Evaluation of Machine Learning and Deep Learning Models in Forecasting Time Series of Variable Lengths

Forecasting a heterogeneous set of time series in the domain of perishable products presents a significant challenge due to volatility, data sparsity, and variable availability of historical data. This thesis systematically evaluates the forecast accuracy of machine learning and deep learning models in a real-world application using sales data from an Austrian food tech company. To address the limited number of original time series and ensure a robust comparison of models across varying amounts of historical data, a Fixed-Length Time Series Splitting framework is introduced, which generates multiple standardized windows from each series. Using this framework, the performance of gradient-boosted decision trees, neural networks, and pretrained transformer models is compared on a seven-day-ahead forecasting task under both local and global modeling paradigms. The impact of series length, engineered features, various statistical characteristics, and hyperparameter optimization on forecast accuracy is examined, using mean absolute scaled error as the primary metric. This thesis provides empirical evidence that can guide model selection for the task of forecasting time series with variable lengths. The results show that a global neural network model provides the most accurate and robust forecasts overall, while a local gradient-boosted decision trees model proved to be the most effective on the shortest series with a few weeks of history, addressing the cold start problem.