Empirical Study of Neural Network and other Machine Learning Models in Time Series Forecasting using Sales Data

Sammanfattning

Time series forecasting is a critical component of decision-making in various domains, from finance to retail. With the increasing complexity of real-world datasets, selecting the most appropriate modeling techniques has become crucial for accurate predictions. This thesis evaluates the performance of a traditional gradient boosting ensemble modeling, which constructs a stronger model by sequentially adding weaker models in a gradient descent manner to enhance accuracy and robustness. It also assesses a Neural network modeling that implements a mix of backward and forward residual links and a deep stack of fully-connected layers in a hierarchical structure. These models are evaluated using three distinct datasets, each differing in their structure, features, and origin. Through comprehensive analyses, including hyperparameter optimization, cross-validation, and evaluation of model performance metrics, insights into the strengths and limitations of these approaches are provided. Findings reveal that traditional gradient boosting ensemble methods like LightGBM demonstrate commendable speed and competitive performance, particularly in datasets with regular patterns. Conversely, Neural Network models like N-BEATS excel in capturing complex temporal dependencies but may face challenges due to longer training times and computational demands. While N-BEATS generally exhibit higher accuracy, LightGBM showcases superior efficiency and speed, which is valuable in scenarios with limited computational resources or real-time forecasting requirements. The study also delves into the effectiveness of different cross-validation strategies, shedding light on their impact on model performance assessment. The comparison between traditional and deep learning models highlights the nuanced trade-offs between accuracy, computational resources, and model complexity, emphasizing the importance of selecting the most suitable forecasting model based on specific dataset characteristics and forecasting requirements. In addition, this thesis investigates the effectiveness of a public python library Darts, that is effective in handling panel and multivariate hierarchical datasets, expanding the evaluation to cover different time series forecasting challenges. In conclusion, this thesis contributes to improved understanding of the performance of traditional and deep learning models in time series forecasting, providing practical insights for decision-makers in various industries.