Abstract

Having an accurate workload forecast is important for workload capacity planning, since the goal of capacity planning is to optimally allocate resources to current and future demand requirements. Capacity planning alongside workload forecasting determines employee headcount, backlog levels, and scheduling requirements. There are a growing number of studies in recent years that show that machine learning (ML) outperforms traditional statistical benchmarks. The thesis evaluates whether Light Gradient Boosting Machines (LightGBM), K-Nearest Neighbors (KNN), Support Vector Machines (SVM), and feedforward neural networks (NN) consistently outperform simple statistical benchmarks for short-term and long-term workload forecasting. Based on the findings of the experiments, one recommendation to managers is that workload forecasting should be based on simple statical methods rather than ML models. For both the budget (long-term) and schedule (short-term) forecasting task, simple statistical methods outperformed KNN, LightGBM, SVM, and NN. The high level of noise that is present in workload time-series makes it unlikely that ML models will outperform simple statistical benchmarks. A second recommendation to managers is to incorporate hierarchical reconciliation using minimum trace ordinary least squares to improve forecasting accuracy while making the forecasts coherent.
Key Words: workload, hierarchical reconciliation, budget forecasting, schedule forecasting, LightGBM, Neural Networks, K-Nearest Neighbors, Support Vector Machines, workload capacity planning, forecasting benchmarks