In the age of big data, performing fast and accurate time series forecasting is a challenging task for business. The Theta method has fascinated researchers and practitioners because of its simplicity and extraordinary performance in the M3 competition. Despite the academic insights that have been contributed to the Theta literature, the full potential of Theta has not been reached yet. On the one hand, although the Theta has been shown to forecast well for different types of data, most relevant studies have adopted the traditional ways of handling seasonality and trends in Theta and left largely unexplored the extensions of these usages. On the other hand, controlled by θ parameter, the magnitude of the short-term and the long-term patterns differs for different Theta lines, which shares similarities to the concept of multiple Temporal Aggregation (TA). TA generates multiple aggregated series encoded with different patterns and collectively uses information obtained from those surrogate series to benefit forecasting. To the best of our knowledge, no one has systematically explored the integrated use of Theta and TA. To close this gap and to further exploit Theta’s potential, expand its framework, and improve its forecasting performance, we first proposed a hybrid Theta method in which several extensions that focus on the seasonality test, decomposition, and extrapolation of the trend curve were explored. Then, the proposed Theta method was integrated with TA and the value of their interaction was investigated. This proposed hybrid Theta method and its integration with TA were next evaluated in an extensive empirical study. The results (a) demonstrate by using a diverse set of real-life series the superior performance of the hybrid Theta method in comparison to all examined variants of the Theta method and available benchmarks, (b) show that the interaction of Theta and TA is not always beneficial, especially for trended data, thus requiring application of specific remedial strategies, and (c) suggest that TA is more effective for improving quantile forecasts than for point forecasts derived by the Theta method.
|Date of Award||14 Sep 2022|
|Supervisor||Fotios Petropoulos (Supervisor), Jooyoung Jeon (Supervisor) & Gunes Erdogan (Supervisor)|