| Summary: | Time series forecasting has a long track record in many application areas. In forecasting research, it has been illustrated that finding an individual algorithm that works best for all possible scenarios is hopeless. Therefore, instead of striving to design a single superior algorithm, current research efforts have shifted towards gaining a deeper understanding of the reasons a forecasting method may perform well in some conditions whilst it may fail in others. This thesis provides a number of contributions to this matter. Traditional empirical evaluations are discussed from a novel point of view, questioning the benefit of using sophisticated forecasting methods without domain knowledge. An own empirical study focusing on relevant off-the shelf forecasting and forecast combination methods underlines the competitiveness of relatively simple methods in practical applications. Furthermore, meta-features of time series are extracted to automatically find and exploit a link between application specific data characteristics and forecasting performance using meta-learning. Finally, the approach of extending the set of input forecasts by diversifying functional approaches, parameter sets and data aggregation level used for learning is discussed, relating characteristics of the resulting forecasts to different error decompositions for both individual methods and combinations. Advanced combination structures are investigated in order to take advantage of the knowledge on the forecast generation processes. Forecasting is a crucial factor in airline revenue management; forecasting of the anticipated booking, cancellation and no-show numbers has a direct impact on general planning of routes and schedules, capacity control for fareclasses and overbooking limits. In a collaboration with Lufthansa Systems in Berlin, experiments in the thesis are conducted on an airline data set with the objective of improving the current net booking forecast by modifying one of its components, the cancellation forecast. To also compare results achieved of the methods investigated here with the current state-of-the-art in forecasting research, some experiments also use data sets of two recent forecasting competitions, thus being able to provide a link between academic research and industrial practice.
|
|---|
