【统计学】【2010.01】时间序列预测的组合：何时以及为什么能使人受益？

在这里插入图片描述

本文为英国伯恩茅斯大学（作者：Christiane Lemke）的博士论文，共139页。

时间序列预测在许多应用领域有着悠久的历史。在预测研究中，有人指出，找到一种最适合所有可能情况的算法是没有希望的。因此，当前的研究工作并没有致力于设计一个单一的优越算法，而是转向更深入地理解一种预测方法在某些条件下可能表现良好而在其他条件下可能失败的原因。这篇论文对这个问题提供了一些贡献。

传统的经验评估是从一个新的角度进行讨论，质疑在没有领域知识的情况下使用复杂预测方法的好处。一个实证研究侧重于相关的离线预测组合方法强调了相对简单的方法在实际应用中的竞争力。此外，还提取了时间序列的元特征，利用元学习自动发现和利用特定应用数据特征与预测性能之间的联系。最后，讨论了通过多样化函数方法、参数集和用于学习的数据聚合级别来扩展输入预测集的方法，并将结果预测的特性与各个方法和组合的不同误差分解联系起来。为了充分利用预测生成过程的知识，研究了高级组合结构。预测是航空公司收入管理中的一个关键因素；预测预期的订票、取消量和无票数量将直接影响到航线和时刻表的总体规划、票价等级的运力控制和超订限制。本文与德国汉莎航空公司（Lufthansa）合作，在一个航空公司数据集上进行了实验，目的是通过修改算法来改进当前的预测方式。为了将本文所研究方法的结果与预测研究的现状进行比较，一些实验还使用了最近两次预测竞赛的数据集，从而能够提供学术研究和工业实践之间的联系。

Time series forecasting has a long trackrecord in many application areas. In forecasting research, it has beenillustrated that finding an individual algorithm that works best for allpossible scenarios is hopeless. Therefore, instead of striving to design asingle superior algorithm, current research efforts have shifted towards gaininga deeper understanding of the reasons a forecasting method may perform well insome conditions whilst it may fail in others. This thesis provides a number ofcontributions to this matter. Traditional empirical evaluations are discussedfrom a novel point of view, questioning the benefit of using sophisticatedforecasting methods without domain knowledge. An own empirical study focusingon relevant off-theshelf forecasting and forecast combination methodsunderlines the competitiveness of relatively simple methods in practicalapplications. Furthermore, meta-features of time series are extracted toautomatically find and exploit a link between application specific datacharacteristics and forecasting performance using meta-learning. Finally, theapproach of extending the set of input forecasts by diversifying functionalapproaches, parameter sets and data aggregation level used for learning isdiscussed, relating characteristics of the resulting forecasts to differenterror decompositions for both individual methods and combinations. Advancedcombination structures are investigated in order to take advantage of theknowledge on the forecast generation processes. Forecasting is a crucial factorin airline revenue management; forecasting of the anticipated booking,cancellation and no-show numbers has a direct impact on general planning ofroutes and schedules, capacity control for fareclasses and overbooking limits.In a collaboration with Lufthansa Systems in Berlin, experiments in the thesisare conducted on an airline data set with the objective of improving thecurrent net booking forecast by modifying one of its components, thecancellation forecast. To also compare results achieved of the methodsinvestigated 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 industrialpractice.