Kontrol Teorisi Uygulamaları ile Zaman Serisi Öngörülerinin İyileştirilmesi

Abstract

The success of business management and functions largely depends on the correct functioning of the planning process in demand management and therefore the reduction of uncertainties in demand planning. In this thesis work; made propositions state that demand planning is a process just like, but an abstract one, physical industrial processes; that inputs and outputs of demand planning can be defined and measured as demand forecasts and actual sales respectively, and that accordingly, control systems used in physical industrial processes can also be used to reduce errors in demand planning. Starting from this point of view, the propositions are generalized, approach for improvement of time series forecasts via control theory applications have been asserted. Along with this approach, it has been hypothesized that time series forecast models can be abstracted and defined as black box just via the input and output relationship, and thus control systems can be integrated with forecast models. In order to verify the asserted approach; first, prediction methods in general, time series and statistical time series analysis have been mentioned and ARIMA models has been emphasized. Then, control theory, control loops and modelling of control systems have been referred and PID control has been discussed in detail. Later on, the relationship between control theory and time series forecasts were explained and PID controlled forecast model which is designed in line with the proposed approach were presented in detail. Finally, demand forecasting applications on demand and sales datasets consisting of nonstationary time series of different types were implemented via the use of PID controlled forecast model approach in conjunction with ARIMA models. Experimental findings obtained from the applications have been shown that PID controlled forecast models substantially reduced error rates on all applications, when PID controlled forecast models have been compared with lean forecast models in terms of forecasting errors. At the same time, it was resulted that lean forecast models with lower forecasting success were improved to a greater extent with control systems. By the experimental findings, all the propositions which were made about demand planning were supported and the black box hypothesis which was stated along with the proposed approach for forecasting models was confirmed. Thus, the inference have been made that time series forecasts can be improved via control theory applications, independently of the type of time series and the internal structure of forecast models. As the conclusion of this thesis; the approach for improvement of time series forecasts via control theory applications has been gained to the literature of forecasting as a novel methodological framework, and therewithal a brand new field of application for control theory and control systems has been introduced.