Gmdh-type neural network algorithms for short term forecasting


Tezin Türü: Yüksek Lisans

Tezin Yürütüldüğü Kurum: Orta Doğu Teknik Üniversitesi, Fen Edebiyat Fakültesi, İstatistik Bölümü, Türkiye

Tezin Onay Tarihi: 2015

Öğrenci: OSMAN DAĞ

Danışman: CEYLAN YOZGATLIGİL

Özet:

Group Method of Data Handling (GMDH) - type neural network algorithms are the heuristic self-organization method for modelling the complex systems. GMDH algorithms are utilized for the variety of purposes, which are identification of physical laws, extrapolation of physical fields, pattern recognition, clustering, approximation of multidimensional processes, forecasting without models and so on. In this study, GMDH - type neural network algorithms were applied to make forecasts for time series data sets. We mainly focused on development of free software. For this purpose, we developed an R package GMDH. Moreover, we integrated different transfer functions, sigmoid, radial basis, polynomial, and tangent functions, into GMDH algorithm. We proposed an algorithm in which all transfer functions are used simultaneously or separately if desired. Also, we used regularized least square estimation for the estimation of weights to overcome multi-collinearity problem. The methods were illustrated on real life datasets having different properties to see the prediction and forecasting performance of the algorithm. We included ARIMA models and exponential smoothing methods for the comparison purpose. GMDH algorithms show the same or even better performance than the other methods.