Comparison of computational intelligence models on forecasting ATM demands

---

Tezin Türü: Yüksek Lisans

Tezin Yürütüldüğü Kurum: Marmara Üniversitesi, Türkiye

Tezin Onay Tarihi: 2019

Tezin Dili: İngilizce

Öğrenci: Onur Gürkan Gültekin

Danışman: Ali Fuat Alkaya

Açık Arşiv Koleksiyonu: AVESİS Açık Erişim Koleksiyonu

Özet:

Usage of ATMs (Automated Teller Machines) is crucial for people to reach cash money instantly when they need it. To satisfy this demand of the customers, banks have to determine when to visit and how much to load to each ATM of the bank. The primary question that needs to be answered turns out to be forecasting how much money will be withdrawn from the ATMs in the next days. In this study, we exploit computational intelligence techniques to solve this problem via using past transactions and their features. This study shows that our mission is a difficult task since our ATM data, taken from one of the largest banks in Turkey, have too volatile withdrawal patterns when compared to ATM data taken from UK banks. This is the reason why classical time series forecasting models do not perform well on our ATM data, although we know that they perform competitively on ATM data taken from the UK banks. According to this fact, we implemented different computational intelligence models with extensive search of their parameter space for finding the best performing models and present the detailed comparison with respect to their accuracy and time performance. The results show that SVR (Support Vector Regression) adapted from Support Vector Machines and ANN (Artificial Neural Networks), already known as universal approximators, outperform the others. On the other hand, SPSA (Simultaneous Perturbation Stochastic Approximation) which uses gradient approximations for updating its weights rather than direct gradient calculations shows promising performance to be a good sign for future improvements.