Journals / Black Sea Journal of Engineering and Science / 2026 / Cilt: 9 Sayı: 5

The Performance of RNN Architectures in ATM Cash Demand Forecasting: LSTM, GRU, Bi-LSTM, and CNN-LSTM

ATM Nakit Talep Tahmininde RNN Mimarilerinin Başarımı: LSTM, GRU, Bi-LSTM ve CNN-LSTM

Havanur Şenocak(ATATÜRK ÜNİVERSİTESİ)0009-0003-6841-6520
Tuba Irmak(ATATÜRK ÜNİVERSİTESİ)0000-0003-4749-5226
Pages
2156–2174
Publication Date
2026-09-15
🏛️Bibliographic metadata and full text access for this article are provided via DergiPark.
DergiPark Record ↗
✨

AI Summary & Key Takeaways

AI Insight
⚡ TL;DR(Too Long; Didn't Read)

In conclusion, the study demonstrated that RNN-based models can be effectively used for cash demand forecasting.

📌 3 Key Takeaways

🎯
Objective & Hypothesis:This study comparatively examines the prediction performance of four recurrent neural network (RNN)-based deep learning architectures (LSTM, GRU, Bi-LSTM, and CNN-LSTM) using a dataset containing daily cash withdrawal data from five different ATMs that is openly accessible on Kaggle.
📊
Key Findings & Data:The analysis revealed significant scale differences among ATMs, weekly periodicity, strong autocorrelation structures, and concurrent structural breaks after 2015.
💡
Conclusion & Impact:In conclusion, the study demonstrated that RNN-based models can be effectively used for cash demand forecasting.
🔬Design:Modeling & Simulation
📍Scope:2015 Verileri
⚙️Variables:Nakit yönetimi · RNN mimarileri · LSTM

Abstract

This study comparatively examines the prediction performance of four recurrent neural network (RNN)-based deep learning architectures (LSTM, GRU, Bi-LSTM, and CNN-LSTM) using a dataset containing daily cash withdrawal data from five different ATMs that is openly accessible on Kaggle. Exploratory data analysis was conducted to understand the structural characteristics of the data. The analysis revealed significant scale differences among ATMs, weekly periodicity, strong autocorrelation structures, and concurrent structural breaks after 2015. Based on these findings, the analysis period was limited, and new lagged, moving statistical, and calendar-based features were generated. During the modelling phase, each architecture was trained using two different input structures. The first was a univariate structure that used only historical withdrawal amounts, while the second was a multivariate structure consisting of 27 features derived from lagged variables, moving statistics, and calendar/holiday information. Additionally, a total of 120 experiments were conducted during the training process using a forward validation approach under 7-, 14-, and 30-day back sight windows. The findings showed that the multivariate input structure produced lower SMAPE values than the univariate structure across all four architectures, and the best results were consistently achieved with the multivariate structure. At the model level, the lowest average SMAPE value was obtained with the multivariate GRU architecture (36.63%), while the best ATM-level prediction performance was achieved by the Bi-LSTM architecture for the Big Street ATM (19.12%). The effect of back sight window length on model performance remained limited; however, a 14-day window yielded optimal results for most architecture. In conclusion, the study demonstrated that RNN-based models can be effectively used for cash demand forecasting. Furthermore, differences in demand behavior among ATMs were found to have a decisive impact on model performance, and it was not easy for a single model to achieve a similar level of success across all ATMs.

Özet

Bu çalışmada, Kaggle’da açık erişimli olarak paylaşılan, beş farklı ATM’ye ait günlük nakit çekim verilerini içeren veri seti kullanılarak tekrarlayan sinir ağı (Recurrent Neural Network – RNN) tabanlı dört derin öğrenme mimarisinin (LSTM, GRU, Bi-LSTM ve CNN-LSTM) tahmin başarımı karşılaştırmalı olarak incelenmiştir. Verinin yapısal özelliklerinin anlaşılabilmesi için keşifsel veri analizleri yapılmış ve analiz sonucunda ATM'ler arasında belirgin ölçek farklılıkları, haftalık periyodiklik, güçlü otokorelasyon yapıları ve 2015 yılı sonrasında eşzamanlı yapısal bozulmalar tespit edilmiş; bu bulgular doğrultusunda analiz dönemi sınırlandırılarak gecikmeli, hareketli istatistiksel ve takvim temelli yeni öznitelikler üretilmiştir. Modelleme aşamasında her bir mimari iki farklı girdi ile eğitilmiştir, bunlar; yalnızca geçmiş çekim tutarlarını kullanan tek değişkenli ve gecikmeli, hareketli istatistik ile takvim/tatil bilgilerinden türetilen 27 öznitelikten oluşan çok değişkenli girdi yapılarıdır. Ayrıca, 7, 14 ve 30 günlük geri bakış pencereleri altında ileri yönlü doğrulama yaklaşımıyla yapılan model eğitimi sürecinde toplam 120 deney gerçekleştirilmiştir. Bulgular, çok değişkenli girdi yapısının dört mimaride de tek değişkenli yapıya kıyasla daha düşük SMAPE değerleri ürettiğini ve en iyi sonuçların tamamının çok değişkenli yapıda gerçekleştiğini göstermiştir. Model bazında en düşük ortalama SMAPE değerine çok değişkenli GRU mimarisi (%36,63) ulaşırken, ATM bazında en iyi tahmin Bi-LSTM mimarisi ile Big Street ATM'sinde elde edilmiştir (%19,12). Geri bakış penceresinin uzunluğunun model başarımı üzerindeki etkisi sınırlı düzeyde kalmış; çoğu mimaride 14 günlük pencere optimal sonuç vermiştir. Çalışma sonunda, RNN tabanlı modellerin nakit talep tahmininde etkili bir şekilde kullanılabileceği görülmüştür. Ayrıca ATM'ler arasında gözlenen talep davranışı farklılıklarının model performansları üzerinde belirleyici olduğu ve tek bir modelin tüm ATM'lerde benzer başarı düzeyine ulaşmasının kolay olmadığı görülmüştür.

🔖

Cite This Article

Şenocak, H., Irmak, T., Şirin, B. T. (2026). The Performance of RNN Architectures in ATM Cash Demand Forecasting: LSTM, GRU, Bi-LSTM, and CNN-LSTM. Black Sea Journal of Engineering and Science, Cilt: 9 Sayı: 5, 2156–2174. https://doi.org/10.34248/bsengineering.1966267

📄 Full Text PDF Viewer

📄

PDF görüntüleyici hazırlanıyor...

References (29)

  1. Andrawis, R. R., Atiya, A. F., & El-Shishiny, H. (2011). Forecast combinations of computational intelligence and linear models for the NN5 time series forecasting competition. International Journal of Forecasting, 27(3), 672–688. https://doi.org/10.1016/j.ijforecast.2010.09.005
  2. Berus, Y., & Yakut, Y. B. (2024). Derin öğrenme (1D-CNN, RNN, LSTM, BiLSTM) ile enerji tüketim tahmini: Diyarbakır AVM örneği. Dicle Üniversitesi Mühendislik Fakültesi Mühendislik Dergisi, 15(2), 311–322. https://doi.org/10.24012/dumf.1415055
  3. Bhandari, R., & Gill, J. (2016). An artificial intelligence ATM forecasting system for hybrid neural networks. International Journal of Computer Applications, 133(3), 13–16. https://doi.org/10.5120/ijca2016907863
  4. Cedolin, M., Orhan, D., & Genevois, M. (2024). Statistical and artificial intelligence based forecasting approaches for cash demand problem of automated teller machines. Academic Platform Journal of Engineering and Smart Systems, 12(1), 21–27. https://doi.org/10.21541/apjess.1360151
  5. Chakladar, D. D., Dey, S., Roy, P. P., & Dogra, D. P. (2020). EEG-based mental workload estimation using deep BLSTM-LSTM network and evolutionary algorithm. Biomedical Signal Processing and Control, 60, Article 101989. https://doi.org/10.1016/j.bspc.2020.101989
  6. Chen, Q., & Huang, G. (2021). A novel dual attention-based BLSTM with hybrid features in speech emotion recognition. Engineering Applications of Artificial Intelligence, 102, Article 104277. https://doi.org/10.1016/j.engappai.2021.104277
  7. Cho, K., van Merriënboer, B., Gulçehre, Ç., Bahdanau, D., Bougares, F., Schwenk, H., & Bengio, Y. (2014). Learning phrase representations using RNN encoder–decoder for statistical machine translation. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP) (pp. 1724–1734). Association for Computational Linguistics. https://doi.org/10.3115/v1/D14-1179
  8. Fallahtafti, A., Aghaaminiha, M., Akbarghanadian, S., & Weckman, G. R. (2022). Forecasting ATM cash demand before and during the COVID-19 pandemic using an extensive evaluation of statistical and machine learning models. SN Computer Science, 3(2), Article 164. https://doi.org/10.1007/s42979-021-01000-0
  9. Fareed, S. (2020). Data set of ATM transaction of XYZ bank [Data set]. Kaggle. Retrieved September 10, 2025, from https://www.kaggle.com/datasets/saadfareed/data-set-of-atm-transaction-of-xyz-bank
  10. Gençalp, E. (2020). ATM demand prediction with machine learning methods (Master's thesis, Galatasaray University). National Thesis Center of the Council of Higher Education (YÖK).
  11. Gu, J., Wang, Z., Kuen, J., Ma, L., Shahroudy, A., Shuai, B., & Chen, T. (2018). Recent advances in convolutional neural networks. Pattern Recognition, 77, 354–377. https://doi.org/10.1016/j.patcog.2017.10.013
  12. Jadwal, P. K., Jain, S., Gupta, U., & Khanna, P. (2017). K-means clustering with neural networks for ATM cash repository prediction. In S. Satapathy, Y. V. Joshi, & N. Modi (Eds.), Information and Communication Technology for Intelligent Systems (ICTIS 2017) - Volume 2 (pp. 588–596). Springer. https://doi.org/10.1007/978-3-319-63645-0_70
  13. Javanmard, M. (2016). An approach to improve forecasting cash demand at ATMs using fuzzy logic. International Journal of Computer Science and Information Security, 14(11), 404–411.
  14. Hochreiter, S., & Schmidhuber, J. (1996). LSTM can solve hard long time lag problems. Advances in neural information processing systems, 9.
  15. Kamini, V., Ravi, V., & Kumar, D. N. (2014). Chaotic time series analysis with neural networks to forecast cash demand in ATMs. In 2014 IEEE International Conference on Computational Intelligence and Computing Research (pp. 1–5). IEEE. https://doi.org/10.1109/ICCIC.2014.7238495
  16. Kaysal, K., Yurttakal, A. H., & Hocaoğlu, F. O. (2023). Hibrit derin öğrenme yöntemi kullanılarak hiperparametre optimizasyonu ile yenilenebilir elektrik enerjisi tahmini. Niğde Ömer Halisdemir Üniversitesi Mühendislik Bilimleri Dergisi, 12(3), 770–777. https://doi.org/10.28948/ngumuh.1263782
  17. Kılınç, H. Ç., & Öztürk, Y. (2022). Hibrit gri kurt optimizasyonu ile kapılı tekrarlayan birim modeli kullanılarak günlük akım tahmini. Avrupa Bilim ve Teknoloji Dergisi, (35), 259–267. https://doi.org/10.31590/ejosat.1062777
  18. LeCun, Y. A., & Bengio, Y. O. (1998). Convolutional networks for images, speech, and time series. In M. A. Arbib (Ed.), The Handbook of Brain Theory and Neural Networks (pp. 255–258). MIT Press.
  19. Nemeshaev, S., & Tsyganov, A. (2016). Model of the forecasting cash withdrawals in the ATM network. Procedia Computer Science, 88, 463–468. https://doi.org/10.1016/j.procs.2016.07.465
  20. Poorzaker Arabani, S., & Ebrahimpour Komleh, H. (2019). The improvement of forecasting ATM cash demand of Iran banking network using convolutional neural network. Arabian Journal for Science and Engineering, 44(4), 3733–3743. https://doi.org/10.1007/s13369-018-3647-7
  21. Rafi, M., Wahab, M. T., Khan, M. B., & Raza, H. (2020). ATM cash prediction using time series approach. In 2020 3rd International Conference on Computing, Mathematics and Engineering Technologies (iCoMET) (pp. 1–6). IEEE. https://doi.org/10.1109/iCoMET48628.2020.9103780
  22. Rostamian, A., & O’Hara, J. G. (2022). Event prediction within directional change framework using a CNN-LSTM model. Neural Computing and Applications, 34(20), 17193–17205. https://doi.org/10.1007/s00521-022-07687-3
  23. Şanlıalp, İ., & Yıldız, M. (2025). Sistem çağrısı verilerinde derin öğrenme mimarileri kullanılarak anomali tespitinin değerlendirilmesi: Performans ve enerji verimliliği. Süleyman Demirel Üniversitesi Fen Bilimleri Enstitüsü Dergisi, 29(2), 388–396. https://doi.org/10.19113/sdufenbed.1655066
  24. Sarveswararao, V., Ravi, V., & Vivek, Y. (2023). ATM cash demand forecasting in an Indian bank with chaos and hybrid deep learning networks. Expert Systems with Applications, 211, Article 118645. https://doi.org/10.1016/j.eswa.2022.118645
  25. Teddy, S. D., & Ng, S. K. (2011). Forecasting ATM cash demands using a local learning model of cerebellar associative memory network. International Journal of Forecasting, 27(3), 760–776. https://doi.org/10.1016/j.ijforecast.2010.02.013
  26. Tosun, M. F., & Şentürk, A. (2023). Kasislerin yakıt tüketimine etkisinin RNN, LSTM, GRU tekrarlayan derin öğrenme algoritmaları ile tespiti. Journal of Intelligent Systems: Theory and Applications, 6(1), 12–23. https://doi.org/10.38016/jista.1141359
  27. Tsujimoto, T., Takahashi, Y., Takeuchi, S., & Maeda, Y. (2016). RNN with Russell's circumplex model for emotion estimation and emotional gesture generation. In 2016 IEEE Congress on Evolutionary Computation (CEC) (pp. 1427–1431). IEEE. https://doi.org/10.1109/CEC.2016.7744023
  28. Venkatesh, K., Ravi, V., Prinzie, A., & Van den Poel, D. (2014). Cash demand forecasting in ATMs by clustering and neural networks. European Journal of Operational Research, 232(2), 383–392. https://doi.org/10.1016/j.ejor.2013.07.027
  29. Zha, W., Liu, Y., Wan, Y., Luo, R., Li, D., Yang, S., & Xu, Y. (2022). Forecasting monthly gas field production based on the CNN-LSTM model. Energy, 260, Article 124889. https://doi.org/10.1016/j.energy.2022.124889