Optimizing Cash Management in ATM Networks using Time Series Models and Spatial Analysis
Date Issued
2024
Author(s)
Aliyev, Musadiq
Abstract
Effective cash management across ATM networks is essential for banks to maintain optimal
service levels. However, in ABB, the largest state bank of Azerbaijan, cash management is
conducted manually. The main objective of this thesis is to develop an optimal strategy for
ABB’s ATM cash demand forecast by integrating time series models and spatial analysis. The
significance of this study is that, with the help of a robust strategy, the bank will reduce
operational costs, improve service levels, increase customer satisfaction, and enhance
profitability. Cash demand forecasting experiments were conducted on ABB’s ATMs to
achieve those goals. Various methods were used to build eight different models. Different
time series, machine learning models, and Meta’s Prophet model were used. The
best-performing model was SARIMAX, with SMAPE score of 5.89% and RMSE score of
83537. Next, spatial analysis was performed by grouping ATMs into five distinct sub-regions
based on their distance from the city center. This analysis was done for Baku city to consider
the various spatial factors influencing cash demand in different locations. Then, one ATM
from each region was selected, and the cash demand for the last nine months was forecasted
using SARIMAX. The MAPE score spans between 4.7% and 6.6% which demonstrates that
the model is reliable even with different locations, populations, and atm distributions. The
forecasted MAPE scores will be utilized as model calibration factors to adjust the fluctuations
of the model's predictions in production. The integration of forecast results with spatial
analysis to formulate strategies represents the novelty presented in this thesis.
service levels. However, in ABB, the largest state bank of Azerbaijan, cash management is
conducted manually. The main objective of this thesis is to develop an optimal strategy for
ABB’s ATM cash demand forecast by integrating time series models and spatial analysis. The
significance of this study is that, with the help of a robust strategy, the bank will reduce
operational costs, improve service levels, increase customer satisfaction, and enhance
profitability. Cash demand forecasting experiments were conducted on ABB’s ATMs to
achieve those goals. Various methods were used to build eight different models. Different
time series, machine learning models, and Meta’s Prophet model were used. The
best-performing model was SARIMAX, with SMAPE score of 5.89% and RMSE score of
83537. Next, spatial analysis was performed by grouping ATMs into five distinct sub-regions
based on their distance from the city center. This analysis was done for Baku city to consider
the various spatial factors influencing cash demand in different locations. Then, one ATM
from each region was selected, and the cash demand for the last nine months was forecasted
using SARIMAX. The MAPE score spans between 4.7% and 6.6% which demonstrates that
the model is reliable even with different locations, populations, and atm distributions. The
forecasted MAPE scores will be utilized as model calibration factors to adjust the fluctuations
of the model's predictions in production. The integration of forecast results with spatial
analysis to formulate strategies represents the novelty presented in this thesis.