ATM Cash demand forecasting in an Indian Bank with chaos and deep learning

TL;DR

This paper models chaos in Indian bank ATM withdrawal data, incorporates day-of-week as an exogenous variable, and compares various forecasting methods including deep learning models, demonstrating improved accuracy with chaos modeling and feature augmentation.

Contribution

It introduces a chaos-based approach combined with deep learning and exogenous features for improved ATM cash demand forecasting in Indian banks.

Findings

Chaos modeling improves forecast accuracy across models

Deep learning models perform comparably to RF in forecasting

Adding day-of-week features enhances prediction performance

Abstract

This paper proposes to model chaos in the ATM cash withdrawal time series of a big Indian bank and forecast the withdrawals using deep learning methods. It also considers the importance of day-of-the-week and includes it as a dummy exogenous variable. We first modelled the chaos present in the withdrawal time series by reconstructing the state space of each series using the lag, and embedding dimension found using an auto-correlation function and Cao's method. This process converts the uni-variate time series into multi variate time series. The "day-of-the-week" is converted into seven features with the help of one-hot encoding. Then these seven features are augmented to the multivariate time series. For forecasting the future cash withdrawals, using algorithms namely ARIMA, random forest (RF), support vector regressor (SVR), multi-layer perceptron (MLP), group method of data handling (GMDH), general regression neural network (GRNN), long short term memory neural network and 1-dimensional convolutional neural network. We considered a daily cash withdrawals data set from an Indian commercial bank. After modelling chaos and adding exogenous features to the data set, we observed improvements in the forecasting for all models. Even though the random forest (RF) yielded better Symmetric Mean Absolute Percentage Error (SMAPE) value, deep learning algorithms, namely LSTM and 1D CNN, showed similar performance compared to RF, based on t-test.