ATM cash demand forecasting in an Indian bank with chaos and hybrid deep learning networks

•Deep learning models, LSTM, 1D-CNN outperformed other models w.r.t SMAPE.•RF, LSTM & 1D-CNN + LSTM are statistically similar in non-chaos-observed-ATMs.•In chaos-observed-ATMs, RF and LSTM was statistically significant.•GRU displayed less variability in predictions than other techniques w.r.t. SMAPE.•Including exogenous dummy variable significantly improved the predictions.