Page 100 - Special Topic Session (STS) - Volume 4

STS566 K. Prokopenko et al.

Complex seasonal autoregressive model
compared to machine learning methods for cash
volume forecasting
K. Prokopenko, B. Bruijnis, M. Symotiuk
Giesecke+Devrient Currency Technology GmbH, Prinzregentenstrasse 159, 81677 Munich,
Germany

Abstract
As expert in cash volume forecasting Giesecke+Devrient has invested in
software that uses algorithms for cash volume forecasting based on data
available from bank branches, cash devices and retail stores. This paper
describes the validation of the data and the algorithms applied to this data.
Based on statistical validated data from cash handling locations in Western
Europe the seasonal behaviour of cash usage has been described. Based on
the trends found three forecasting models have been validated: Complex
Seasonal ARMA (CSARMA), LSTM recurrent neural network and XG Boost
decision trees. The latter two are standard machine learning (ML) algorithms
while CSARMA is a new approach based on complex seasonal autoregressive
model of global trend, annual and weekly shapes. The forecast validation
shows that for this data set the CSARMA algorithm outperforms the ML based
algorithms.

Keywords
Cash volume forecasting; Seasonality; Time series; Machine learning; Auto
regression; Neural networks; Decision trees

1. Introduction
Despite rumours about the cashless society in several countries like
Georgia the absolute cash usage is still growing [5]. In other countries cash
usage is in decline with Sweden as one of the most appealing examples.
However, despite this movement to other means of payment, a fully cashless
society is not there yet [6].
The combination of decreased cash usage and the desire to allow cash to
remain as payment method demands Central banks and other actors in the
cash supply chain to minimise the cost of handling cash while keeping cash
availability without cash-outs. One cash cost improvement that doesn’t require
the infrastructure to be changed is the improvement of the cash usage
forecast. In this paper the focus is on customer incoming and outgoing
payments.
Using data of several companies in Western Europe Giesecke+Devrient has
developed algorithms that predicts cash requirements in the market
considering seasonality. This paper describes the decision making towards the
algorithm to use. By knowing the demand from the end customer Central

89 | I S I W S C 2 0 1 9