The goal was to improve efficiency in cash handling and reduce the cost of cash-in-transit (CIT) services by switching from fixed delivery schedules to demand-based planning. However, this goal was only partially achieved. One of the main challenges was the lack of financial incentives for the CIT provider, who benefits from a higher number of stops. As a result, the full savings potential was not realized.

This thesis investigates how historical cash register data from the HRCC system can be used to improve cash management in a data-driven way. The analysis is based on machine learning and statistical methods to better understand cash usage patterns and optimize two areas: the daily cash float at the beginning of each cashier shift, and the timing of CIT pickups based on real needs rather than fixed schedules.

The results show that cash demand clearly depends on the day of the week, the time of day, and the location. Based on past data, starting floats were calculated using the 99% quantile of cash out-flows for each denomination and for each weekday. This makes it possible to set floats that are high enough to cover almost all situations, without keeping too much cash in the system. For CIT scheduling, a fixed threshold per denomination was defined. If this threshold is passed, a pickup is triggered. This method helps avoid unnecessary CIT visits while still making sure that there is enough space in the machine. The thesis shows that cash handling can be improved by using the available data in a smart way. The models used give clear results and can be applied in practice. They offer a useful way to re-duce cash stocks and lower costs. In the future, these methods could be expanded further, for ex-ample by also including CIT deliveries or improving the rules for triggering a pickup.