Models / Cash & ATM Operations
ATM Cash Demand Forecaster
Forecasts cash withdrawals per ATM, 14 days ahead.
MAPE (daily, per terminal)
11.8%
WAPE (network)
6.4%
Cash-out days avoided (pilot)
-38%
Idle cash reduction (pilot)
-14%
About this model
Forecasts daily cash withdrawals for every ATM and cash recycler in the network, 14 days ahead, so cash-in-transit orders carry less idle cash and fewer machines run dry. Accounts for payday cycles, public and state holidays, school terms and nearby branch closures.
Intended use
Cash replenishment planning by Group Operations and the cash-in-transit vendor schedule. Aggregated at terminal level; no customer data involved.
Training data lineage
Global LightGBM model trained in Vertex AI Pipelines on three years of terminal-level withdrawal aggregates from casa-transactions, joined to terminal attributes from branch-atm-network in BigQuery. Retrained monthly; each run is logged in Vertex AI Experiments and promoted through the model registry.
Limitations & bias notes
Newly installed terminals rely on similar-site priors for their first eight weeks and run roughly 25% higher error. Unplanned outages and off-site machines in shopping malls during sales events are the largest error sources.
Ownership and sensitivity
Who approves access
No approval needed. Self-service: subscribe and use, logged for chargeback.
In Safe to try
Open to all staff. Aggregates, policy text, public series and synthetic data.
Self-service: subscribe and use, logged for chargeback.
- Aggregated only
- No row-level customer data; aggregates with small-cell suppression.
Try it
Live sandboxterminal=ATM-KL-0214&horizon=12
fixtureFeedback
Details
- Updated
- 2026-07-08
- Latest version
- 2.1.0
- Licence
- Group Reuse
- Access
- Open to all BUs
- Framework
- LightGBM
- Language
- N/A