Daily cash position for a card-payments business over the next 13 weeks, built from the drivers that move cash: card volume, settlement timing, merchant payout schedules, payroll, vendor runs and taxes. Move the levers to stress it.
Applies to all volume after today. Receipts and merchant payouts both move, but payouts lag.
Share of settled volume charged back three days later.
Paid on the first business day of week
Policy floor. P10 days below it are flagged.
ActualForecastP10–P90Buffer
Week ending
Receipts
Merchant payouts
Payroll
Other out
Net
Ending balance
Low P10
$ millions. "Other out" is vendors, estimated tax, chargebacks and any unplanned payment, less interest received. Weekly balances end on Fridays, the weekly payout day, which is why each week's low point is usually its last day.
Which forecasting method to trust
Method
Daily MAPE in
Daily MAPE out
Weekly MAPE in
Weekly MAPE out
Week-13 balance error
Daily MAPE covers the first two weeks of each forecast; weekly MAPE covers all 13 weeks. Balance error is the mean absolute miss on cumulative net cash by week 13.
Seasonal naive (same weekday last week) breaks on every holiday and month end.
Holt-Winters follows the level but cannot see payroll, tax dates or holiday pile-ups.
Gradient boosting on total inflows and outflows learns the calendar and gets both totals within about 7%, but its inflow and outflow errors do not cancel, so the balance drifts.
Driver-based forecasts card volume once, then derives receipts, payouts and chargebacks from it using the settlement rules. A volume miss moves both sides together, which is why its balance error is a fraction of the others'.
How it works
Card volume is forecast by calendar day with gradient boosting on the ratio to its trailing 28-day level, so growth and seasonality (paydays, holidays, the November–December peak) carry through.
Volume settles T+1 on business days, so weekend and holiday volume piles onto the next business day. Daily merchants are paid a day after settlement; weekly merchants are paid in a Friday batch.
Payroll, vendor runs and estimated taxes come from the treasury calendar; interest uses last month's amount; unplanned payments are not forecast.
The P10–P90 band is σ·√(business days ahead), with σ fitted on backtest errors of cumulative net cash.
All data is synthetic, generated with known settlement rules so each method can be checked against the truth. Built with AI-assisted coding (Claude); modeling choices, validation and review are my own.