26%Real bank data from an open research dataset
Next quarter's receipts, with a range that holds
less error than a 13 week average, forecasting weekly receipts 1 to 13 weeks ahead
- What was observed
- 1,139 accounts, forecast at two dates held out from tuning; most account weeks had no receipts at all.
- What a generic approach says
- Carry the last 13 weeks' average forward.
- What the engine read
- A forecast from each account's own history cut the error, weighted by amount, by 26%, and its 90% range for the quarter's total held 92% and 90% of the time at the two dates.
- What it meant for the decision
- A receipts forecast with a range that holds gives a renewal or a limit a figure to plan on, rather than a single line.