FIN.18:5.1 - Continue the forecast comparison through the actual funding rule

The four periods above are independent constructed decision windows. Their actual closing cash is measured before any funding action taken in response to the forecast. A’s absolute errors are 2, 10, 12 and 1, for an average of 6.25. B’s are 4, 1, 5 and 2, averaging 3. B has the smaller average here, but its predictions of 3 and −3 are still above the shortage outcomes of 2 and −8. Borrowing only the predicted amount needed to reach reserve 5 would leave cash at 4 and 0 in those two windows. Correctly flagging a shortage has not established an adequate funding amount.

Suppose instead that the actual available response in each independent window is a net draw of 15 before the payment cutoff, used whenever predicted cash is below 5. The full cost of a draw over that window and its repayment is stipulated as 1, paid at repayment after the measured closing time. The forecast arrives in time, repayment is separately feasible, and no other effect differs between methods. A draws in the third window, bringing its actual cash there to 7, but misses the second window’s shortage of 3. B draws in both, bringing cash to 17 and 7.

For this comparison only, assign an additional financial loss of 9 to a missed shortage; all relevant costs are included without overlap. A’s response cost is 1 + 9 = 10, while B’s is 2. If B’s extra data and operating burden costs 5 over these four uses, its total is 7 and the gain over A is 3. If that burden is 9, B’s total is 11 and the apparent gain disappears. The assumed loss, available line and repeated-use horizon are part of the decision, not universal forecasting weights.

Now add a different operating condition: a necessary source for B becomes available only after the draw cutoff. Its statistical accuracy no longer establishes this funding result. A timely simpler procedure may be preferable, or B may remain useful for another horizon. Four selected windows still cannot establish general performance; a further trial is justified only if its attainable answer can change the actual adoption decision.