Measure cash forecast error when net cash is near zero

Percentage error is difficult to interpret when actual net cash flow is zero or changes sign. A business can receive and pay substantial amounts in the same week, leaving a small net difference. Dividing the forecast miss by that difference can produce an enormous percentage even when the error is modest relative to the money moving through the business.

Use a set of measures that matches the treasury decision: dollar error in receipts, dollar error in payments, error in the lowest available cash balance, and a consistent measure of directional bias. Retain the forecast horizon beside every score. Predicting next week's receipts and predicting receipts three months ahead are different tasks and should not share an unexplained average.

Forecasting: Principles and Practice explains why percentage errors become undefined or unstable around zero and distinguishes forecast errors from fitted residuals. The examples here apply those measurement principles to cash planning; they do not establish an acceptable error tolerance for a particular business.

In an illustrative week, actual receipts are $500,000 and payments are $490,000, leaving $10,000 of net inflow. The forecast had receipts of $480,000 and payments of $485,000, leaving negative $5,000 of net flow. Net flow is therefore $15,000 better than forecast, which becomes a 150% error if divided by the $10,000 actual net inflow.

The operating explanation is more informative: receipts were $20,000 higher and payments were $5,000 higher than forecast. Those movements partly offset. A different week could have both receipts and payments wrong by $100,000 in the same direction and show almost no net error. Net accuracy alone would conceal substantial weaknesses in the underlying schedules.

Use one sign convention in the review. For example, record error as actual minus forecast for both receipts and payments, while making clear that an excess payment reduces cash. Calculate the cash effect separately so a positive payment error cannot be read as a favorable cash result.

Mean absolute error averages the size of misses without allowing positive and negative errors to cancel. If four receipt forecasts miss by positive $20,000, negative $10,000, positive $5,000 and negative $5,000, their absolute errors total $40,000 and the average absolute error is $10,000.

The signed errors total positive $10,000, giving an average signed error of positive $2,500. Under the actual-minus-forecast convention, that indicates average underforecasting of receipts. It does not mean the typical miss is only $2,500. The absolute-error result and the bias result answer separate questions.

Break down persistent bias before altering the entire model. One recurring customer may pay earlier than assumed, or a payroll estimate may include a payment that the bank actually processes the following week. A global adjustment can hide those causes and create a new bias elsewhere.

For positive receipt series, a team may divide total absolute forecast error by total actual receipts over the evaluation period. That produces a scale-relative measure without dividing separately by each small weekly value. Document whether the calculation covers receipts, payments or another strictly defined population.

Do not apply that formula to signed net flows without considering cancellation. Positive and negative actual weeks can sum to a very small number even when the underlying activity is large. Dividing by the sum of absolute net flows is a different metric and should have a different label and written definition.

A portfolio percentage can also hide errors in smaller but critical categories. Large routine receipts may dominate the denominator while tax, debt or payroll payments carry the liquidity consequence. Preserve dollar-error detail for material obligations, even when an aggregate metric appears to improve.

Treasury may care most about whether the forecast identified a cash shortfall early enough to respond. Review the predicted and actual minimum available balance, the date of the trough and the lead time before any required funding action. A forecast that gets the monthly net movement right but misses a midmonth trough can still be operationally poor.

Available cash needs its own definition. Restricted balances, amounts held by processors and undrawn facilities subject to conditions should not be quietly combined with immediately usable bank cash. Evaluate the same measure that the team used to make its payment and funding decisions.

Record changes that occurred after the forecast was issued. An emergency collection effort or an accelerated financing draw may prevent the predicted shortfall. That intervention should appear in the review; otherwise a useful warning could be scored as a forecasting failure merely because management acted on it.

Build a two-period view for large timing errors. If a $60,000 receipt forecast for week one arrives in week two, the weekly absolute errors may total $120,000 even though the combined two-week receipt amount was correct. Keep that weekly score because timing affects liquidity, but identify the single shifted receipt in the explanation. Calling it two independent collection failures would misstate the operating issue and could direct effort toward the wrong customer process.

Save each issued forecast before replacing it. Match its periods and categories to the actual cash record, then distinguish timing misses from amount misses. A payment moving from Friday to Monday affects two weekly columns but may have no cumulative error across the combined period.

Set tolerances in dollars and response time according to the business's cash buffer and operational needs. There is no universal percentage that makes a forecast suitable for every company. A $20,000 miss has a different consequence when unrestricted cash is $30,000 than when it is $3 million.

The review should end with a specific model change or data correction where the evidence supports one. Preserve the prior forecast, the actuals cutoff and the reason for any adjustment. Future scores can then show whether the correction reduced the relevant error without relying on revised history.