How to Measure Financial Forecast Accuracy

Financial forecast accuracy is measured by comparing what you predicted with what happened, summarizing the errors in a few numbers and checking whether the forecast beats a simple alternative. The point is not to grade people but to learn where forecasts are reliable, where they are biased and whether the process is improving. This guide defines the main measures, calculates each on one data set so you can see how they differ, explains when to use which and shows how to set up an accuracy report. It supports the forecasting section of budgeting and forecasting and builds on financial forecasting.

The data

A hypothetical team forecast monthly revenue (in thousands of dollars) and compared it with actuals.

The error convention here is forecast minus actual, so a positive error is over-forecasting and a negative error is under-forecasting.

The measures

Mean absolute error (MAE)

MAE = Sum of absolute errors / Number of periods

48 / 6 = 8.0. On average the forecast was off by $8,000 a month. MAE is in the same units as the data and easy to explain. It treats all errors equally.

Mean absolute percentage error (MAPE)

MAPE = Average of (absolute error / actual)

44.75% / 6 = 7.46%. MAPE lets you compare accuracy across items of different sizes. Its weakness is that small actuals produce huge percentage errors, and a zero actual makes it undefined. It also treats a 10 percent error on a tiny item the same as on a major one.

Weighted absolute percentage error (WAPE)

WAPE = Sum of absolute errors / Sum of actuals

48 / 663 = 7.24%. WAPE weights errors by size, so large items count more. It is often preferable to MAPE for portfolios of items with different volumes. Accuracy is sometimes reported as 1 minus WAPE (92.8 percent here).

Root mean squared error (RMSE)

RMSE = Square root of (Average of squared errors)

Squared errors: 100, 25, 100, 25, 64, 100 = 414. Average = 69. Square root = 8.31. RMSE penalizes large errors more than MAE. It is higher than MAE (8.31 vs 8.0) because errors vary in size. If RMSE is much larger than MAE, a few big misses are driving the error.

Bias (mean error)

Bias = Sum of errors / Number of periods
Bias % = Sum of errors / Sum of actuals

Sum of errors = -8. Bias = -8 / 6 = -1.33. Bias % = -8 / 663 = -1.2%. A negative bias means the forecast has been lower than actual on average. Here the bias is small relative to the MAE (8), which suggests the errors are mostly random, not systematic. See forecast bias in financial planning.

Tracking signal

Tracking signal = Cumulative error / MAE

-8 / 8.0 = -1.0. A tracking signal near zero indicates no persistent bias. Values beyond about plus or minus 4 are a common flag for review, though you should set your own threshold based on experience.

Summary

Which measure to use

Use more than one. Accuracy and bias tell different stories.

Benchmarking against a naive forecast

A forecast is only good if it beats a simple alternative. The common benchmarks:

• Naive: next period equals the last actual.
• Seasonal naive: next period equals the same period last year.
• Moving average of recent periods.

In the example, a naive forecast would have used the previous month's actual (months 2 to 6): forecasts of 110, 115, 100, 125 and 118 against actuals of 115, 100, 125, 118 and 95. The errors are 5, 15, 25, 7 and 23; sum 75; MAE = 15.0. The team's forecast MAE for months 2 to 6 was (5 + 10 + 5 + 8 + 10) / 5 = 7.6. So the team's forecasts were about twice as accurate as the naive benchmark, which is evidence that the process adds value.

The improvement is (15.0 - 7.6) / 15.0 = 49 percent lower error than the naive method. This idea is called forecast value added (FVA): measure each step in the process (statistical baseline, then adjustments by sales, then by management) and see whether it improved accuracy. Steps that do not add value can be dropped.

Level of aggregation and horizon

Accuracy depends on the level you measure.

Measure at the level where decisions are made, and measure accuracy at the horizon of those decisions (for example, accuracy of the forecast made three months before the period, matching a three-month supplier lead time). A forecast measured a week before the period is not the one that drove the purchase order.

Setting up an accuracy report

Keep a snapshot of every published forecast. Forecasts that are overwritten cannot be evaluated.

Setting targets

Do not set accuracy targets by copying an industry number. Set them from your own history and the benchmark: aim to beat the naive forecast by a margin, improve year over year and keep bias near zero. Some items are inherently less predictable; compare like with like.

Using the results

1. Find where errors are largest by dollars, not only by percent.
2. Check bias by owner, product and channel.
3. Review the biggest misses for causes.
4. Update methods or inputs where a pattern exists.
5. Report accuracy alongside the forecast, so users know how much to trust it.

For product businesses, forecast accuracy at item level sets safety stock and therefore cash. See demand planning and inventory optimization. For the economics, see unit economics.

Common mistakes

• Reporting only MAPE.
• Ignoring bias.
• Not saving forecast snapshots.
• Comparing accuracy across different horizons.
• Using percentage errors on tiny items.
• Not benchmarking against naive.
• Treating accuracy as a ranking of people instead of a diagnostic.

Related guides

To understand and correct systematic over- or under-forecasting, read forecast bias in financial planning. To choose a method that fits your data, see how to choose the right forecasting method for your data.

FAQs

How do you measure forecast accuracy?

Compare forecasts with actuals using measures such as mean absolute error, weighted absolute percentage error, root mean squared error and bias, and benchmark against a naive forecast.

What is a good forecast accuracy?

It depends on the item, level and horizon. Compare with a naive benchmark and your own history rather than a generic target.

What is the difference between MAPE and WAPE?

MAPE averages the percentage errors by period; WAPE divides total absolute error by total actuals, so large items count more.

What is forecast bias?

The average tendency to over- or under-forecast, calculated as the mean of forecast minus actual.

What is a tracking signal?

Cumulative forecast error divided by mean absolute error, used to detect persistent bias.