Forecast Accuracy Metrics: How to Measure and Choose

To measure forecast accuracy, subtract each actual result from its forecast, then summarize those errors with a metric that fits your business. Use WAPE or MAE for most revenue and expense forecasts, MAPE when every value is comfortably above zero, and always track bias next to them. A metric tells you how far off you were. Bias tells you which way you keep missing.

This guide is for owners, controllers and finance leads who forecast revenue, cash, expenses or demand and want to know whether the forecast deserves trust. We walk through the main metrics, calculate each on one small dataset, and show how to pick one without overcomplicating things.

TL;DR

• Pick one accuracy metric (WAPE or MAE for most finance teams) and one bias measure, and report both every month.
• MAPE is popular but breaks when actual values are zero or near zero. Scaled or weighted alternatives such as MASE and WAPE handle that better.
• Judge a forecast only on data it never saw. A model that fits history well can still forecast badly.
• There is no universal "good" percentage. Compare against your own past accuracy and a simple baseline.
• Simpler, explainable forecasts often hold up as well as complex ones, so favor models your team can defend.

Table of Contents

• What forecast accuracy actually measures
• A worked example you can copy
• The core metrics, one by one
• Which metric should you use?
• Test your forecast on data it has not seen
• What is a good forecast accuracy percentage?
• Keep the forecast explainable
• Our perspective
• Build the scorecard in five steps
• Common mistakes that distort accuracy numbers
• Need help turning forecasts into decisions?
• FAQ
• Sources

What forecast accuracy actually measures

A forecast is a number you commit to before the facts arrive. Accuracy is the gap between that number and what happened. Every metric below starts from the same raw material: the forecast error for each period.

We define it as error = forecast minus actual. A positive error means you over-forecast. A negative error means you under-forecast. Some textbooks flip the sign, so state your convention once and keep it everywhere. Mixing conventions is the most common reason two analysts report opposite bias for the same data.

The OpenStax data science text groups the common measures into scale-dependent (MAE, RMSE), percentage-error (MAPE, sMAPE), scaled (MASE) and weighted (WAPE) families. Which family fits depends on whether you need to compare across series of different sizes, or to punish particular kinds of misses such as large outliers or systematic bias. No single metric wins in every case.

A worked example you can copy

Suppose a services company forecast monthly revenue (in thousands of dollars) and then compared it with actuals for six months.

We will reuse these numbers for every metric.

The core metrics, one by one

MAE: mean absolute error

Add the absolute errors and divide by the number of periods. Here: 55 / 6 = $9.2K. MAE is in the same unit as the thing you forecast, so it is easy to explain: "we miss by about $9,200 a month on average." Its weakness is scale. A $9,200 miss is small for a $1M business and huge for a $50K one, so you cannot compare MAE across different lines of business.

RMSE: root mean squared error

Square each error, average the squares, then take the square root. Here: the squares sum to 525, divided by 6 is 87.5, and the root is $9.35K. Because squaring amplifies large misses, RMSE rises sharply when you have one or two bad months. Use it when a big miss is far more painful than several small ones, for instance when a single cash shortfall can breach a loan covenant. It is the less intuitive metric to explain to a board.

MAPE: mean absolute percentage error

Divide each absolute error by that month's actual, then average. Here: 51.6% / 6 = 8.6%. MAPE is the most widely quoted metric because a percentage is instantly understood. It has two known problems. It cannot handle zero actuals, and it treats a miss on a tiny month the same as a miss on a huge one. The AutoGluon time series documentation notes that MAPE is mathematically problematic when data contains zeros, since it divides by the actual value. That matters for project-based revenue, seasonal products and any line that is sometimes nil.

WAPE: weighted absolute percentage error

Divide the total absolute error by the total actual. Here: 55 / 655 = 8.4%. WAPE keeps the readability of a percentage but weights big months more heavily, and it survives the occasional zero. For most finance teams forecasting revenue or spend across many customers, products or departments, it is the most practical headline number. The same AutoGluon guidance points practitioners toward WAPE or MASE over MAPE when series have different scales or contain zeros.

MASE: mean absolute scaled error

MASE divides your forecast's average error by the average error of a naive benchmark, such as "next month equals the same month last year" for seasonal data. A MASE below 1 means you beat the naive guess. Above 1, you would have done better doing nothing clever. It works across series of any scale, which makes it the best choice when you compare dozens of lines. The cost is that it needs a benchmark and is harder to explain to non-specialists.

Bias: the direction of your misses

Sum the signed errors and divide by total actuals. Here: +5 / 655 = +0.8%, a slight over-forecast. Our six months look accurate by every metric above, and the bias is nearly zero because the over- and under-forecasts offset. That is good news. A forecast with an 8% WAPE and a bias of +7% is a different problem: it is consistently optimistic, and the fix is a change in assumptions rather than more modeling.

The AutoGluon documentation draws the same line: accuracy metrics capture the size of errors, bias captures direction, and you should measure both. It also warns that bias alone cannot rank models, because a forecast that is too high half the time and too low half the time can have zero net bias and still be badly wrong.

Which metric should you use?

Researchers writing in the Journal of the Operational Research Society describe a gap between the measures statisticians favor and the ones practitioners actually need. Their advice is to match the metric to the business goal. If inventory cost is what you care about, choose a measure that weights higher-volume items more heavily. If your concern is cash runway, pick the metric that reflects cash dollars missed, not an average of percentages that treats a $2K line like a $2M one.

Test your forecast on data it has not seen

The most reliable accuracy number comes from genuine forecasts. Rob Hyndman's paper on measuring forecast accuracy is blunt: how well a model fits history says little, and accuracy can only be judged on data not used to build the model. A spreadsheet that has been tuned until it reproduces last year perfectly will look wonderful in-sample and then disappoint.

In practice, that means:

1. Save each forecast as it was issued, with its date. Do not overwrite it when actuals arrive.
2. Compare the saved forecast with the actual once the period closes.
3. Record the horizon. A forecast made one month out and one made twelve months out are different animals and need separate accuracy numbers.
4. If you build statistical models, hold back the most recent periods as a test set.

One technical warning for anyone using machine learning tools: a review of econometric methods cautions that standard K-fold cross-validation mixes future and past observations in time series. Use rolling or expanding windows that respect chronological order.

What is a good forecast accuracy percentage?

There is no honest universal answer. A mature subscription business with contracted revenue should forecast next month's revenue within a few percent. A project business with a lumpy pipeline will not. A new product with no history can miss by far more and still be forecast reasonably.

Use three yardsticks instead of a published benchmark:

• Your own trend. Is WAPE improving quarter over quarter?
• A naive baseline. Does your forecast beat "same as last month" or "same as last year"? If not, the added effort is not paying off.
• Decision tolerance. How large a miss can you absorb? If a 10% revenue shortfall means missing payroll, 10% is your ceiling no matter what any benchmark says.

Benchmarks quoted online vary widely by industry and are often published by software vendors, so treat them as rough context only.

Keep the forecast explainable

It is tempting to chase accuracy with ever more elaborate models. The evidence says to be careful. The M3 forecasting competition results found that statistically sophisticated methods do not necessarily produce more accurate forecasts than simpler ones. For a business, that supports a plain rule: if you cannot explain why the forecast says what it says, you cannot fix it when it is wrong.

This is the same logic behind treating AI as decision support rather than an autopilot. Let tools produce the number, but keep a person responsible for the assumptions behind it, and judge both by the same accuracy and bias scorecard.

Our perspective

When we look at a company's forecasting, the metric is rarely the real problem. The usual culprits sit upstream: books that are weeks behind, a chart of accounts that blends unlike revenue streams, and assumptions that live in one person's head. You cannot measure accuracy on actuals you do not trust. That is why we start with clean, timely bookkeeping, then build the forecast, then score it.

Our practical advice for owners:

• Pick one headline accuracy metric and one bias number. Resist a dashboard of eight.
• Score the forecast the month the books close, and write down why the biggest miss happened. The explanation is worth more than the percentage.
• Treat persistent bias as a management conversation, not a math problem. If you over-forecast every quarter, the sales assumptions are too generous.
• Update the forecast on a rolling basis instead of defending a number set in January.

Our approach to fractional CFO work is modular, so a business can bring in forecasting and variance review without hiring a full-time executive.

— Angelica

Build the scorecard in five steps

1. Define what you forecast and the horizon. Revenue by month, cash by week, expense by department. Be specific.
2. Freeze the forecast. Save a dated copy before the period starts.
3. Close the books. Use reconciled actuals, not estimates.
4. Calculate the metrics. WAPE or MAE, plus bias, in a simple spreadsheet. Add a naive baseline for comparison.
5. Review and adjust. Look at the two or three biggest misses, name the cause, and change the assumption that caused it.

If you want to go deeper on the forecast itself, our guides cover rolling forecasts for FP&A teams, financial forecasting models and how to build a forecasting process for a smaller business. For SaaS companies, revenue forecasts that reconcile to ASC 606 deserve their own treatment.

Common mistakes that distort accuracy numbers

• Averaging percentages across unlike lines. A 40% miss on a $3K line drags a simple average far more than it hurts the business. Weight by dollars.
• Reporting accuracy without bias. Offsetting errors can hide a forecast that is wrong in every single period in the same direction, or hide a healthy one that only looks noisy.
• Changing the forecast after the fact. Revised forecasts are fine, but score the original and each revision separately.
• Ignoring the horizon. A one-month-ahead number flatters you. Report accuracy at the horizon where decisions are made.
• Chasing a benchmark. Compare against your baseline and your tolerance, not someone else's industry average.

Need help turning forecasts into decisions?

Accurate forecasting depends on accurate books. AmCFO provides bookkeeping, QuickBooks setup and cleanup, budgeting, forecasting and financial analysis, with fractional CFO oversight, so your forecast starts from numbers you can trust. If you want a second pair of eyes on your forecast process, see our fractional CFO services or contact us through our site.

FAQ

How do you measure the accuracy of a forecast?

Subtract each actual from its forecast, then summarize the errors with a metric such as MAE, WAPE or MAPE. Always score forecasts made before the period started, using closed actuals. Add a bias figure so you know whether you lean high or low.

What is a good MAPE for forecasting?

It depends on how predictable the underlying business is, so no single figure applies everywhere. A better test is whether your MAPE improves over time and beats a naive baseline such as last year's same month. Remember that MAPE misbehaves when actual values are zero or very small.

What is the difference between forecast accuracy and forecast bias?

Accuracy describes how large your errors are. Bias describes whether they lean in one direction. A forecast can have zero bias because over- and under-forecasts cancel out and still be far from the actuals, so track both.

Should I use MAPE or WAPE?

Use WAPE when you forecast many items or when some actuals can be zero, because it weights larger values and avoids division by zero. MAPE is fine for a single positive series when your audience wants a simple average percentage. In our worked example they landed close together, at 8.6% and 8.4%.

Sources

• 5.4 Forecast Evaluation Methods, OpenStax
• Forecasting Time Series: Evaluation Metrics, AutoGluon documentation
• On the selection of forecasting accuracy measures, Journal of the Operational Research Society
• Measuring forecast accuracy, Rob J. Hyndman
• Econometrics of Machine Learning Methods
• M3 Competition summary

Recommended

• Financial Forecasting Models: A Practitioner's Guide
• Business Forecasting Methods for Financial Leaders
• Financial Modeling Explained: A Guide for Analysts
• Rolling Forecast Explained for FP&A Teams: Quick Setup