You measure forecast accuracy by comparing what you predicted to what actually happened, then summarizing those differences with a metric that fits your situation. The most widely used metrics are Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Mean Absolute Percentage Error (MAPE), but each one tells you something slightly different, and picking the wrong one can make a bad forecast look good or a good forecast look bad.
Start With the Error
Every accuracy metric begins with the same raw ingredient: the forecast error for each period. That’s simply the forecast value minus the actual value. If you predicted 500 units of sales and 480 actually sold, your error for that period is +20. If you predicted 500 and 530 sold, it’s -30. The sign tells you the direction: positive means you overshot, negative means you undershot.
What you do with those individual errors determines which metric you’re using and what kind of insight you get.
Mean Absolute Error (MAE)
MAE is the simplest and most intuitive accuracy metric. You take the absolute value of each period’s error (dropping the positive or negative sign), then average them. If your errors across four months were +20, -30, +10, and -15, the absolute values are 20, 30, 10, and 15, and the MAE is 18.75.
MAE tells you the average size of your misses in the same units as your data (dollars, units, headcount, whatever you’re forecasting). That’s its strength and its limitation. An MAE of 10 is excellent if your typical demand is 1,000 per period, but terrible if demand is only 12. Because MAE isn’t scaled to the size of what you’re measuring, it works best when you’re evaluating a single product, single revenue line, or single time series where you already have context for what “big” and “small” mean.
Root Mean Square Error (RMSE)
RMSE works like MAE’s stricter sibling. You square each error, average the squared values, then take the square root. The squaring step means large errors get amplified before averaging, so a single major miss pulls the RMSE up much more than it would pull up the MAE.
This makes RMSE the right choice when big misses are far more costly than small ones. If a 200-unit shortfall triggers a stockout that shuts down a production line, you care more about avoiding that one catastrophic miss than about shaving a few units off your average error. RMSE captures that priority. Like MAE, it’s expressed in the same units as your data and isn’t scaled to the level of demand, so the same contextual limitation applies.
Mean Absolute Percentage Error (MAPE)
MAPE converts each error into a percentage of the actual value before averaging. If you forecasted 500 and the actual was 480, that period’s percentage error is about 4.2%. Averaging those percentages across all periods gives you a single number you can compare across products or business lines of very different sizes, which is why MAPE is one of the most commonly reported accuracy metrics in business.
It has a serious flaw, though. Because each error is divided by that period’s actual value, any period where actual demand is very low (or zero) produces a wildly inflated percentage. Forecast 10, sell 2, and your error for that period is 400%, which can dominate the entire average. This skew tends to reward forecasts that consistently undershoot demand, since undershooting keeps the denominator relatively large. For products with lumpy or intermittent demand, MAPE can be misleading enough to avoid entirely.
A common alternative is Weighted MAPE (WMAPE), which divides the total absolute error by total actual demand across all periods rather than averaging individual percentages. This prevents low-demand periods from hijacking the result.
What the Numbers Look Like in Practice
There’s no universal threshold for “good” accuracy because it depends heavily on what you’re forecasting. According to Gartner data cited by the Institute for Supply Management, median demand forecast error in the food and beverage industry runs about 25%, with top-quartile performers around 20%. In durable consumer products, the median error is roughly 50%, because demand is lumpier and harder to predict.
Revenue forecasts at stable subscription businesses might hit single-digit percentage errors, while new-product launch forecasts routinely miss by 30% or more. The right benchmark is your own recent track record: measure your current accuracy, then try to improve it incrementally rather than chasing someone else’s number.
Measuring Bias Separately
Accuracy metrics like MAE and RMSE strip away the direction of your errors by using absolute values or squaring. That means they can’t tell you whether your forecasts consistently lean too high or too low. A forecast that overshoots by 100 every single month and one that randomly bounces between +100 and -100 will produce the same MAE, but they represent very different problems.
Bias is simply the average of the raw errors with the sign included. If you consistently get positive errors, your forecasts are systematically too high. Negative on average means you’re systematically too low. A bias near zero doesn’t guarantee accuracy (random large errors can cancel out), but it tells you the forecast isn’t tilted in one direction.
A practical way to spot bias early is to watch the sequence of errors. Four or more consecutive errors in the same direction (all positive or all negative) gives you roughly 90% confidence that something systematic is happening, not just random noise. That’s a useful trigger for investigating whether your model, your assumptions, or your data have shifted.
Picking the Right Metric
No single metric works perfectly in every situation. Here’s a practical way to choose:
- You’re tracking one product or one forecast at a time: MAE is the clearest starting point. It’s easy to calculate, easy to explain, and treats all errors equally.
- Large misses are especially costly: Use RMSE. Its squaring mechanism flags the forecasts with occasional big blowups, even if the average miss is small.
- You need to compare accuracy across products of different sizes: Use WMAPE rather than standard MAPE. It gives you a percentage you can compare across a $10 million product line and a $200,000 one without letting low-volume periods distort the picture.
- You suspect the forecast leans one direction: Track bias alongside your main accuracy metric. Accuracy tells you how far off you are; bias tells you whether you keep missing in the same direction.
Using the wrong metric can make some teams or models appear better, or worse, than they actually are, simply because of how the math works. If you evaluate a forecasting model with MAPE on a product that has frequent zero-demand periods, the model will look awful regardless of how well it handles the rest of the pattern.
Setting Up a Measurement Process
Calculating a metric once isn’t the same as measuring forecast accuracy. To get real value, you need a repeatable process.
First, decide on a consistent time lag between the forecast and the actuals you compare it to. If your team produces a forecast on the first of each month for that month’s sales, always compare to that same month’s final numbers. Mixing forecast horizons (comparing a 30-day-out forecast one month and a 7-day-out forecast the next) makes trends in accuracy impossible to read.
Second, track your chosen metric over time, not just as a snapshot. A single month’s MAE doesn’t tell you much. A rolling 6- or 12-month trend shows whether your forecasting process is improving, degrading, or holding steady. Plot it on a simple chart alongside any changes you’ve made to models, data sources, or methodology so you can see what’s actually helping.
Third, segment your results. An aggregate accuracy number across all products or all regions can hide pockets of terrible performance. Break your accuracy metric down by product category, business unit, or customer segment. You’ll often find that 80% of your total error comes from a small number of items, which tells you exactly where to focus your improvement efforts.
Finally, pair your accuracy metric with bias tracking. If accuracy is getting worse but bias is near zero, your forecast is noisy but not systematically broken. If bias is growing, you have a structural problem, like a model that hasn’t been updated for a market shift, that won’t fix itself with more data alone.