What Is a Good Forecast Accuracy Percentage?

There is no single good forecast accuracy percentage. Accuracy, usually reported as 100% minus MAPE or WAPE, depends on the level you measure at, the time horizon and how volatile demand is. A good forecast beats a simple naïve forecast and your own past accuracy, at the level where decisions are actually made.

Key takeaways

• The same forecast can score close to 99% for a product family per month and below 80% per item per week. Always state the level and horizon with the number.
• WAPE (total absolute error divided by total actuals) is a safer basis for an accuracy percentage than MAPE, which breaks on low or zero sales.
• Judge accuracy against a benchmark: a naïve forecast, last year's accuracy, or the error you can afford given lead times and safety stock.
• Track bias separately. A forecast that is consistently 10% high can look "accurate" on average while filling warehouses.
• Measure at the decision level: weekly per distribution centre for replenishment, monthly per product family for capacity, quarterly for budgets.

What is forecast accuracy?

Forecast accuracy is how close predicted values are to what actually happened. Most teams calculate a forecast error first and then express accuracy as its complement: if the error is 15%, accuracy is 85%.

The trouble is that "error" can be calculated in several ways, and each gives a different percentage for the same forecast. Before comparing your number with anyone else's, check which formula, which aggregation level and which time lag they used.

How do you calculate forecast accuracy?

These are the formulas you will meet most often. A is the actual value and F the forecast for each item and period.

Forecast accuracy % is then usually 100% − WAPE, floored at zero. The forecasting textbook by Hyndman and Athanasopoulos explains why percentage errors such as MAPE are "infinite or undefined" when actuals are zero and extreme when they are close to zero (Forecasting: Principles and Practice, section 5.8). That is the main reason demand planners move from MAPE to WAPE. AWS's forecasting documentation describes WAPE the same way and notes it is less sensitive to outliers than RMSE because it uses absolute rather than squared errors (Amazon Forecast accuracy metrics).

A worked example: one forecast, three accuracy numbers

Take three products in one week (illustrative numbers):

• Total level: forecast 160 against actual 162. Error 1.2%, so accuracy is 98.8%.
• Item level, WAPE: 38 ÷ 162 = 23.5% error, so accuracy is 76.5%.
• Item level, MAPE: 11.1%, 28.6% and 400%, averaging about 147% error. Accuracy is meaningless, purely because product C sold 2 units.
• Bias: (160 − 162) ÷ 162 = −1.2%, close to neutral.

Nothing about the forecast changed between those lines. Only the measurement did. This is why a vendor claiming "95% accuracy" tells you very little until you know the level, the metric and the horizon.

What is a good forecast accuracy percentage?

A good number is one that is better than a simple benchmark and good enough for the decision it supports. Published "industry averages" rarely state their formula or level, so treat them with caution and build your own baseline instead.

Use three tests:

1. Does it beat a naïve forecast? A naïve forecast just repeats the last value, or the value from the same period last year for seasonal items. If your model cannot beat that, it is not adding value. MASE makes this explicit: in the textbook's definition, a scaled error below one means the forecast is better than the naïve one (FPP3).
2. Is it improving against your own history? Compare this quarter's WAPE with the same quarter last year, at the same level and lag.
3. Is the remaining error affordable? Translate error into safety stock, lost sales or idle capacity. If cutting error by two points costs more than the extra stock it saves, stop there.

A useful reality check from the same textbook: on quarterly Australian beer production, a plain seasonal naïve method scores a MAPE of 3.17%, which would be reported as about 97% accuracy. Smooth, seasonal series are easy. Volatile, intermittent ones are not, and a lower percentage on them can still be the better forecast.

What drives the number up or down

How to improve forecast accuracy

• Clean the history. Remove stock-out periods (sales of zero because shelves were empty), duplicates and one-off orders before training.
• Start simple. A seasonal naïve or exponential smoothing model is a fair baseline and easy to explain. Add complexity only when it beats the baseline in backtests.
• Use backtesting. Score models on periods they have not seen, ideally with rolling-origin cross-validation, rather than on the data they were fitted to.
• Add drivers that you actually know in advance: price changes, promotions, holidays, short-range weather for seasonal goods.
• Measure the value of manual overrides. Compare the statistical forecast with the overridden one. Sales and planner adjustments often add bias.
• Agree one number across functions. Sales, marketing, operations and finance should review the same forecast, with known assumptions, on a fixed cadence.
• Plan for the error you cannot remove. Sometimes shorter lead times, smaller batches or flexible capacity pay off more than another model.

Machine learning and cloud forecasting tools

Machine learning models help most when you have many related series and useful external drivers, such as thousands of SKUs across many stores. They help less on short, sparse histories. On AWS, note that Amazon Forecast is no longer available to new customers (AWS documentation). New projects typically use SageMaker or open-source libraries, with the same accuracy metrics described above. Whatever the tool, keep the naïve baseline in your reports so you can see what the model adds.

How Opsio helps

Opsio's predictive analytics consulting and data analytics and BI teams help organisations set up the data pipelines, backtesting and dashboards behind a trustworthy accuracy number. When models move to production, our MLOps practice handles retraining, monitoring and drift alerts in the cloud. For the wider picture of how ML differs from other AI disciplines, see computer vision vs machine learning.

Frequently asked questions

Is 80% forecast accuracy good?

It depends on the level. At item-by-week level for a mixed retail range, 80% (100% − WAPE) can be strong. At company-by-month level it would usually be weak. Compare it with a naïve forecast at the same level before deciding.

How do I calculate forecast accuracy in Excel?

Put actuals in column A and forecasts in column B. Calculate absolute error per row with =ABS(A2-B2). Then accuracy = 1 − SUM(absolute errors) ÷ SUM(actuals), which is 1 − WAPE. Format it as a percentage and floor negative values at zero.

What is the difference between forecast accuracy and forecast bias?

Accuracy measures the size of the errors regardless of direction. Bias measures their direction: whether you forecast too high or too low on average. You need both, because errors in opposite directions cancel out in bias.

Why is my MAPE so high when the forecast looks fine?

Usually because some items have very small actuals. A miss of 8 units on an item that sold 2 is a 400% error, which drags the average up. Switch to WAPE or MASE, or segment slow movers.

Can forecast accuracy be negative?

Mathematically, 1 − MAPE or 1 − WAPE can go below zero when errors exceed actuals. Most teams floor the figure at 0% and look at the underlying error instead.