What Is Forecast Accuracy and How Do You Measure It?

You forecast 1,000 units, sell 700, and now you are sitting on dead stock that will move at a discount in four months. The forecast was not a small amount wrong. It was wrong in a way nobody measured until the warehouse made it obvious, which is the actual problem.

Forecast accuracy is how close your demand forecast came to actual sales over a period. You measure it by comparing forecast against real units sold, per SKU, and expressing the gap as a percentage or in units. Higher accuracy means fewer stockouts, less dead stock, and buying decisions you can defend.

Key takeaways

• Accuracy is not the same as a good forecast: a tight number on a product nobody buys is worth less than a rough number on your best-seller.
• Measure at the level you buy: catalog-wide accuracy hides the SKU-level misses that actually cost money.
• Size and direction are separate questions: how far off you were, and whether you lean the same way every month.
• Errors cost in both directions: under-forecast and you stock out, over-forecast and you mark down.

What does forecast accuracy actually mean?

Forecast accuracy is the distance between what you predicted and what happened, measured over a defined period and at a defined level of detail. Both of those qualifiers matter. Accuracy over a quarter says something different from accuracy over a week, and accuracy across your whole catalog says something very different from accuracy on one product.

Three words get used interchangeably in this area and should not be. Error is the raw gap for one period: you forecast 100, you sold 120, the error is 20 units. Accuracy is what you get when you summarise those errors into a readable score. Bias is whether those errors lean consistently in one direction rather than scattering either side of zero. A forecast can be accurate on average and badly biased, which is the combination that quietly builds overstock while the dashboard stays green.

Accuracy is not the same as a good forecast

A high accuracy score on the products that barely sell is not an achievement, and this trips up more teams than it should. If your catalog is mostly slow movers, forecasting them all at zero will give you an impressive-looking overall score and tell you nothing useful. What matters is accuracy weighted by what the product is worth: the SKUs carrying your revenue deserve tight forecasts, and a two-point improvement there is worth more than a twenty-point improvement across the long tail. Judge the score against the value at stake, not against the number of products it covers.

Measure at the level you buy

Accuracy aggregated across the catalog is close to meaningless, because errors in opposite directions cancel. Over-forecast one product by 200 units, under-forecast another by 200, and the total looks perfect while both decisions were wrong. Since you place purchase orders per SKU, measure per SKU. Roll up afterwards if leadership wants one number, but do the diagnosis at the level where the buying happens. This is the single most common measurement mistake in demand planning, and it is entirely self-inflicted.

Why does forecast accuracy matter for your inventory?

Because the forecast is the input to every inventory number you calculate, and an error in it propagates to all of them at once. Your reorder points, your safety stock, your coverage reads, and your purchase quantities are all derived from expected demand. Get that wrong and everything downstream is confidently, precisely wrong.

The cost runs in both directions, which is what makes accuracy worth measuring rather than simply padding. Under-forecast and you stock out: you lose the sale, you lose the ad spend that drove the click, and sometimes you lose the customer to whoever had it in stock. Over-forecast and you overbuy: cash sits on a shelf, storage costs accrue, and the stock eventually moves at a markdown that eats the margin you booked it for. Neither failure is cheap, and padding the forecast upward simply trades the loud failure for the quiet one.

How do you measure forecast accuracy?

You compare forecast to actual over a period and summarise the gap, and there are three readings worth having rather than one. Each answers a different question, and using only one is how blind spots survive.

• The percentage read: expresses each period's miss as a percentage of what sold, then averages. It travels well between teams and across products of different sizes, but it distorts badly on low-volume items where dividing by a near-zero number inflates the result.
• The unit read: averages the size of your misses in actual units. It stays readable on slow movers and maps directly onto a purchase order, which is why buyers tend to prefer it.
• The direction read: sums the signed errors instead of the absolute ones, so pluses and minuses can cancel. If the running total drifts steadily one way, you have bias rather than noise.

Each of those has its own method, its own arithmetic, and its own failure mode, and this page deliberately does not derive any of them. The comparison of the three main magnitude measures, including which suits which SKU profile, is in forecast accuracy metrics. The raw building block underneath all of them is covered in forecast error, and the direction read has its own guide in forecast bias.

Rule of thumb: track one magnitude measure plus direction. Two numbers change decisions; six numbers get ignored.

What is a good forecast accuracy score?

There is no universal target, and any figure quoted without asking what you sell is guessing. What counts as good depends on three things: how volatile the product's demand is, how far ahead you are forecasting, and how much history you have. A steady staple with three years of clean sales data should land far tighter than a seasonal item in its first year.

As a rough frame, steady sellers with reliable history often sit in the range of ten to thirty percent average error, seasonal items run higher because timing errors magnify quickly, and new products with no history run higher still. Those are orientation, not targets. The useful benchmark is your own trend: is the error on your top revenue SKUs smaller this quarter than last, and is it smaller than the method you replaced would have produced on the same data? That comparison is answerable and honest. An industry benchmark from a different catalog is neither.

This is also where the difference between measuring accuracy and improving it becomes practical. Conative AI's proprietary deep-learning models forecast at the product level and read live marketing signals, ad spend, sales velocity, and campaign events, alongside sales history, so the forecast reflects where demand is heading rather than where it's been. Forecast accuracy is tracked per SKU so the trend is visible rather than asserted, and any forecast falling outside its accuracy guardrails gets flagged instead of applied quietly. See a demo of the accuracy view on the inventory planning platform.

How do you improve forecast accuracy once you measure it?

Measuring is the easy half. Improving comes down to three levers, in this order. Fix the data first, because every method inherits it: reconcile duplicate SKU records, correct returns booked to the wrong period, and mark the weeks a product was out of stock so the model does not learn that a stockout week was genuinely low demand. That last one is the highest-value cleanup most brands have never done. Segment second, so your effort concentrates where the money is rather than spreading evenly across a catalog. Correct bias third, because a forecast leaning the same way every month has a cause you can name and remove, whereas random error mostly does not.

Only after those three is it worth arguing about method. Swapping a moving average for something more sophisticated on top of dirty, unsegmented data usually produces a more expensive version of the same misses. The full playbook of improvement levers, including the human ones, lives in how to improve demand forecast accuracy.

Frequently asked questions

What's the difference between forecast error and forecast accuracy?

Error is the raw gap for one period: forecast 100, sold 120, error 20 units. Accuracy is what you get when those errors are summarised into a score across periods. Error is the ingredient; accuracy is the dish. You need the errors recorded per period before any accuracy measure means anything.

Should I measure accuracy in percentages or units?

Both, for different jobs. Percentages compare cleanly across products of different sizes and travel well in a report, but they become unstable on low-volume items. Units stay readable on slow movers and map straight onto a purchase order. Most planners report a percentage upward and buy against units.

What's a good accuracy score for an eCommerce brand?

It depends on catalog volatility and forecast horizon, so treat any single figure with suspicion. Steady sellers with clean history land far tighter than seasonal or newly launched products. The benchmark worth using is your own trend on your top revenue SKUs against the method you replaced.

How often should you measure forecast accuracy?

On the cycle you buy, which is monthly for most eCommerce brands, with a lighter check on fast movers in between. Measuring more often than you act produces noise. Measuring less often means a bad pattern runs for months before anyone sees it.

Can AI improve forecast accuracy?

It can, by weighing more signals than sales history alone and refreshing more often than a manual rebuild. Neither is a guarantee, since a model inherits the quality of the data it learns from. The honest test is measuring it against your current method on your own history.

What causes low forecast accuracy?

Most often dirty inputs rather than a weak method: stockout weeks recorded as low demand, promotions absent from the data, duplicate SKU records, and returns booked to the wrong period. Genuine volatility and short history account for the rest. Fixing the data usually beats changing the model.