Measuring Forecast Accuracy: How Clean Is Your Crystal Ball?
Forecast accuracy measures how close your forecast landed to what actually happened, using an error metric such as MAD, MAPE, WAPE (also sold as WMAPE) or RMSE. There is no single correct metric. Use MAD when you are tracking one item in real units, WAPE when you need one honest number across a mixed-volume portfolio, and RMSE when large misses cost disproportionately more than small ones. Whichever metric you pick, measure it at the lag your organization actually plans against. A number computed at the wrong lag can be mathematically correct and still tell you nothing useful.
What does forecast accuracy actually measure?
Forecast accuracy measures the size of the gap between your forecast and what happened, with the direction of that gap thrown away on purpose. A forecast that missed high by 200 units and one that missed low by 200 units score identically on accuracy, because accuracy only asks how far off we were, never which way. That second question has its own name, forecast bias, and its own metric, and the two get confused constantly. The 2016 version of this very post is a case in point: it went out under a measuring-forecasting-accuracy title and was, start to finish, a bias-formula explainer. That mix-up is common across the industry, not just something we got wrong once. Accuracy and bias are both essential and neither substitutes for the other. We cover bias formulas, the sign-convention traps, and how teams game them in a separate piece; this one stays focused on accuracy.
What are the main ways to measure forecast accuracy?
Four metrics cover nearly every situation you will run into, and each one has a specific condition under which it quietly stops telling the truth.
| Metric | Formula | Good for | Where it breaks |
|---|---|---|---|
| MAD / MAE (mean absolute deviation or error) | Average of |Actual − Forecast|, in real units | Single-item tracking in units your team already thinks in; feeds safety stock math | Cannot compare across items of different volume; a MAD of 50 means nothing without knowing the item’s typical demand |
| MAPE (mean absolute percentage error) | Average of |Actual − Forecast| ÷ Actual, as a percentage | Single, reasonably stable item over time; intuitive to non-technical stakeholders | Explodes or divides by zero near low or zero actuals; punishes over-forecasting harder than under-forecasting, which quietly trains planners to forecast low (Hyndman & Koehler, 2006, warn against MAPE under exactly these conditions) |
| WAPE / WMAPE | Sum of |Actual − Forecast| ÷ Sum of Actual | Portfolio or category-level reporting; survives mixed volumes and near-zero SKUs | Top sellers dominate the number, so a good WAPE can sit on top of a genuinely chaotic tail (full teardown in our MAPE vs WMAPE vs WAPE piece) |
| RMSE (root mean square error) | Square root of the average of (Actual − Forecast)² | Comparing statistical models against each other; situations where a big miss costs far more than several small ones | Expressed in squared-then-rooted units, not a clean percentage, so it is hard to explain in an S&OP room; oversensitive to a single outlier period (Hyndman & Koehler, 2006) |
The Institute of Business Forecasting treats MAPE as the industry default and bias as its required companion metric, which matches what we see in practice: most organizations under-invest in the second half of that pairing.
Which accuracy metric should you actually use?
It depends on the decision the number is meant to support, not on which metric is easiest to calculate. A safety-stock adjustment for one SKU wants MAD, in the same units the inventory system already uses. An executive KPI covering a category or a whole business wants WAPE, because it survives the long tail without a spreadsheet full of exceptions. A data science team choosing between two forecasting models wants RMSE, because it correctly punishes the model that occasionally misses badly over the one that is mediocre but consistent. Reach for MAPE only when you are watching a single, fairly stable item over time and everyone in the room already understands its limits.
The chart below walks through this as a decision tree: what decision is the number feeding, then which metric survives that decision’s specific failure mode.
What counts as a good forecast accuracy score?
There is not a universal number, and the guides that hand you one (70% is fine, 80% is good) are quietly hiding three variables that matter more than the number itself: the lag the score is measured at, the level it is aggregated to, and how predictable the underlying demand actually is. A WAPE of 85% on a portfolio of laundry detergent and airline seats is a different achievement than 85% on impulse-buy items at the checkout till, even though the number reads identically on a dashboard.
In our experience, forecast accuracy is only truly critical for a just-in-time business, where there is no buffer to absorb the miss. For everyone else carrying inventory as a buffer, accuracy is an optimization lever, not a pass or fail test, and how hard to push that lever depends on what a miss actually costs you. A business selling a cheap, substitutable commodity can live with a weekly accuracy in the 50% range and barely notice. A service business, where capacity itself is the real buffer, might find even 80% is not good enough. Context decides, not a benchmark table: how much inventory-carrying headroom you have, and what a stockout actually costs you. Run a temperature-controlled warehouse where electricity is one of the biggest line items, or carry thin inventory cover, and 80% or better becomes the number you need to sweat those assets properly.
We have seen this most clearly running an ice-cream factory. Pre-building production for the summer season is an expensive bet, and going out of stock on strawberry two weeks into the season can wreak havoc on the plant and put the full year’s financial target out of reach. That is a business where a forecast miss has nowhere to hide. A slower-moving, lower-stakes category can absorb the same percentage miss and nobody downstream ever notices.
How do you make an accuracy number actionable?
Two adjustments turn a single accuracy figure from a scoreboard into a diagnostic tool. First, measure at more than one lag, matched to how far ahead your supply chain actually has to commit (a four-week manufacturing lead time and a fifty-two-week capacity decision need separate accuracy readings, because they are really answering separate questions). Second, segment by ABC class or predictability before you average anything, because one company-wide number always hides more than it reveals. We go deep on both cuts, with a worked example, in the MAPE vs WMAPE vs WAPE piece. The short version: the lag and the segmentation matter more than which metric you pick.
What mistakes wreck a forecast accuracy program?
The same handful of mistakes show up across most stalled programs. Measuring at only one lag, usually whichever one makes the number look best that month. Publishing one company-wide figure and letting it hide a rotten tail behind a handful of strong sellers. Chasing accuracy while ignoring bias entirely, so a forecast can hit a respectable WAPE while running consistently high or low and quietly filling the warehouse. And treating the metric as the destination rather than the diagnostic, so the team spends its energy defending last month’s number instead of fixing the process that produced it.
What target should you actually set?
Picking the right metric solves only half the problem. The other half is what target you set once you are measuring the right thing, and that is where most programs quietly go wrong.
The first mistake is treating the accuracy number itself as the point. It is not. The question that actually moves a program forward is why the forecast was not 100%, not how far off it was. That is where the real problems unravel, and in our experience most of them are internally driven: a promotion that never made it back into the model, a sales number that got sandbagged, a new-product ramp nobody flagged upstream. Chase the why and the accuracy follows. Chase the accuracy number alone and you get planners gaming the metric instead of fixing the cause behind it.
The second mistake is setting the target on an absolute basis rather than a loss-reduction basis. We would rather see a target framed as reduce the error by X% than hit 85%, because a loss-reduction target keeps paying off even after the easy wins are gone, while an absolute target invites people to stop the moment they clear the bar. Better still, tie the target to the business metric its improvement actually drives. Are you after better asset utilization? A better return on promotional spend? Faster inventory turns? Something else? Forecast accuracy should be moving that number. If it is not, the metric itself is the bigger problem, and that is worth resolving before anything else on this list.
Two forecasting books worth your shelf
We are recommending these based on their scope and reputation in the field, not a first-hand read we are claiming here. Both build the underlying math well enough to help you pick the right metric for your own situation, which is exactly what this article has been arguing for. Charles Chase’s Demand-Driven Forecasting covers the statistical methods and the organizational process around them together, which most forecasting books do not attempt. Hyndman and Athanasopoulos’s Forecasting: Principles and Practice is the textbook the accuracy-metric research in this article draws from, and it is also free to read online if you want to check before you buy.
