Demand Forecasting Error Statistics for Manufacturers (2026)
Most forecast-error numbers online are repeated without a source. We traced each statistic to its original study, labelled how strong the evidence is, and added worked examples that show how forecast error becomes stockouts and excess stock.
Search for demand forecasting error statistics and you will find the same handful of numbers repeated across dozens of vendor sites, rarely with the original study behind them. Some are real and useful. Others lose the caveats their authors attached.
This guide collects the figures that hold up. Each one is traced to its source, with the year, the sample and a note on how much weight it can bear. It also shows how forecast error turns into stockouts and excess stock, using a worked example you can rerun on your own numbers.
It is written for operations and planning heads at Indian manufacturers who still forecast in Excel.
Demand forecasting error statistics at a glance
Seven demand forecasting error statistics are worth knowing. The table shows where each comes from and how much it can prove.
| What was measured | Result | Source | Kind of evidence |
|---|---|---|---|
| Individual forecasts that failed to beat a naive "same as last period" forecast | Around 50% | Morlidge, Foresight, 2014 | Company forecast data |
| Planners' manual adjustments that improved accuracy | On average at 3 of 4 companies; small and upward adjustments often did not | Fildes and others, International Journal of Forecasting, 2009 | More than 60,000 forecasts from four supply-chain companies |
| Indian supply chain leaders naming poor forecast accuracy a top planning challenge | 66%, the most common of seven options | KPMG in India, 2018 | Survey of 76 leaders |
| Same survey: exploring planning and forecasting tools | 44%, the most common technology option | KPMG in India, 2018 | Survey of 76 leaders |
| Error reduction from AI-driven forecasting | 20 to 50 percent, with up to 65 percent fewer lost sales and cases of product unavailability | McKinsey, 2022 | A statement of what is possible; the text gives no sample |
| Large organisations using AI-based forecasting by 2030 | 70% | Gartner, September 2025 | A prediction, not measured adoption |
| Distortion of demand signals up the supply chain | Order variance can exceed sales variance and grows at each step upstream | Lee, Padmanabhan and Whang, Management Science, 1997 | Peer-reviewed research; no single percentage |
Read the last column before you quote a number. A survey of 76 leaders and a study of more than 60,000 forecasts are different kinds of evidence, and a prediction is not a measurement.
How often forecasts fail: what company data shows
Around half of forecasts lose to a naive guess
Steve Morlidge's 2014 article in Foresight presents evidence that many companies fail to beat a naive forecast, one that simply repeats the last period's actual. His finding is that around 50% of individual forecasts fail to meet that benchmark.
It is also the cheapest test you can run. If "same as last month" misses by less than your forecast does, the model, the spreadsheet or the review meeting is adding error.
Manual overrides help less often than planners expect
Fildes, Goodwin, Lawrence and Nikolopoulos analysed more than 60,000 forecasts and their outcomes from four supply-chain companies, in a paper published in the International Journal of Forecasting in 2009. Planners' adjustments to the system forecast improved accuracy on average at three of the four companies.
The detail is less flattering. Larger adjustments tended to help more, smaller ones often reduced accuracy, and upward adjustments were much less likely to help than downward ones, which the authors read as optimism bias.
So before sales adds another 10% for the festive season, check whether last year's adjustment improved accuracy.
In India, forecast accuracy tops the planning problem list
KPMG in India's 2018 survey of 76 supply chain leaders in automotive, auto ancillary, pharmaceuticals, consumer goods and durables, and engineering found that 66% named poor forecast accuracy as a top planning challenge. It was the most common of seven options, ahead of lack of end-to-end supply chain visibility at 51%.
The same survey found 44% of respondents exploring planning and forecasting tools, the most common technology option. The sample is small and the data is from 2018, so read it as a direction, not a benchmark.
How demand forecasting errors cost manufacturers revenue
Forecast error costs money in two directions. Forecast too low and you run short: lost orders, expedited freight, overtime and a customer who calls your competitor. Forecast too high and cash sits in stock that can age, expire or be written off.
We could not find a sourced rupee figure for Indian manufacturers, so this section shows you how to calculate your own. The mechanics are simple.
The bullwhip effect: your demand signal is already distorted
In a 1997 paper in Management Science, Hau Lee, V. Padmanabhan and Seungjin Whang showed that orders can vary more than the sales behind them, and that the distortion grows at each step up the supply chain. They named four causes: demand signal processing, the rationing game, order batching and price variations.
Their Sloan Management Review article gives the classic case. Consumers used Pampers diapers at a steady rate and retail sales fluctuated, but not excessively, while distributors' orders varied more, and the orders Procter & Gamble placed with its own suppliers varied more again.
If you sell through distributors, dealers or OEMs, your incoming orders are not end demand. Where you can, forecast from their sales or stock data as well as from their purchase orders.
From forecast error to safety stock: a worked example
Safety stock is the buffer that covers forecast error during the replenishment lead time. NC State's Supply Chain Resource Cooperative describes the safety factor, z, as the number of standard deviations above average lead-time demand that you stock to, with 1.65 for a 95% cycle service level. When you replenish from a forecast, the spread to cover is the forecast error.
Take an item that sells 1,000 units a month on average. Assume monthly errors are independent and roughly normal, so the spread over a lead time of L months grows with the square root of L. Safety stock is then 1.65 × the monthly error × √L.
The table shows the safety stock needed for 95% service.
| Monthly forecast error (standard deviation) | 1-month lead time | 4-month lead time |
|---|---|---|
| 150 units (15% of average demand) | 248 units | 495 units |
| 300 units (30% of average demand) | 495 units | 990 units |
Double the forecast error and the buffer doubles. Quadruple the lead time, as with imported raw material, and it doubles again. At ₹800 a unit, the gap between the best and worst cells is 742 units, or about ₹5.9 lakh of cash in one item.
Now look at service. Say you set the 248-unit buffer when the error was 150 units, and the error later rises to 300. That buffer now covers 0.83 standard deviations, a service level of about 80% instead of 95%, so a stockout in one cycle in twenty becomes one in five.
These are illustrations on assumed numbers, not statistics. Rerun the arithmetic on your own items, and see our guide to forecasting volatile demand for how to set safety stock when lead times and demand both move.
What AI forecasting can realistically change
A widely quoted number in this area comes from McKinsey's February 2022 article by Jorge Amar, Sohrab Rahimi, Zachary Surak and Nicolai von Bismarck. It says AI-driven forecasting in supply chain management can reduce errors by 20 to 50 percent, and that this can mean up to 65 percent fewer lost sales and cases of product unavailability.
The same article puts warehousing cost reductions at 5 to 10 percent and administration cost reductions at 25 to 40 percent. Read the wording closely: "can" and "up to" describe what is possible, the article frames the gain against traditional spreadsheet-based methods, and its text does not say how many companies the ranges come from.
Gartner's September 2025 prediction points the same way: 70% of large organisations will use AI-based forecasting by 2030. It also names what slows adoption, which is no clear vision among planning leaders, incomplete or hard-to-reach data, and staff used to traditional practice.
If your sales history sits in Tally and spreadsheets, cleaning it up comes first. Then measure any tool against the naive forecast, not against last year's spreadsheet. Our guide to the top demand forecasting tools for manufacturers covers what to look for.
How to measure forecast error without fooling yourself
MAPE breaks on the items that matter
Mean absolute percentage error is the most quoted accuracy figure. The forecasting textbook Forecasting: Principles and Practice lists its weaknesses: it is infinite or undefined when actual demand is zero, extreme when demand is close to zero, and it penalises over-forecasts more heavily than under-forecasts.
Spare parts, made-to-order jobs and new SKUs often have months with no demand, which is exactly where MAPE fails. Hyndman and Koehler's 2006 paper found that many common accuracy measures break in situations like these, and proposed the mean absolute scaled error, which compares your error with a naive forecast.
A three-item example: MAPE, WAPE and bias
Here are three items with last month's actuals and forecasts.
| Item | Actual | Forecast | Error (units) | Error (%) |
|---|---|---|---|---|
| A, a fast mover | 1,000 | 900 | 100 | 10% |
| B | 50 | 25 | 25 | 50% |
| C | 10 | 20 | 10 | 100% |
MAPE averages the percentage errors and gives 53%. WAPE, also written WMAPE, divides the total units missed by total demand, 135 ÷ 1,060, and gives 13%. Bias, forecast minus actual, is −115 units or −11%, so you under-forecast overall.
Same forecasts, three verdicts. MAPE lets two tiny items dominate, while WAPE follows the units you actually make and sell, and bias shows the direction, so track both. Our guide to demand forecasting methods goes deeper on MAPE, WMAPE and bias.
Totals hide errors, and other people's benchmarks mislead
Add items together and the errors cancel. If item A is 100 units over and item B is 100 units under, the family total is exactly right while both items are wrong. Measure error at the level where you buy or make.
Published accuracy benchmarks are no better a guide. Stephen Kolassa's analysis in Foresight concluded that published accuracy surveys cannot be trusted as benchmarks, and advised companies to look to internal benchmarks instead. The benchmark that matters is your own naive forecast and your own error last year.
Run a forecast-error audit this week
You need twelve months of history and a spreadsheet. Pull it from Tally or your ERP and follow these steps:
- Pull twelve months of forecast and actual sales or consumption for your top items.
- Add a naive column that uses last month's actual as this month's forecast.
- Calculate WAPE and bias for your forecast and for the naive one. If the naive forecast wins, fix the process before you buy anything.
- Flag the months where someone overrode the forecast, and check whether those overrides improved accuracy.
- Convert the error into safety stock and rupees for your ten largest items, as in the example above.
If your history is stuck in Tally or spreadsheets, our Tally to ERP guide and spreadsheet to ERP guide cover how to move it into one system.
If the audit shows the forecast is fine but purchases still arrive late, the gap sits between the forecast and the purchase order. That is what our demand planning software is for: it nets the forecast against stock and open orders through the bill of materials and drafts the purchase for your approval. If you only need the forecast number, see our demand forecasting software page, and for S&OP across sales, operations and finance, our supply chain planning software.
Demand forecasting error: common questions
What is a good forecast accuracy for a manufacturer?
There is no valid universal number, because published accuracy surveys are not reliable benchmarks. A useful target is to beat the naive forecast every month, then to cut WAPE at the level where you buy and make.
What is the average forecast error?
We could not find a trustworthy cross-company average. In Morlidge's data, around half of individual forecasts failed to beat a naive forecast, so start by checking yours against that baseline.
How do demand forecasting errors cost manufacturers revenue?
Under-forecasting leads to stockouts, lost orders, expedited freight and overtime, while over-forecasting ties up cash in stock that can expire or be written off. Safety stock rises in line with forecast error: in our example, doubling the error doubled the buffer. For shelf-life products, see our guides to demand planning for pharma and demand planning for FMCG.
Does AI reduce forecast error?
McKinsey says AI-driven forecasting can cut error by 20 to 50 percent, but that is a statement of potential rather than a measured average. Test any tool on your own history against a naive forecast before you rely on the number.
Should I use MAPE or WAPE?
Use WAPE for a range with many small and large items, because MAPE is undefined at zero demand and lets small items dominate. Track bias alongside it.
Conclusion
The demand forecasting error statistics agree on three points: many forecasts lose to a naive guess, manual adjustments need testing, and forecast error flows straight into safety stock and cash. Measure yours against the naive forecast first. To see how Likwid turns a forecast into purchase orders on your own data, talk to our team, or read about our demand planning software.