The Demand Forecasting Accuracy Report 2026: What Happens When AI Forecasts Compete Against Incumbents

1. Accuracy against the incumbent forecast

Absolute accuracy figures travel badly. A 25% MAPEMAPE The unweighted mean absolute percentage error. It treats every item-period equally. is a strong result for a spare parts distributor with thousands of slow-moving items and a weak one for a beverage brand's top ten SKUs. Product mix, demand intermittency, and the level you forecast at (SKU, SKU-location, category) all move the number before the forecasting method has any influence on it. The benchmark that holds up is improvement over the forecast a company already runs, on that company's own data and in that company's own metric.

On that benchmark, the AI forecast won every comparison in the study. In each benchmarked deployment, the AI forecast produced less error than the incumbent, whether that incumbent was the baseline out of the ERP or planning system, a consensus number out of S&OP, or a planner's spreadsheet. One deployment ran the comparison month by month for a full year. Twelve months, twelve wins.

Error reduction ranged from 14% to 56%, with an average of 32% across the 15 deployments.

At the top of the range, 56% came from Rimports, a home fragrance distributor selling through Amazon and other channels, and the portfolio with the highest starting error in the set. Its manual forecasts had been producing both overstock and stockouts, and its Amazon rankings and storage costs were paying for it. Weighted MAPE fell from 146% to 64%. Weighted MAPE (WMAPEWMAPE Divides total absolute error by total actual demand across the portfolio, so high-volume items count for more than long-tail ones.) divides total absolute error by total actual demand across the portfolio, so high-volume items count for more than long-tail ones. A starting point above 100% means the incumbent's total absolute error was larger than the demand it was trying to predict, which is what intermittent, low-volume demand does to a forecast. At 64%, more than half the incumbent's error was gone, concentrated on the items that carry the volume.

The floor of the range belongs to the deployment that ran the month-by-month comparison. A health and wellness manufacturer's weighted MAPE went from 40% with its internal forecast to 34.3% with the AI forecast, a 14% reduction, and the AI forecast came in lower in every one of the twelve months across its most significant SKU segments. The smallest improvement in the study was also the most consistent one.

Between those two, Kenvue reduced MAPE by 37% in its first year, with models built on orders, shipments, and point-of-sale signals and a human-in-the-loop process that combines the AI's predictions with planner expertise. MAPE, the unweighted mean absolute percentage error, treats every item-period equally. A 37% cut on a large, established portfolio is a broad improvement, not a handful of outliers fixed.

Some companies measure against their own target rather than a textbook metric, and the benchmark works the same way. A pool equipment manufacturer had been forecasting from sales targets rather than demand signals, with an internal tool that produced recurring overstock and understock across SKUs. Measured against the company's own accuracy target, automated forecast accuracy went from 25% to 50%. The modeling effort went to the highest-revenue SKUs first, and the sales-led planning team was never asked to change how it worked.

3. Planner effort and the touchless threshold

Forecast accuracy gets the headlines. Planner time is where most teams feel the change first.

Most planning teams know the shape of the problem before anyone measures it. The planning tool produces a baseline. The planners don't trust it, so they adjust it, SKU by SKU. Next month brings the same adjustments to the same SKUs. The manual work never shrinks, because the forecast underneath it never improves.

Manual planning effort fell by 70% on average across the 15 deployments. At Dorman, across more than 100,000 SKUs, manual data cleansing alone dropped by about 70%: the work of preparing history before any forecast can run. At Mars, 75% of forecast volume qualified as touchless: forecasts that can go into the plan without a planner's edit, leaving planners to review only the remaining quarter, the items where their judgment changes the outcome. As Kristen Daihes, Mars's Vice President of Global Supply Chain, put it, the planning team "can already 'not touch' ~75% of the volume."

We call that second figure the touchless threshold. Once three-quarters of forecast volume can pass without a human edit, demand planning stops being a review of everything and becomes exception management. The planner's week changes shape. Hours move from touching every SKU to investigating the ones the model flags: new launches, promotions, accounts with a known change coming.

Outside product demand, the pattern holds. CAA Club Group, Canada's largest automobile club, had been building roadside call volume forecasts by hand, a full week of work per set. Hourly forecasts across nearly 600 micro-regions and five service types now refresh twice a week, daily during winter storms, and run staffing and dispatch at every club facility. Time spent generating forecasts fell 30%, with no data science headcount added.

Ask any planner what they would do with 70% of their forecast prep time back. The answer is rarely "review more SKUs."

5. What actually drives demand

Of all the findings here, this is the one we expect to be quoted most, because it contradicts what most planning teams believe about their own business.

Nanit, a maker of smart baby monitors, began its project with more than 20 variables its team was confident drove demand. The model narrowed them to the four that carried predictive signal. Forecasting got more accurate and simpler at the same time: 80% forecast accuracy across SKUs and channels, from a company that had been running homegrown models in Python and Excel.

Which four matters less than the pattern. They were specific to that business, and anyone who copied them would be forecasting someone else's demand. The pattern is the transferable part. Many of the drivers a planning team tracks are correlated with each other, or with a demand history that already contains their effect, so they add noise rather than signal. A model that tests every candidate against held-out demand can say which ones earn their place. A planner working from experience can only say which ones feel important.

In practice, the consequence shows up in two places. Teams stop sourcing and maintaining data feeds for variables that don't move the forecast. And the forecast becomes easier to explain to the people who have to trust it, because four drivers fit on one slide.

Where this leaves the incumbent forecast

Companies rarely go looking for a forecasting model. They go looking for a forecast they can run the business on, usually after a period of making purchasing, production, and safety stock decisions against a number the planning team had quietly stopped trusting. The benchmark is how you find out whether that distrust is justified, and by how much.

Running one is a small project. It takes two to three years of SKU-level order or shipment history, the incumbent forecast for the same periods, and an agreed error metric. Every accuracy figure in this study came from one, and every one showed the same direction of result. The cost of not running it is another planning cycle spent making the same manual fixes to a forecast that has never been tested against an alternative.

Next year's edition will add the deployments benchmarked over the coming year. If your team runs its own comparison, we would like to see the result, whichever way it goes.

Limitations