AI Demand Forecasting: Methods, Data and Accuracy

By Tara Minh · Updated Oct 11, 2026

AI demand forecasting is the use of machine learning models to predict how much of a product or service customers will want over a future period, using not just past sales but also prices, promotions, calendars, weather and the behavior of related products. The output is a number per item, location and period, usually with a range around it, that feeds purchasing, production, staffing and inventory decisions.

If you're new to the discipline itself, start with the non-AI basics in demand forecasting for ecommerce and demand forecasting for manufacturing. This piece assumes you already forecast something and want to know three things. When does a machine learning model actually beat a statistical one? Which outside data is worth feeding it? And how do you measure accuracy in a way that survives contact with finance?

Key Facts

• In the M5 retail competition, built on 42,840 Walmart sales series, all of the top 50 submissions beat the best statistical benchmark by more than 14%, and the winner by 22.4% (Makridakis, Spiliotis and Assimakopoulos).
• The same paper calls M5 the first M competition where all top-performing methods were pure machine learning and significantly beat every statistical benchmark and combination. In the earlier M4 competition, the two best methods were hybrids that still built on statistical foundations.
• All winning M5 submissions used external information such as prices, promotions and calendar events. Adding promotion and event data to exponential smoothing made it about 6% more accurate, and adding it to ARIMA about 13% more accurate.
• MAPE breaks down when actual demand is zero or near zero, and it penalizes over-forecasts more heavily than under-forecasts (Hyndman and Athanasopoulos). Intermittent demand needs a different yardstick.
• Accuracy and bias are separate questions. A forecast can be right on average and still wrong every single week.

Statistical vs Machine Learning Forecasting

Both families do the same job: learn patterns from history and project them forward. They differ in how much they assume and how much they can take in.

Statistical methods fit a model to one series at a time. Exponential smoothing captures level, trend and seasonality. ARIMA models describe how today's value depends on recent values and errors. Croston-type methods handle items that sell rarely. They're transparent, fast, work with two or three years of history, and a planner can explain them in a sentence. Their weakness is that they mostly look at one series and its own past.

Machine learning methods learn from many series at once and from many input columns. Gradient-boosted trees (LightGBM was the tool of choice in M5) and neural networks can combine past sales with price, promotion flags, day of week, store attributes and the sales of neighboring products, and they pick up non-linear effects, such as a discount that lifts demand only on weekends. The M5 authors describe this "cross-learning" across thousands of related series as one of the main sources of the gains (Makridakis et al.).

The honest summary from the competition record is that machine learning wins when you have a lot of related series and rich drivers, and the margin shrinks where the data is thin. Even in M5, the winning model's edge over a Croston benchmark at the product-store level, the hardest and noisiest level, was small, while the gains at higher aggregation levels were large. The lowest level is mostly randomness, and no model removes randomness.

Combinations usually beat single models

One finding has held across every M competition: averaging several forecasts beats picking one. The M5 winner was an equal-weighted combination of LightGBM models trained at different levels, and several other top teams combined models too (Makridakis et al.). For a planner, that's practical advice. Keep your statistical model, add a machine learning model, and average them before you replace anything. It's cheap insurance, and it gives you a baseline to beat.

Data and External Signals

Forecast quality is mostly a data question, and the operational data piece in this collection covers readiness in depth. For forecasting specifically, the inputs fall into four groups.

1. Demand history. Sales or shipments by item, location and day or week. Use demand, not sales: if an item was out of stock, the zero in your history isn't a lack of demand. Flag stockout periods or the model will learn that the product is unpopular.
2. Your own levers. Price, promotions, marketing campaigns, assortment changes, new store openings. These explain spikes that history alone can't.
3. Calendar and events. Holidays, paydays, school terms, local festivals, trading-day counts. The M5 authors note that models relying only on history can't account well for holidays, special events, promotions, prices and possibly the weather, and that those factors can distort the historical pattern itself (Makridakis et al.).
4. Outside signals. Weather, competitor pricing, search trends, economic indicators, supplier lead times. Add these one at a time and keep only those that improve accuracy on data the model hasn't seen.

The evidence for external variables is concrete. In M5, the exponential smoothing benchmark that used promotion and event data was about 6% more accurate than the same model without it, and the ARIMA version with external variables was about 13% more accurate than its plain counterpart. Every winning team used external information, and several found price-related features especially important (Makridakis et al.).

One trap: a signal you won't know at forecast time is useless. Next week's actual weather isn't available when you plan, only the weather forecast is, so train the model on forecasts, not on what really happened. Same for promotions: use the plan that existed when you ran the forecast.

Forecast at the level you decide at

Decisions happen at different levels. A buyer orders by SKU, a plant schedules by product family, finance plans by region. Forecasts built separately at each level won't add up. The standard fix is reconciliation, which adjusts forecasts so that SKU numbers sum to family numbers and family numbers sum to the total. Hyndman and Athanasopoulos note that reconciliation usually improves the base forecasts for almost all levels, and that it gives planners numbers that agree with each other (Hyndman and Athanasopoulos).

Accuracy Metrics Planners Should Use

No single metric works for every item. Use two or three and know what each hides. Definitions follow the standard textbook treatment (Hyndman and Athanasopoulos), where error is actual minus forecast.

WAPE is the same as MAE divided by average demand. It's often called weighted MAPE, because big items count for more, which usually matches what the business cares about. MASE, introduced by Hyndman and Koehler, scales errors by how well a simple benchmark does on the training data, so it works for any item, including those with zero sales (Hyndman and Athanasopoulos). The M5 competition scored teams on a related scaled measure, weighted RMSSE.

A worked example (illustrative)

Four weeks of demand for one product: actuals of 100, 80, 120 and 40 units, against forecasts of 110, 70, 100 and 60.

• Absolute errors are 10, 10, 20 and 20, so MAE is 15 units.
• WAPE is 60 divided by 340 total units, or about 17.6%.
• MAPE is the average of 10%, 12.5%, 16.7% and 50%, or about 22.3%. The small week dominates it.
• Bias is zero, because forecasts and actuals both total 340 units.

Same forecast, three different stories. Bias says it's fine, WAPE says you're off by about a sixth, and MAPE says it's worse because of one low week. This is why you track bias and a volume-weighted error together.

Two more habits help. Measure at the lag where you commit: if you order eight weeks ahead, judge the eight-week-ahead forecast, not the one made last Friday. And always compare against a naive benchmark (last week, or the same week last year). If your model can't beat it, it isn't earning its cost.

How to Implement AI Demand Forecasting, Step by Step

1. Pick the decision, not the model. Choose one decision the forecast drives, such as weekly replenishment for the top 500 SKUs or monthly production for six product families. Note the horizon and the level.
2. Set the baseline. Record your current forecast's WAPE and bias at that level for the last 6 to 12 months, and add a naive benchmark. This is the number to beat.
3. Prepare the history. Clean returns and duplicates, mark stockouts and one-off events, and align calendars. Fix demand history before you buy anything.
4. Add drivers in layers. Start with price, promotions and calendar flags. Add weather or other outside signals one by one, and keep each only if it improves out-of-sample accuracy.
5. Run a statistical model and a machine learning model side by side. Test on periods the models haven't seen, using rolling origins (forecast, move forward a period, forecast again) rather than a single split.
6. Combine, then compare. Average the two and score all three against the baseline. Break results out by item class, because winners differ for fast movers, slow movers and new products.
7. Reconcile across levels. Make SKU, family and total forecasts agree before they reach the planners.
8. Keep the human override, with a record. Let planners adjust for known events, and log every override with a reason. Later you can check whether overrides added accuracy or only added noise.
9. Monitor and retrain. Track bias and WAPE weekly, alert on drift, and retrain on a schedule. A model that was right last year is not guaranteed to be right after a price change or a new competitor.

The AI forecasting agent article covers how this can be packaged as an agent that refreshes forecasts and flags exceptions, and the AI in supply chain management hub shows where forecasts feed the rest of the chain.

Examples by Function

These are illustrative scenarios, not case studies.

Retail and ecommerce. A merchandising team forecasts weekly demand per SKU and store, with price, promotion flags and holidays as inputs. They use a machine learning model for the main catalog and keep a simple statistical model for slow movers.

Manufacturing. A planner forecasts demand by product family, reconciles to SKU, and uses the result in the master schedule. Forecast bias matters most here, since a steady over-forecast quietly builds inventory.

Inventory and replenishment. The forecast and its uncertainty range set safety stock. A wider range means more buffer, a narrower one means less. See AI inventory optimization for the link between the two.

Sales and operations planning. Finance, sales and operations compare a statistical baseline, a machine learning forecast and the sales team's view, then agree one number. The AI in sales and operations planning article covers that cycle.

Services and staffing. A support or field-service team forecasts ticket or job volume by hour and skill, then builds the roster. Calendar and campaign data usually explain most of the variation.

Pitfalls to Avoid

Judging by MAPE alone. It looks simple and misleads on low-volume items, since a miss of 2 units on a base of 4 is 50%. Use WAPE and bias for the portfolio, and a scaled measure for intermittent items.

Leaking the future. Training on data the model wouldn't have had at forecast time (actual weather, final promotion calendars) produces excellent test scores and poor live results.

Ignoring stockouts. Sales history during a stockout understates demand, so the model under-forecasts and the stockouts repeat.

Chasing the item-level number. The most granular level is the noisiest. Don't promise 90% accuracy at SKU-store-day. Judge the forecast at the level the decision is made.

Skipping the benchmark. Without a naive comparison you can't tell whether a complicated model is better or just different.

Overriding too often. Manual adjustments feel responsible but often add bias. Track them.

Treating the model as finished. Demand shifts. Assign an owner for monitoring, retraining and retirement.

No plan for new products. Items with no history need analog-based or attribute-based methods, and a human check, as no pattern exists to extrapolate.

What to Do Next

Pick one decision, record today's WAPE and bias, and run a machine learning model beside your current method for a full cycle. Judge it on out-of-sample data, at the horizon you order at, and by item class. If it wins in some classes and not others, deploy it where it wins and keep the simpler tool elsewhere.