[Retail](https://bfcs.ai/industry/retail/) · [Better](https://bfcs.ai/better/)

# Can AI Improve Demand Forecasting?

A large open forecasting competition on Walmart data found machine learning well ahead of the statistical methods retailers have long relied on. Look closer and most of that gain sits in chain and store totals, not on the individual shelf. Promotions and shocks such as the March 2020 rush still break forecasts, so the human override matters as much as the model.

Updated **October 2026** · 6 min read · Evidence-graded · 6 sources

## The BFCS verdict

Every article scored on the same four dimensions

| Dimension | Verdict | Why |
| --- | --- | --- |
| Better | Yes, in normal weeks · Moderate confidence | Large gains in totals; small gains on the shelf · In an open competition on Walmart data, the winning machine-learning entry beat the best statistical benchmark by 22.4% on average, but by only 3.3% at product-by-store level (our calculation from the organisers' scores). The evidence comes from one retailer's pre-pandemic data. |
| Faster | Partly · Low confidence | Automates the routine; exceptions still take people · The winning entry forecast 30,490 product-store series in a standard cloud notebook without a graphics processor. We found no independent study measuring planner hours saved, and promotions and shocks still need human judgement. |
| Cheaper | Partly · Low confidence | Waste is the prize; independent cost evidence is thin · Food wasted at the Australian retail stage is valued at billions of dollars a year, and ordering is one named lever. In our reading, peer-reviewed field evidence that AI forecasting cuts that cost is still scarce. |
| Safer | Conditional · Governance-gated | Safe when someone owns the override · A forecast trained on normal weeks cannot see a rare shock coming. Supply stays safe when shocks are detected early, human overrides are recorded and checked, and ordering limits are a deliberate decision. |

## The problem, in numbers

A retail forecast is a bet placed thousands of times a day. Order too little and the shelf is empty; order too much and fresh food goes in the bin. Australia's 2018-19 food waste baseline put wholesale and retail waste at 527,000 tonnes, part of a national total of 7.6 million tonnes, about 70% of it edible. (Source 6) The same study put the cost at the retail stage at about $4.4 billion a year, and it names ordering processes and replenishment sizes among the levers for cutting it. (Source 6)

Then there are the weeks no model has seen. In March 2020 Australian retail turnover rose 8.5% in a month, seasonally adjusted, and supermarket and grocery turnover rose 23.0%. Flour sales rose 147.9% on February and canned and dry mix soup 180.6%. (Source 1) A forecast built on the previous few years had little to go on.

### The evidence, in four numbers

22.4% · lower error than the best statistical benchmark · M5 winning entry, average across 12 levels of Walmart sales; our calculation from the organisers' published scores

3.3% · the same margin at product-by-store level · the level at which store orders are placed; same source and calculation

23.0% · one-month rise in Australian supermarket and grocery turnover · March 2020, seasonally adjusted (ABS)

527,000 · tonnes of food wasted in Australian wholesale and retail · 2018-19 baseline (FIAL National Food Waste Strategy Feasibility Study, 2021)

## What improves, and why

Traditional retail forecasting looks at one product's sales history at a time, often with a manual uplift added for promotions. (Source 5) Machine learning changes two things. It learns across thousands of products and stores at once, so a slow seller can borrow patterns from similar items. And it takes in the signals that move demand: price, promotion, day of the week, holidays and local events. (Source 5)

That is where the accuracy gains come from. A review of retail forecasting research reports that simple time-series methods do well when nothing is on promotion, but that models with more inputs improve accuracy substantially once promotions start. (Source 5) One study it cites, covering 988 products, found that methods using price information beat single-series methods by typically more than 10%, with larger gains in promotional periods. (Source 5) The review also notes that forecasters adjust forecasts by hand, and that in one non-food chain, which adjusted some 20% of its forecasts, the adjustments added no value. (Source 5) A model that learns the promotion effect from data can replace some of that guesswork.

## What the evidence actually says

The clearest public test we found is the M5 competition, run on Walmart sales data. Teams forecast 42,840 sales series, from total chain sales down to 30,490 single products in single stores, for 3,049 products across ten US stores, 28 days ahead. (Source 3), (Source 4) The winning entry, built from gradient-boosted tree models and run on a standard cloud notebook without a graphics processor, scored 22.4% better on the competition's error measure than the best statistical benchmark, by our calculation from the organisers' published scores. (Source 4) All of the top 50 entries beat that benchmark on average. (Source 4)

The headline hides a gradient, and the gradient is the story for anyone placing orders.

### How far the M5 winner beat the best statistical benchmark, by level (Source 4)

Total chain sales · 53.3%

By store · 31.1%

By store and department · 21.4%

By product · 4.6%

By product and store · 3.3%

Percentage reduction in scaled error (WRMSSE) for the winning entry against bottom-up exponential smoothing, our calculation from the organisers' published scores. The gain shrinks as the forecast gets closer to the shelf.

At total chain level the winner's error was 53.3% lower than the benchmark. For a single product in a single store, the level at which replenishment happens, it was 3.3% lower. (Source 4) The competition guide describes these series as intermittent, with sporadic sales and lots of zeros. (Source 4) No model can say which day the one jar sells. Machine learning makes the plan for the business much sharper; it makes the shelf only a little sharper.

> The forecast that improves most is the one for the whole chain. The forecast that fills the shelf improves least, and it is the one that decides whether food is thrown away.

The competition also had calm data: its sales history ended in June 2016. (Source 4) Shocks look different. In a study of 42 stores of a large North American grocery chain, weekly transactions in panic-buying categories doubled or even tripled in mid-March 2020, and the authors describe panic buying as creating false demand signals that are hard to predict because the event is rare. (Source 2) Their answer was not a better forecast but a model that detects the anomaly early, so the retailer can respond with measures such as purchase limits. (Source 2)

## Where it falls short

-   **Small gains where orders are placed.** At product-by-store level the M5 winner was only 3.3% better than a well-chosen statistical method. (Source 4)
    
-   **Unstable weeks are under-researched.** The retail forecasting review we drew on says the effects of intermittency and market instability have yet to be fully explored. (Source 5)
    
-   **Accuracy is not stock.** In the chain whose overrides added no value, many adjustments were made for stocking and service-level reasons rather than accuracy. (Source 5) In our reading, an error score does not measure waste. Choose the measure that matches the cost.
    
-   **Shocks are outside the training data.** Panic buying is rare by definition, (Source 2) and a model trained on normal weeks has no pattern to match it against.
    

## Where governance decides the outcome

Forecasting looks like a back-office tool, but in a supermarket it decides whether essential goods are on the shelf. Three rules keep it safe. First, decide who may override the model and record every override with a reason, then check whether overrides improve results; in one chain studied, they added no value. (Source 5) Second, watch for shocks separately from forecasting. A model that flags abnormal buying early gives people time to act, but the response, such as a purchase limit, is a decision for accountable managers, not the model. (Source 2) Third, measure what matters. Track waste, stockouts and availability alongside forecast error, because in our reading a model can score well on error and still fill the bin. In our reading, peer-reviewed evidence that AI forecasting reduces retail waste in the field is still scarce, so treat vendor waste claims as targets to verify against your own figures.

## How to think about adoption

Start where the evidence is strongest: chain, store and department forecasts that drive labour, distribution and buying. The M5 gains were largest there. For store ordering, expect modest accuracy gains and judge them by waste and availability, not error scores alone. Feed the model price, promotion and event data, because that is where it beats the old methods. Keep a person responsible for promotions and for unusual weeks, and test the model against a simple benchmark every season. If it cannot beat bottom-up exponential smoothing on your own data, it is not yet earning its place.

## How we read this evidence

BFCS.ai does not fabricate figures. The M5 margins are our own calculations from the organisers' published scores, and they describe one retailer's pre-pandemic data. We found no peer-reviewed field study measuring the waste or cost effect of AI forecasting in Australian retail; where that is the case we say so.

## Sources

1.  Australian Bureau of Statistics, "Retail Trade, Australia, March 2020", released 6 May 2020. [abs.gov.au](https://www.abs.gov.au/statistics/industry/retail-and-wholesale-trade/retail-trade-australia/mar-2020)
    
    **What we took from it:** Seasonally adjusted turnover rose 8.5% in March 2020; food retailing rose 24.1% and supermarkets and grocery stores 23.0%. Supermarket scanner data in original terms show non-perishable goods up 39.0% on February and perishables 21.6%; a chart in the release shows flour up 147.9% and canned and dry mix soup up 180.6%, March on February.
    
2.  Adulyasak, Y., Benomar, O., Chaouachi, A., Cohen, M. C. & Khern-am-nuai, W., "Using AI to detect panic buying and improve products distribution amid pandemic", AI & Society, published online 15 April 2023 (peer-reviewed journal), doi:10.1007/s00146-023-01654-9. [pmc.ncbi.nlm.nih.gov](https://pmc.ncbi.nlm.nih.gov/articles/PMC10105357/)
    
    **What we took from it:** Point-of-sale data from 42 stores of a large North American grocery chain, January 2018 to May 2020, for toilet paper, canned soup and household cleaners. Weekly transactions in week 11 of 2020 doubled or tripled relative to weeks 1 to 8; panic buying causes short-term shocks and false demand signals and cannot be easily predicted because it is a rare event. The 56.74% improvement in access to essential products is a simulation on historical data (a one-item limit on toilet paper, 12 to 19 March, in one representative store), not a measured result.
    
3.  Makridakis, S., Spiliotis, E. & Assimakopoulos, V., "M5 accuracy competition: Results, findings, and conclusions", International Journal of Forecasting 38(4), 2022, pp. 1346-1364 (peer-reviewed), doi:10.1016/j.ijforecast.2021.11.013. [econpapers.repec.org](https://econpapers.repec.org/RePEc:eee:intfor:v:38:y:2022:i:4:p:1346-1364)
    
    **What we took from it:** The abstract describes forecasting 42,840 hierarchical unit-sales series of Walmart, with 30,490 point forecasts required at the lowest level. The publisher's full text refused automated access; this article relies on the abstract and on the organisers' published data (note 4).
    
4.  M5 organisers' official repository (Mcompetitions/M5-methods on GitHub): [github.com](https://github.com/Mcompetitions/M5-methods)
    
    **What we took from it:** "Scores and Ranks.xlsx", "M5-Competitors-Guide.pdf" and the first-placed team's README. The guide describes 3,049 products sold in ten stores across California, Texas and Wisconsin, a history from 29 January 2011 to 19 June 2016, a 28-day forecast horizon, explanatory variables including prices, promotions and special events, and series characterised by intermittency with lots of zeros. The winning entry (WRMSSE 0.520 averaged across 12 levels) used LightGBM models in a Kaggle notebook on CPU; the best benchmark on the average score, bottom-up exponential smoothing (ES\_bu), scored 0.671, and all top 50 entries scored below it. At some single levels another benchmark scored better than ES\_bu; the margins below are all against ES\_bu. Margins of 22.4% overall, 53.3% at total level, 31.1% by store, 21.4% by store and department, 4.6% by product and 3.3% by product and store are BFCS.ai calculations from the published scores.
    
5.  Fildes, R., Ma, S. & Kolassa, S., "Retail forecasting: research and practice", Lancaster University Management School Working Paper 2018:04, MPRA Paper No. 89356, October 2019 revision (working-paper version; the peer-reviewed version is International Journal of Forecasting 38(4), 2022, pp. 1283-1318, doi:10.1016/j.ijforecast.2019.06.004, whose full text is subscription-only, and the accepted-manuscript link in Lancaster's repository served an unrelated paper when checked). [mpra.ub.uni-muenchen.de](https://mpra.ub.uni-muenchen.de/89356/)
    
    **What we took from it:** The journal version's abstract says adequate evidence on machine learning methods has not yet accumulated. Second-hand findings it reports: simple time-series methods perform well without promotions, but models with more inputs improve accuracy substantially during them (Gur Ali et al., 2009); multivariate methods using price information beat univariate methods with gains typically above 10% in an evaluation of 988 SKUs, and typically higher in promotional periods (Ramos and Fildes, 2017); a non-food retail chain adjusted some 20% of its forecasts and the adjustments added no value, possibly because many were motivated by stocking and service-level considerations rather than accuracy (Fildes et al., 2009). Many retailers use a base-times-lift approach, a simple baseline plus promotional adjustments that include managers' judgement; data pooling can help at lower levels of the product hierarchy. The authors say intermittency and the effects of market instability have yet to be fully explored, and wide-ranging evidence of the benefits of machine learning is needed before believing the hype.
    
6.  Food Innovation Australia Ltd (FIAL), "The National Food Waste Strategy Feasibility Study: Final Report", 2021 (delivered by Edge Environment with Lifecycle Strategies, WRAP and 3Keel). [openaccess.city.ac.uk](https://openaccess.city.ac.uk/id/eprint/27351/1/FIAL%20NFWS%20Feasibility%20Study%20Report_FINAL.pdf)
    
    **What we took from it:** 2018-19 baseline: 7.6 million tonnes of food waste a year, about 70% edible, at an estimated cost of AUD$36.6bn; wholesale-retail waste of 527 thousand tonnes (Table 3) and a cost of $4,402 million for the retail stage (Table 4, hotspot analysis). Its "retail operational efficiency" intervention covers digital pricing and stock management, ordering processes and replenishment sizes.
    

Putting this to work

## Working through this in your own organisation?

BFCS.ai is published by Inhouse CX. We help small and mid-sized organisations decide where AI earns its place, and put the governance in place that makes it safe to use.

[Talk to Inhouse CX →](https://inhouse.cx/engage/)

Continue the thread

- [Can AI Speed Up Fulfilment?](https://bfcs.ai/articles/can-ai-speed-up-fulfilment/): More in Retail · RETAIL · FASTER · 8 min · Read →
- [Can AI Reduce Shrinkage Costs?](https://bfcs.ai/articles/can-ai-reduce-shrinkage-costs/): More in Retail · RETAIL · CHEAPER · 7 min · Read →
- [Can AI Predict Maintenance Needs?](https://bfcs.ai/articles/can-ai-predict-maintenance-needs/): Also Better · MANUFACTURING · BETTER · 7 min · Read →

---

Canonical page: https://bfcs.ai/articles/can-ai-improve-demand-forecasting/
This is the Markdown version of the page, generated from its HTML at build time.
Site guide for agents: https://bfcs.ai/llms.txt · Every page: https://bfcs.ai/sitemap.xml