AI and the New Demand Forecasting Stack

There's a quote I've seen attributed to various people over the years (statisticians, economists, consultants who really should know better) that goes something like: "the best forecast is one built on the best data." Technically true, but it glosses over the harder problem, which is that the data most readily available to retailers is historical, and history is a worse oracle than it looks when conditions change quickly.

Most demand forecasting systems have been built around a simple assumption: the future will broadly resemble the past. Buy what sold. Adjust for growth. Account for seasonality. It's a reasonable approximation for stable, high-velocity SKUs in stable conditions, and it works right up until it doesn't.

What the pandemic actually proved

The pandemic is the obvious example, and it's been cited so often in supply chain discussions that it risks becoming wallpaper. But it's worth being specific about what failed and why.

When consumer behaviour shifted in March 2020 (essentials surged, discretionary collapsed, categories like toilet paper and pasta became subjects of genuine national anxiety) models trained on two or three years of pre-pandemic data had nothing useful to say. There was no historical pattern that resembled a global pandemic. The exponential smoothing and ARIMA models that underpin most legacy forecasting tools are structurally unable to handle structural breaks; they smooth and average, which is exactly the wrong thing to do when something unprecedented is happening.

Retailers relying solely on historical models saw forecast accuracy collapse. The question that followed is what's driven most of the serious investment in this space since: what do you actually need to forecast demand well during disruption?

The signal problem

The core limitation of historical-only forecasting isn't the maths; it's what data goes in. Historical sales capture what happened; they say very little about why, and almost nothing about what's about to happen. Modern AI forecasting systems are built around a much wider signal set:

Internal signals: real-time point-of-sale data (not just end-of-day batches), inventory positions across the network, the promotional and pricing calendar, website and app engagement, customer service query patterns that often foreshadow demand shifts before they show at the till.

External signals: hyperlocal weather (not regional; a city-level temperature forecast can drive measurable differences in clothing demand across a single distribution area), social media trend analysis, macroeconomic indicators, competitor activity, event calendars, news and media sentiment.

The architectural shift is from a model that asks "what did we sell?" to one that asks "what's driving demand right now, and what's about to change?" That's a different kind of system, and it requires a different kind of data infrastructure.

The technical stack

Modern forecasting systems tend to combine several approaches rather than picking one:

Foundation models are pre-trained on large retail datasets (sometimes millions of SKU-location-time combinations) and then fine-tuned for specific product categories or trading patterns. They're better at generalising across novel conditions than models trained only on a single retailer's own history.

Ensemble methods combine multiple forecasting approaches (gradient boosting, neural networks, classical time-series models) and weight them based on recent performance. No single method is reliably best; the ensemble is more robust.

Causal inference is the approach that gets least attention but arguably matters most for promotions and pricing. Rather than learning that "sales go up in week 12," a causal model tries to identify what actually causes the uplift, and crucially, what would happen if you changed the promotion mechanic. This is an active area of academic research (there's a body of work applying double machine learning to retail pricing that's starting to reach production) and it matters enormously for any retailer running a complex promotional calendar.

Reinforcement learning is used in the most sophisticated implementations to continuously update model parameters based on outcomes. Each replenishment cycle provides feedback; the system adjusts. There's a European grocery chain documented in academic literature that processes 425,000 SKU-location combinations overnight using this kind of architecture, with results suggesting €218 million in annual inventory reduction while maintaining 98.7% product availability. Those are exceptional numbers; most deployments are less dramatic. But they illustrate what the architecture is capable of.

McKinsey's research on AI-driven forecasting suggests the technology can reduce forecast errors by 20–50% compared to traditional approaches, with knock-on reductions in lost sales and product unavailability of up to 65%. Those ranges are wide because outcomes depend heavily on starting accuracy, data quality, and how well the implementation is done.

What UK retailers are actually doing

The JLP case is probably the most clearly documented example in British retail. John Lewis Partnership has deployed 18 production-grade AI models covering forecasting, pricing, labour planning, and checkout optimisation. The business value attributed to that portfolio is over £40 million. Barry Hostead, their Director of Data Management, made the point that before their Dataiku deployment, getting a model into production took months; now it takes a couple of weeks. That velocity shift matters because a forecasting model that can't be updated when conditions change isn't much better than the historical model it replaced.

One of their concrete outputs is the Bakery Optimisation Tool: before it existed, in-store partners were downloading spreadsheets and manually estimating what to bake. Now the model generates store-specific recommendations based on historical sales and real-time signals. It's a mundane example (baked goods forecasting) but that's rather the point. The value of these systems usually comes from hundreds of small, unglamorous improvements rather than one transformative deployment.

M&S has taken a similar path in food, deploying Relex Solutions' AI platform across its food estate. Relex is a Finnish supply chain software company with a particularly strong track record in fresh and perishable categories, where the cost of getting forecasting wrong (wastage, stockouts) is highest. The deployment is one of the larger known implementations of AI-driven supply chain planning in UK grocery.

Tesco has used AI models to simulate promotional demand effects, enabling preemptive replenishment rather than reactive firefighting. That's directionally important for any large supermarket: promotional planning is where forecasting fails most visibly and most expensively.

What the implementation actually involves

There are a few things that don't get said enough about AI forecasting deployments.

Data quality is the real project. The most sophisticated model cannot overcome poor input data. Implementing AI forecasting usually surfaces existing data problems (inconsistent SKU definitions, gaps in POS history, promotional flags that were never properly captured) that have been quietly poisoning every previous model. Fixing those problems is often more work than the model itself, and it pays dividends beyond forecasting.

High-impact SKUs first. Not every product needs a cutting-edge forecast. The economics of improving accuracy on a stable, slow-moving, low-margin SKU are different from improving accuracy on a high-velocity seasonal product where a stockout loses real revenue. Concentrating implementation effort on the SKUs where accuracy matters most is how sensible retailers approach this.

Implementation timelines are longer than vendors suggest. One analysis of 12 retail organisations transitioning to AI-driven forecasting platforms found an average full-scale implementation time of 8.7 months. Vendors will often quote shorter timelines for initial go-live; achieving the actual forecasting gains requires the data integration work, the model tuning, and the feedback loops to be operational. It takes time.

Humans still need to be in the loop. AI forecasting is good at patterns. It's less good at contextual knowledge: the buyer who knows a supplier is about to have production problems, the merchandiser who understands why a product sells differently in different regions, the planner who remembers that last year's promotional result was distorted by a distribution failure. Systems that make it easy for experienced people to intervene and override matter as much as the models themselves.

Where this is heading

The direction that serious vendors are moving towards is what some are calling generative assortment planning: systems that don't just forecast demand for existing products but identify gaps: demand signals that suggest a product category should exist but doesn't. A few early adopters are experimenting with this, mostly larger retailers with the data infrastructure to support it. It's at the pilot stage for most; I'd be cautious about anyone claiming production scale on this yet.

The nearer-term evolution is agentic architectures: multi-agent systems where demand sensing, forecasting, and replenishment decisions happen continuously rather than in overnight batch cycles. The shift from periodic planning to continuous inference is real and it's coming. Whether it represents a fundamental change or a better version of what we already have is a question the industry hasn't fully answered.

For most retailers, though, the opportunity still lies in the fundamentals: integrating more signal sources, improving model accuracy on high-impact SKUs, building proper feedback loops, and making sure forecasts actually translate into better buying and allocation decisions. The technology is no longer the constraint. Getting the data infrastructure right usually is.

---

For more on the operational layer that sits downstream of forecasting, see Warehouse Robots and the Autonomous Future. The overlap between demand signals from personalisation engines and forecasting inputs is covered in AI Personalisation in 2026: State of Play.