Your stores are overstocked on sweaters because inventory arrived before the cold snap hit. Your shelves are empty on the pasta aisle because a one-day promo moved weeks of volume in a day, and the order cycle had no way to see it coming. Your DC is sitting on slow-moving SKUs that tie up cash while a cluster of stores marks down what should have been profitable margin. And your order cycle — still built around last quarter's average sales — has no way to see any of it.
The Forecast You Actually Use
Demand forecasting has been done in retail for decades. The problem is that most of it happens at the wrong grain. You get a regional forecast. You get a monthly forecast. Both are stale the moment the weather shifts, a competitor marks down, or a local event drives foot traffic nobody modeled.
The forecasts that move the order book are built at SKU-store-day granularity. Not "how many units of this category will we sell next month," but "how many units of this exact SKU will move in this store on this day, given the weather forecast, the promo calendar for that week, and how this assortment actually turns in this store cluster versus the chain average."
That specificity matters because it surfaces the leakage. Last-quarter-average forecasting buys safety stock to cover regional variance and calls it done. SKU-store-day forecasting shows you that one cluster runs faster on a particular assortment while another runs slower — and lets you cut safety stock on the slow movers, unlocking cash, while keeping on-shelf availability high on the fast ones. That is the mechanism behind the headline number: an 18% drop in stockouts without burying working capital in safety stock.
But the number alone does not ship a model. What ships a model is explainability.
Why Merchants Hand the Order Book to the Model
A traditional demand-planning tool creates a forecast. A buyer prints it. Maybe the number seems too high or too low. They override it. They have no window into how the model thought about it, so they override on intuition.
The forecasts that move the order book inside retail work backwards. Every demand number traces back to its drivers. Not just a quantity, but the components behind it: the base demand for this SKU and store, the lift the running promotion is expected to add, the dampening or boost the weather forecast implies, the contribution of local events, and the cushion that historical variance justifies. The buyer sees the breakdown, not just the total.
A category manager can look at that, disagree with the weather effect — maybe rain actually drives traffic in their cluster, not away from it — and override the weather contribution without breaking the rest of the model. They can also see that the promo contribution is confident enough to act on as-is. Explainability moves the forecast from something to second-guess into something to work with.
The design choice that made this work — tracing each demand component back to its driver — is what let buyers hand the order book over. Not because the forecast was perfect, but because they could see what it was thinking and correct it surgically instead of throwing the whole number out.
The Assortment Velocity You Already Miss
One of the hardest problems in retail is the planogram reset. You plan an assortment for your stores. It is supposed to move at a given rate. But store clusters are different — one moves faster, another slower. If you reorder on chain-average velocity, the fast cluster runs short before the next reset and the slow cluster builds overstock.
The signal that catches this — which facings actually turn in which clusters — lives in the POS data you are already generating. But almost no planning system uses it, because the calculation is awkward: you have to know which facings are actually loaded, which ones moved, and what the cluster signature is. It requires fusing the planogram data (what should be loaded), the POS data (what actually moved) and the store-cluster metadata. Most systems never join the three.
The models that ship do exactly that. They read the actual facings in each store from the planogram, match them to what moved in the POS, and flag the slow movers — assortments that are not turning at the rate the category manager expected. The payoff is a flag to the reset team that says this facing configuration is underperforming in a given cluster, so consider reducing facings or rotating a different product through that slot.
That is not a revolutionary forecast. But it surfaces leakage that phantom-inventory and overstock problems had been hiding — assortment gaps that quietly leak sales between resets, per cluster rather than chain-average. The point of reading shelf velocity at cluster level is that it makes those silent losses visible while there is still a reset to act on.
Promo Lift: Forecasting the Incremental Spend
A buyer runs a promotion. The units spike. The forecast now has to answer: how much of that spike was new demand it would not otherwise have had, and how much was cannibalization — demand it would have gotten the following week, just pulled forward?
Overestimate incremental lift and you buy too much stock, then get caught holding overstock that has to be marked down when the promo ends. Underestimate it and you stock out and leave money on the table. The job of a good promotion forecast is to land between the two.
The models that ship treat lift causally: they look at the historical promo calendar, compare demand in the weeks before and after each promotion, and learn the true incremental lift rather than just reading the spike. They also learn the carry-forward effect — when a promo pulls demand forward, it accounts for the trough that follows. So the replenishment order placed during the promo is high, but the order placed the week after is lower, because demand has already been partially satisfied.
The math is worth getting right because the stakes cut both ways: overstock on clearance costs margin, stockouts cost customer trust and sales. Forecasting true lift instead of the raw spike means promo spend actually pays back. And it means end-of-life markdowns can be timed more precisely — clearing stock without torching margin, rather than being caught holding inventory long after the promo ended.
- 18%
- Stockout reduction
- One signal
- Demand, replenishment, promo lift
- ~4-6 weeks
- Assess phase to ranked roadmap
One Order System, One Reason
The fatal flaw in most AI-first retail systems is that they sit beside the actual workflow.
A demand-sensing model lives in a new tool. A buyer logs into the ERP where they have always placed orders, then logs into the new AI tool, compares the two, and overrides the AI's number back in the ERP. The two systems drift. The model learns on what was actually ordered, not what was forecast. The whole thing decays.
The systems that move the order book solve this differently: the demand-sensing output writes directly into the ERP and POS systems the buying team already uses. Not a new dashboard. Not a competing forecast. The RetailDemand instrument plugs into the replenishment workflow itself, so when a buyer logs in to place the next order, the recommended quantity is already there — with the drivers shown.
They can override it, and most do on a fraction of SKUs. But the baseline is the model's number. That tight feedback loop — where the model sees what was actually ordered, what the POS revealed, and why — is what keeps retraining clean. The model learns at the order level, not filtered through a parallel human workflow. That integration depth is also what makes the signal sticky: once it is wired into the daily order cycle, nobody wants to rip it out.
Every demand number traces back to its drivers — promo, weather, seasonality, cannibalization — so a category manager can see why the model wants more units and override one store without breaking the chain.
Cold Starts and Seasonal Cliffs
Demand-sensing models are brittle against certain failure modes. A new SKU launches and there is no history, so the model has no baseline to forecast from. The promo calendar for next quarter looks unusually packed, and the last time it looked like that was years ago in a different competitive environment — the model sees the pattern but does not have enough data to be confident.
The production systems that ship account for cold starts by falling back to human input on brand-new SKUs, letting buyers seed an initial forecast the model tunes from there. They track promo calendars separately, so when a novel promo mix appears the model flags its own uncertainty and an analyst sanity-checks the recommendation before it feeds the order.
The other failure mode — the post-promo cliff — is subtler. A major promo ends, demand collapses because the following weeks have no promotional activity and the customer base is temporarily sated. A model trained on the promo period alone will overstock for the weeks after. The production systems catch this by monitoring for the demand trough that follows promotional activity and retraining to account for the carry-forward effect.
Where to Start
Start with the hardest question: where is availability actually bleeding revenue? Not "where would more inventory have been nice," but "where did we run out and lose sales, or where are we carrying overstock that will have to be marked down?"
The Assess phase maps that leakage by SKU, store cluster and season, and separates true stockouts from phantom inventory and forecast bias. You pull your POS history, your inventory records, your out-of-stock flags if you have them — many retailers do not track them precisely — and you ask: for which categories would a short-horizon, explainable demand forecast actually change the order? You also audit your promo calendar and planogram data. How clean is it? Can you connect a promo date to a POS spike? Can you match a planogram to actual shelf facings?
The output is a ranked list of high-velocity categories where demand sensing pays back fastest, mapped against your POS, ERP and replenishment data quality. The assessment typically takes 4–6 weeks. The first candidates are usually the ones with the most volatile demand — apparel, seasonal goods, anything touching weather or local events — because that is where the signal improves the most.
You then pilot on one or two high-velocity categories, building the demand-sensing model on SKU-store-day data, wiring it into your ERP's replenishment workflow, and validating that the recommendation actually reaches the order cycle without being overridden into uselessness. Models account for promo lift, cannibalization and lead-time variability from the start. If the pilot moves stockout and overstock metrics, you harden it toward chain-wide production.
The work is not fast. The technical bar is high because you are fusing heterogeneous data — POS, promo calendars, weather, planogram metadata — and the mesh has to be clean. But the payoff is concrete: fewer empty shelves, less dead working capital, and a forecast that a buyer actually believes.
“SKU-store-day forecasting shows you which assortments run faster in some clusters and slower in others — unlocking cash on the slow movers while keeping on-shelf availability high on the fast ones.”
Get in touch
Put RealAI’s applied-AI team on your hardest data problem.
We help enterprises move from pilots to production — sovereign models, governed data, and agents you can audit. Start with a value-first assessment.