Retail demand forecasting is the practice of estimating how much revenue a specific location will generate by measuring the demand available in its trade area and the share of that demand a particular brand can realistically capture. It is a question about a place, asked before the store exists, answered with signals rather than sales history.
One clarification before anything else, because the term carries two meanings and they belong to different departments. This article covers the location sense: forecasting the revenue of a site you are considering leasing. The other sense is inventory demand planning— forecasting how many units of which SKUs will sell so that replenishment and supply chain can be scheduled. That is a time-series discipline built on item-level history, it sits with supply chain rather than real estate, and almost nothing here applies to it.
A location demand forecast is a weighted combination of a few signals — population, daytime population, income fit, competition, co-tenancy, mobility data, accessibility — calibrated against the stores you already operate. The signals are not the hard part; the weighting is, and the right weighting is specific to your category and brand. Foot traffic is one input, not the answer: a busy corner full of the wrong customer forecasts badly. And the honest measure of a forecast is the width of the band around it, governed by how many comparable stores the model learned from.
What Is Retail Demand Forecasting?
Retail demand forecasting estimates the revenue a specific site will produce, using the characteristics of the surrounding trade area rather than any sales history for that site. It answers two linked questions: how much relevant spending exists within a realistic distance of this door, and what share of it this brand will take given who else is competing for it. The output is a revenue figure with a confidence band, not a traffic count.
The mechanics are less mysterious than vendor language suggests. A model is fitted against your existing stores: for each open unit, the signals around it are measured against what that unit actually sells. The fit produces a set of weights — a statement that in your business income fit matters this much and competitive density that much — and those weights are then applied to the signals around a candidate site. The model families that do this, from analogues to gravity models to machine learning, differ in how they derive the weights, not in what they are trying to do; we compare them in retail sales forecasting methods compared.
Demand, capture, and the gap between them
Keep the two halves separate. Demand is a property of the geography: the people, their spending, when they are present. Capture is a property of your brand in that geography: whether they know you, prefer you, or default to the competitor two doors down. A trade area can be rich in demand and poor in capture — which is the technical description of a site that looks obvious on a map and disappoints after opening.
Which Signals Actually Predict Store Revenue?
The signals that recur in location revenue models are trade-area population, daytime population, income and demographic fit to your existing customer, competitive density, co-tenancy quality, mobility or visit data, and drive-time accessibility. Which of those carries the most predictive weight is category- and brand-specific, and anyone who tells you there is a universal answer is selling something. A breakfast-weighted drive-thru lives on commuter flow and access; a destination specialty retailer may be driven almost entirely by income fit and the quality of its neighbours.
A signal earns its place by varying meaningfully across the sites you are actually choosing between. A variable that reads identically at every candidate on your shortlist contributes nothing, however important it is in principle.
| Signal | What it is really measuring | Where it misleads |
|---|---|---|
| Trade-area population | How many people live inside a realistic catchment | Radius rings count people a river or a highway stops from ever arriving. Use drive time. |
| Daytime population | Who is physically present during your trading hours | Decisive for weekday lunch, near-irrelevant for a weekend-weighted format. |
| Income fit | Match to the income band your current stores actually sell into | Higher is not better. Fit is directional both ways, and median hides the distribution. |
| Competitive density | How much of the available demand is already spoken for | In some categories competitors cluster because demand is there; density can read positive. |
| Co-tenancy quality | Whether neighbours generate trips by the same customer | A full center is not the same as a relevant center. Count overlap, not occupancy. |
| Mobility / visit data | Observed movement to the parcel and its neighbours | Counts devices, not wallets. Says a place is busy, not that it is busy with your buyer. |
| Drive-time accessibility | Friction between the customer and the door | Modelled travel times flatter sites whose real problem is the left turn at 5pm. |
Foot Traffic Is an Input, Not an Answer
Concretely: imagine two candidates. The first sits on a transit-adjacent block with very high measured footfall, dominated by commuters moving through at pace. The second is a quieter suburban parcel with half the counted visits — but those visits are households whose income band and composition match your best-performing stores almost exactly.
A model that ranks on traffic puts the first site ahead, and it will be wrong, because the quantity it ranked on is not the quantity that produces revenue. Visit counts measure presence. Revenue requires presence, relevance, intent and conversion. Foot traffic is genuinely useful — it is the best available evidence of where people actually go, as opposed to where census geography suggests they live — but it belongs inside the model as one weighted input, validated against what your own stores sell. We go deeper on its uses and limits in retail foot traffic analytics.
A useful test when a vendor presents a site: ask what the forecast would be if the traffic number were cut in half, and what it would be if the customer-fit score were cut in half. If only one of those materially moves the answer, you are looking at a traffic ranking wearing a forecast’s clothes.
Demand signals have also become more observable than they were. Proximity-driven “near me” search is a large share of how customers find physical stores, and it is a measurable expression of intent in a geography rather than mere presence in it. It is one more signal with its own biases, and it belongs in the same place as the others: weighted, not worshipped.
How Much Does the Weighting Change the Answer?
More than most people expect, which is the point of the tool below. Move the weight between signals and watch both the modelled index and the ranked contribution list reorder; then shrink the calibration set and watch the confidence band swallow the difference between a strong site and a weak one. To be explicit: this is an illustration of how weighting behaves, built on invented demonstration values. It is not Locate’s model and it does not forecast revenue for any real site.
An illustration of how signal weighting moves a forecast. Not Locate’s model, and not a prediction for your site.
Read this first.The signal scores below belong to one invented demonstration parcel — they are made-up numbers chosen to show the mechanism, not measurements of any real location. The curve that turns a weighted index into dollars is an illustrative assumption, not a fitted coefficient from any model. Nothing here predicts revenue for a site you are considering.
An endcap beside a full-line grocer with steady weekly trips.
Index points contributed = normalised weight × this parcel’s illustrative read. The order changes as you move the sliders.
Shortlist spread is a proxy: the dollar gap between a candidate that reads as well as this parcel’s strongest signal and one that reads as poorly as its weakest. If the band is wider than that gap, the model is not ranking anything.
With only 34 comparable stores, the band is wider than the difference between your best and worst candidate — the model cannot separate them. Anything inside $1,236,889 to $1,812,070 is statistically the same site. Either add comparables, tighten the market-similarity screen, or make the call on judgment and say so out loud.
Illustration only. The parcel’s signal reads, the index-to-revenue curve, the band formula and the cannibalization haircut are all invented for demonstration — they are not Locate’s model, not industry benchmarks, and not a forecast of what any store would sell. Which signals carry real predictive weight is specific to your category, your brand and your customer, and can only be established against your own open stores.
How Accurate Can a New-Store Forecast Be?
Accuracy depends on two things you can check before commissioning the work: how many comparable stores the model is calibrated on, and how similar the new market is to the markets those stores sit in. Those two variables govern the width of the confidence band, and the band — not the point estimate — tells you whether the forecast can do its job. We deliberately quote no accuracy percentage here: any such figure, stated without its calibration set and market context, is marketing rather than measurement.
The practical standard is simpler than a statistic: a forecast is good enough when its band is narrower than the gap between the candidates you are choosing between. If your best and worst shortlisted sites are separated by less than the model’s own uncertainty, the model has not ranked them — it has produced numbers in an order that looks like a ranking.
- Calibration set size. Ask how many open units the model learned from, and how many of those are genuinely similar in format, vintage and market type.
- Market similarity. A model fitted entirely on suburban units is extrapolating when pointed at a dense urban street, and extrapolation is where models fail quietly.
- Held-out testing. Was the model scored on stores it had never seen, or on the same ones it was fitted to? Only the first number means anything.
- Published band. A forecast delivered as a single number with no interval is withholding the most decision-relevant part of the result.
Accuracy is also something you can measure yourself after the fact, which is the only honest long-run check on any provider. Our method for doing that is in grading site selection forecast accuracy.
Why Do Good Models Fail on New Sites?
Because explaining history and predicting a new site are different problems. A model can fit your existing portfolio beautifully by leaning on variables that describe where you have already chosen to open rather than what makes a store work, and that relationship holds right up until you point it at a site outside the pattern. The fit looks like knowledge. It is often just a description of your past real estate committee.
The four recurring failures
- Extrapolation.The candidate sits outside the range of the calibration data on some dimension — far denser, far wealthier, a different format — and the fitted relationship simply does not extend there.
- Selection bias in the training set.Your open stores are the sites you approved, not a random sample of sites. The model learns your committee’s taste along with the economics.
- Unmodelled transfer.In a market you already serve, part of the new store’s forecast is revenue moving across the street from your own unit. A model trained on first-in-market stores will not know to deduct it.
- Proxy drift.A signal that worked as a proxy — visit counts standing in for customer presence, median income standing in for fit — stops working when the new site breaks the correlation that made it a proxy in the first place.
The defences are unglamorous. Flag extrapolation rather than letting the model answer confidently outside its range. Hold out stores and score against them. Model transfer as its own term. And re-fit as the portfolio grows, because the weights that described twenty stores are not the weights that describe eighty. The sequencing of all this inside a live site decision is laid out in new store sales forecasting.
Buy the Weighting, Not the Number
When you evaluate a forecasting provider, the number they hand you is the least informative part of the deliverable. Ask which signals are in the model, what weight each carries for your category, how many of your stores the weights were fitted on, how the band was derived, and what happens when a candidate falls outside the training range. A provider who can answer those five questions is doing analysis; one who leads with an accuracy claim and will not show the band is doing sales.
This is the work Locate does, and we do it next to the brokerage rather than in a separate report: the weighting is fitted to your stores, the band travels with the number, and the site that wins the model is the site we go and negotiate. If you want the analysis and the execution in one place, talk to us.
Common Questions
- What is retail demand forecasting?
- Retail demand forecasting, in the site selection sense, is the practice of estimating how much revenue a specific location will generate by measuring the demand available in its trade area and the share of that demand your brand can realistically capture. It combines population and daytime population, income fit to your actual customer, competitive density, co-tenancy, observed mobility data and accessibility into a model calibrated on the performance of your existing stores. Note that the same phrase is also used for inventory demand planning, which forecasts units of product rather than revenue at a place, and is a separate discipline.
- Which signals actually predict store revenue?
- The signals that consistently appear in location revenue models are trade-area population, daytime population, income and demographic fit to your existing customer, competitive density, co-tenancy quality, mobility or visit data, and drive-time accessibility. Which of those carries the most predictive weight is category- and brand-specific and cannot be stated as a universal rule: a breakfast-weighted drive-thru lives on commuter accessibility and daytime population, while a destination specialty retailer may be driven almost entirely by income fit and co-tenancy. The only way to know your weighting is to fit it against your own open stores.
- How accurate can a new-store forecast be?
- Accuracy depends mostly on two things you can measure before you commission the work: how many comparable stores the model is calibrated on, and how similar the new market is to the markets those stores sit in. A model with a large calibration set of genuinely similar units can produce a band tight enough to rank candidate sites against one another, while a model built on a handful of stores in dissimilar markets produces a band so wide it cannot separate a good site from a mediocre one. Ask for the confidence interval and the calibration set size rather than a single headline accuracy figure.
- Why do good models fail on new sites?
- Because fitting history and predicting a new site are different problems. A model can explain your existing portfolio very well by leaning on variables that merely describe where you have already chosen to open, and then fail on a site outside that pattern, where the relationship it learned does not hold. The common failure modes are extrapolating beyond the range of the calibration data, ignoring transfer from your own nearby stores, and treating visit counts as demand when the visitors are not your customer.
- Is retail demand forecasting the same as inventory demand planning?
- No. Inventory demand planning forecasts how many units of which SKUs will sell over a future period so that supply chain and replenishment can be scheduled, and it is usually built on time-series history at the item level. Retail demand forecasting in the real estate sense forecasts the revenue a location will produce before that location exists, using geographic and demographic signals rather than sales history for that site. Both are called demand forecasting, and the two use different data, different models and different teams.