New-store sales forecasting is the practice of estimating what a specific, not-yet-open retail location will sell in its first stabilized year, using some combination of your existing stores’ performance, trade-area data, competitive geography, and expert judgment. Everything else in this article is a comparison of the ways people do that, and of what each way costs you when it is wrong.
The uncomfortable truth in this field is that the method matters less than the fit between the method and your situation. A machine-learning ensemble in the hands of a 6-unit brand is a liability. A per-capita estimate at a 400-unit chain is negligence. Most arguments about forecasting methods are really arguments about portfolio size, dressed up as arguments about statistics.
There are six practical methods for forecasting a new store’s sales: broker judgment, analog matching, ratio / per-capita, regression, gravity models, and machine-learning ensembles. The right one is mostly a function of how many comparable stores you already operate. Below roughly eight stores, judgment wins. From about fifteen, analogs win. Past thirty to forty with movement data, models win. Accuracy claims are meaningless unless the claimant tells you the sample size and the holdout method behind them.
What Are the Methods of Sales Forecasting?
Six methods account for essentially all new-store forecasting in retail: expert judgment, analog matching, ratio or per-capita models, regression, gravity models, and machine-learning ensembles. They differ in what data they consume, how many existing stores they require, and what they are good at explaining. They are not competing religions; most serious programs run two or three and treat the spread between them as information.
The broker or operator estimate
A seasoned broker walks the site, reads the co-tenancy, watches the turn lane at 5:15pm, and gives you a number. This is not a fallback method, it is a real one, and it consistently outperforms a badly specified model. The reason is that the broker is implicitly weighting variables no dataset carries: the landlord’s reputation, what happened to the last three tenants in that bay, whether the grocery anchor is renewing. Its failure mode is that judgment does not scale, does not document itself, and carries the estimator’s biases forward invisibly.
Analog matching
Analog matching finds the existing stores most similar to your candidate site and uses their actual revenue as the forecast base. It is the workhorse of multi-unit retail because the comparison is to your own P&L, not to a generic category model, which makes it easy to defend to a real-estate committee. Its failure mode is the thin analog pool: when only three of your stores resemble the candidate, you are averaging noise, and when you enter a genuinely new market type you have no analogs at all.
Ratio and per-capita models
These take category spend per household, multiply by trade-area households, and apply an assumed capture rate. They are fast, cheap, and appropriate for market sizing, deciding whether a metro supports four stores or eleven. They are the wrong tool for choosing between two addresses a mile apart, because they cannot distinguish sites within the same trade area. Their failure mode is a capture-rate assumption that nobody can defend, quietly carrying the whole forecast.
Regression
Regression fits an equation relating your stores’ sales to site variables and applies it to the candidate. Its great virtue is transparency: every variable has a coefficient you can interrogate. Its failure mode is overfitting on small portfolios and multicollinearity among variables that all measure roughly the same thing, producing coefficients that flip sign when you add one store.
Gravity models
Gravity models, of which the Huff model is the best known, allocate a trade area’s demand across competing destinations based on their size and the travel cost to reach them. They are the strongest available tool for two specific questions: how much share a new site will take, and how much of it will come from your own existing stores. We cover the mechanics in our guide to the Huff gravity model in retail. The failure mode is the distance-decay parameter: guess it rather than calibrate it against real movement data and the model produces confident, wrong share numbers.
Machine-learning ensembles
An ensemble blends several model types, typically gradient-boosted trees plus a spatial component, across a wide feature set including mobility data, card spend, and site attributes. When it is well fed and honestly validated, it is the best-performing option available. Its failure mode is that it fails silently: with too few stores it memorizes the portfolio, and because the output is a single number with no visible reasoning, nobody in the room can tell that it has. That is the case for understanding how AI revenue forecasts actually work before you buy one.
The Six Methods, Compared
The most useful column in this table is the third one. Method selection in new-store forecasting is primarily a store-count question, and most forecasting mistakes are made by brands using a method their portfolio is too small to support.
| Method | Data needed | Stores needed | Strength | Failure mode |
|---|---|---|---|---|
| Broker / operator estimate | None beyond local knowledge and a site visit | 0 | Catches what no dataset carries; instant | Unscalable, undocumented, carries hidden bias |
| Analog matching | Store-level sales plus trade-area profiles | ~15–20 comparable | Grounded in your own revenue; committee-ready | Thin analog pool; breaks on new market types |
| Ratio / per-capita | Category spend and trade-area population | 0 | Fast market sizing at near-zero cost | Indefensible capture rate; blind within a trade area |
| Regression | Consistent sales plus site and demographic variables | ~30+ | Auditable coefficients you can argue with | Overfits small samples; unstable collinear variables |
| Gravity / Huff | Competitor locations and sizes, travel times | Few, if calibrated externally | Best available read on share and cannibalization | Guessed distance-decay parameter; weak on absolute dollars |
| ML ensemble | All of the above plus mobility and spend signals | ~40+ | Highest ceiling when well fed and validated | Fails silently; opaque to the people signing the lease |
Which Method Fits Your Situation?
Set the sliders and toggles below to your actual situation, not your aspirational one. Be honest about analyst capacity in particular: it is the single input that most often disqualifies the method a team wants to use.
Describe your situation. The tool ranks the six common new-store forecasting methods by how well they fit it.
Mobility data lets you calibrate the distance-decay curve instead of guessing it, which is what makes gravity models work.
Mobility data lets you calibrate the distance-decay curve instead of guessing it, which is what makes gravity models work.
Data needed: Competitor locations and sizes, travel times, and trade-area geography
Realistic error range: Strong on share-of-market and cannibalization, weaker on absolute dollars
Your store count supports real analogs, and analogs are the most defensible forecast you can put in front of a committee.
Data needed: Sales for your existing stores plus trade-area profiles for each
Realistic error range: Tightens steadily as your analog pool grows past roughly 15-20 comparable stores
Cheap and fast for first-pass market sizing, which is the job it is actually good at.
Data needed: Category spend per household and a trade-area population count
Realistic error range: Coarse: fine for market sizing, weak for a single address
Useful as a sanity check on whatever quantitative method you run.
Data needed: None beyond local market knowledge and a site visit
Realistic error range: Wide, but rarely catastrophic in a market the broker knows well
Too few stores: a regression fitted here will look precise and be wrong.
Data needed: Sales for every store plus consistent site, demographic and competitive variables
Realistic error range: Good once you clear roughly 30 stores; unstable and overfit below that
An ensemble needs more history than you have; it will memorize your stores rather than learn from them.
Data needed: All of the above, plus mobility, card spend or similar signals, maintained continuously
Realistic error range: Best-in-class when well fed and validated; opaque and fragile when not
How to read this: scores are a relative fit ranking from your inputs, not a prediction of accuracy. Error ranges are qualitative bands, because real accuracy depends on your category and on how it was measured.
Which Sales Forecasting Method Is Most Accurate?
No method is most accurate in the abstract. Accuracy is a property of a method applied to a specific portfolio, in a specific category, measured a specific way, and changing any of those three changes the answer. The most accurate method for you is the most sophisticated one your data can actually support, which for most brands is less sophisticated than they would like.
This is why published accuracy claims deserve suspicion. A vendor claiming a tight average error is making a statement that means nothing until you know three things.
- Was it measured on held-out stores? A model scored on the stores it was trained on will always look excellent and tell you nothing.
- Measured against which year? First-year sales and stabilized sales are different targets, and the gap between them is the ramp curve.
- Averaged across what? A mean absolute error across 500 similar suburban stores is a much easier number to post than one across 30 heterogeneous urban sites.
- What happened to the misses? A method with modest average error but occasional catastrophic misses is worse for a lease decision than one with slightly higher average error and a tight tail.
The defensible response is to grade your own forecasts against actuals on a fixed schedule. Our guide to grading site-selection forecast accuracy lays out how to do that without letting the scorecard drift into self-congratulation.
Retail real estate is an asymmetric bet. A store that beats forecast by 20% is pleasant; a store that misses by 40% on a ten-year lease is a multi-year drag you cannot easily exit. Choose the method that minimizes catastrophic misses, not the one with the prettiest average.
How Do You Forecast Sales for a New Store With No History?
With no stores of your own, you build a range rather than a number, from three independent sources: a per-capita model from trade-area population and published category spend, whatever you can learn about comparable operators or franchisees in similar trade areas, and the judgment of a broker who has leased your format in that submarket. Where the three agree, you have a usable planning number. Where they diverge, the spread is your honest uncertainty.
The practical discipline for a first or second location is not forecasting precision, it is downside structuring. Size the space, the buildout, and the term so the pessimistic end of your range is survivable. Our guide to new-store sales forecasting goes deeper on building that range, and opening a second location covers the specific case where your one existing store is your only analog.
Demand-side search data is a genuinely useful third input here, because it is available before you open anything. Semrush’s research puts US “near me” searches at roughly 7.1 million per month and rising 29% year over year, with city-level volume available across its keyword database. That will not give you a revenue number, but it will tell you whether the demand you are assuming actually exists in that geography.
The Ensemble Reality of Modern Practice
Sophisticated teams do not pick one method. They run analogs as the primary estimate, a gravity model to check cannibalization and share, and a regression or ensemble as a challenger, then look at the spread. Tight agreement is a green light. Wide disagreement is the actual finding: it means the site is unlike your portfolio in some way the methods are each handling differently, and that is when you send someone to walk it.
The broker estimate belongs in that ensemble, not outside it. If your model says $2.4M and the broker who has leased four stores on that corridor says $1.6M, the useful next question is what the broker knows that the model does not. Often it is something specific and checkable: a competitor lease signed three weeks ago, a road project starting in spring, a landlord who does not maintain the center.
This is the logic behind how Locate works, running quantitative revenue forecasts and brokerage execution under one roof, so the model output and the market read get reconciled before a lease is signed rather than after. If you want that reconciliation on a live site, talk to our team.
Pick the Method Your Portfolio Can Support
The most common forecasting failure in multi-unit retail is not choosing a bad method. It is choosing a method that requires more stores, more data, or more analyst capacity than the brand actually has, and then trusting its output because it came out of a model. Match the method to the portfolio, run more than one, treat disagreement as information, and grade yourself against actuals every year. That discipline is worth more than any single technique on this list.
Common Questions
- What are the methods of sales forecasting for a new retail store?
- Six methods cover almost all of new-store forecasting practice: an experienced broker or operator's estimate, analog matching against your own comparable stores, ratio or per-capita models based on category spend and trade-area population, regression models fitted to your portfolio, gravity models such as Huff that allocate demand by distance and competitor pull, and machine-learning ensembles that blend many signals. Most mature programs run two or three of them and compare the answers rather than trusting one.
- Which sales forecasting method is most accurate?
- There is no method that is most accurate in the abstract, because accuracy depends on how many comparable stores you have, how clean your data is, and what you are measuring accuracy against. With fewer than about eight stores, an experienced operator's judgment usually beats any model you could fit. Between roughly 15 and 30 stores, analog matching tends to win. Past 30 to 40 stores with mobility data, regression and ensembles pull ahead. Any vendor quoting a single accuracy figure without naming the sample and the holdout method is quoting marketing, not measurement.
- What is analog forecasting?
- Analog forecasting projects a new store's sales by finding existing stores whose trade areas, competition, format, and customer profile most closely resemble the candidate site, then using their actual sales as the basis for the estimate. It is the most defensible method for most multi-unit brands because the comparison is to your own real revenue rather than to a generic model. Its weakness is that it only works if you have enough genuinely comparable stores, and it struggles when you enter a market unlike anything in your portfolio.
- How do you forecast sales for a new store with no history?
- With no store history of your own, you borrow someone else's structure: build a per-capita or category-spend model from trade-area population and published category spend, triangulate it against the revenue of comparable independent operators or franchisees where that is knowable, and weight it heavily against the judgment of a broker who has leased similar formats in that submarket. Treat the output as a range, not a number, and size the lease commitment so the downside of the range is survivable.
- What is regression forecasting?
- Regression forecasting fits a statistical equation that relates your existing stores' sales to measurable site variables such as trade-area population, income, visibility, parking, co-tenancy, and competitive density, then applies that equation to a candidate site. It produces an explicit, auditable coefficient for every variable, which makes it easy to defend in a real-estate committee. The catch is that it needs roughly 30 or more stores with consistent data before the coefficients stabilize; fitted on fewer, it looks precise and is quietly wrong.