Search retail sales forecasting, and almost everything on page one is about inventory: how many units of a given SKU will move off the shelf next week, so the replenishment system knows what to order. That is an important discipline, and tools like RELEX, o9 and Oracle are built for it. It just isn't the forecast that decides whether you sign a ten-year lease on an empty box.
The forecast that decides a lease answers a completely different question: what can this brand do at this specific address, where your register hasn't run yet? You cannot project from the site's prior sales because there are none. You only have your existing fleet, the trade area around the pin, and whatever honest track record your past projections left behind.
This guide focuses squarely on that second job. We will take an illustrative candidate site from analog selection to a forecast range and a rent ceiling, then show how to backtest the model against stores you already run before putting real capital behind it.
Two jobs share the name retail sales forecasting
Demand forecasting starts with a store's existing register tape and projects by SKU, day and promotion. When a demand forecast misses, you sit on excess inventory or run out of stock, and you adjust your reorder next week. Location forecasting starts from zero at the curb and borrows patterns from the rest of your fleet. When a location forecast misses, you pay for the error across a decade of occupancy costs.
The data inputs reflect that split. A demand model feeds on point-of-sale history, prices, promotions and weather. A location model feeds on the trade area: households and median income within the drive time, daytime population, vehicle and foot traffic, co-tenancy, competitor proximity, and distance to your own stores. If a vendor's forecasting page talks about safety stock, replenishment cadence and SKUs, it describes a demand tool, no matter how many times it says store. For that work, read our guide to retail demand forecasting.
What public chains say a new store will do
Most public chains with an expansion plan share a new-store economic model with Wall Street. They outline average unit volume, buildout capex and payback periods. That is a location forecast, generalized across an entire fleet. Your job is the same forecast, made for one specific parcel.
| Company | What it tells investors about a new store | Which way the ramp runs |
|---|---|---|
| Five Below | About $2 million in sales in the first full year, on about $0.4 million of cash investment, with a target payback of about one year (10-K) | Not stated as a curve |
| Ollie's | Targets about $4.0 million in first-year sales and about a two-year payback (10-K) | Opens high: new stores have higher first-year sales than the average existing store |
| Chipotle | New restaurants run at about 80% of the system average (Q3 2025 call) | Opens low: around 36 months to ramp sales and profitability (10-K) |
| Tractor Supply | New store productivity in the 65% to 70% range (Q1 2026 call) | Opens low and builds |
| CAVA | Model volume of $2.3 million in year one, rising to $2.5 million in year two (Q4 2024 call) | Opens low and builds |
| Floor & Decor | Its last full published model called for $14 million to $16 million in first-year sales (FY2021 10-K). Its latest 10-K says recent store classes are seeing "lower first year sales and initial returns" (FY2025 10-K) | Opens low: new stores sell less than stores open more than a year |
Two hard realities from these disclosures apply directly to any site review. First, the ramp does not always curve upward. Brands like Ollie's, Cheesecake Factory and BJ's Restaurants describe their new stores opening hot above their eventual run rate, then settling down. Chipotle, Tractor Supply and CAVA see the opposite: stores start slow and build volume over time. If your forecast assumes the wrong ramp shape, your year-one sales are wrong even if your mature estimate is dead-on.
Second, a fleet model that worked can drift. Newer vintages at Floor & Decor have run below the numbers laid out in earlier fleet models. In July 2008, Starbucks announced roughly 600 US store closures; around 70% of those units had opened since the start of fiscal 2006 (8-K). Both are seasoned, well-run expansion operators. In both cases, the fleet-level model held right up until the sites changed underneath it. That is why a site forecast cannot be an equation you set once and forget. It has to be tested against reality as you grow.
The four ways to forecast a new store's sales
Spatial forecasting has decades of academic and commercial research behind it. The best single overview is Fildes, Ma and Kolassa's paper "Retail forecasting: Research and practice" in the International Journal of Forecasting (2022), available as a free working paper. Their breakdown covers the four methods most teams use.
| Method | How it works | What it needs | Where it breaks |
|---|---|---|---|
| Analog stores (Applebaum, 1966) | Find the stores in your fleet whose trade areas look most like the candidate's and borrow their sales, or their sales per household by distance band | Customer addresses and a fleet varied enough to have true look-alikes | A market unlike any you operate in. An analog that's over- or under-performing passes its error straight into the forecast |
| Gravity model (Huff, 1963) | Splits each neighborhood's spending among all nearby stores by their size and distance, competitors included | Spending by zone, a full competitor list, travel times, and customer data to calibrate it | Data-heavy to calibrate, doesn't transfer between markets, and is usually checked only on stores already open |
| Regression on your stores | Relates each store's sales to trade-area variables, then plugs in the candidate's values | Clean sales by store, and enough stores for the number of variables | Explains existing stores well, then misses on new ones. Treats each site as if your other stores weren't there |
| Machine learning | Gradient-boosted trees and similar models find non-linear patterns across many variables; tools like SHAP show what moved each forecast | The most stores of all, and disciplined hold-out testing | Overfits small fleets. SHAP explains what the model learned, which isn't proof of what causes sales |
Most working forecasts combine several of these: analogs to establish a baseline, statistical or machine learning models to adjust for local nuance, and a trade-area overlap calculation to account for cannibalization. Gravity modeling is more common than people assume; Fildes and his co-authors cite an earlier survey showing that roughly two thirds of retail location planning teams use it.
The paper's most critical insight concerns validation. The authors found little evidence of how accurate these models are "beyond untested claims." They also highlight a structural trap in commercial expansion: sites with the highest forecast outputs are the most likely to get approved and built, so as a cohort they can come in systematically below projections through regression to the mean. Both observations point to the same operational rule: test any model against your own historical openings before you lean on it for your next lease.
A worked forecast, one step at a time
To see how these layers interact, here is how an underwriting forecast comes together for an illustrative candidate site. The numbers and brand are an example, but the arithmetic follows real expansion workflow.
Assume a specialty retail chain of 64 stores. The fleet reaches maturity at year three, averaging roughly $2.0 million annually per mature unit. The deal on the table is a 7,500-square-foot end cap in a grocery-anchored center on the edge of a growing suburb, and the landlord wants an answer in two weeks.
Step 1: Decide which number you're forecasting
Clarify the target metric before opening a worksheet. A lease decision usually needs three distinct outputs: mature annual sales at stabilization (year three for this fleet), first-year sales during the initial ramp, and net new sales to the brand after subtracting sales pulled from neighboring stores. If your unit economics hinge on members, covers or washes rather than register gross, build your projections around that core driver. A fitness brand, for instance, should forecast members, not just revenue.
Step 2: Find the analog stores
Analogs are stores in your current portfolio whose local trade areas match the candidate parcel. Screen on drivers that move your specific sales, not on proximity: total households and median household income within a 10-minute drive, daytime population, the anchor and co-tenants, competitor count and store size. The nearest open location is frequently a misleading analog because it sits in a different kind of trade area.
For this candidate, the five closest trade-area matches in the fleet generated third-year sales of $1.71 million, $1.88 million, $1.96 million, $2.12 million and $2.34 million. The median is $1.96 million. That spread is the first honest thing the forecast tells you: even stores that look alike on paper differ by more than half a million dollars a year.
Step 3: Adjust for what the analogs don't share
Every real estate site has unique trade-area traits. In our example, the candidate parcel has roughly 8% fewer households in its drive-time shed than the analog average, and there is a direct competitor sitting 0.3 miles away, a dynamic none of the five analogs deal with. Offsetting that, the grocery anchor at this center drives higher weekly visit counts than the anchors at the analog locations.
A calibrated model quantifies these trade-offs by evaluating how each variable impacts sales across the whole fleet, rather than relying on a small five-store sample. Applying what the fleet taught the model puts this candidate's mature forecast at $1.90 million, within a projected range of $1.65 million to $2.15 million. That band accounts for both the baseline analog dispersion and how far the model missed when it was tested against holdout stores.
Step 4: Ramp it to year one
Across this chain's portfolio history, first-year stores average roughly 85% of their mature year-three run rate. That establishes a year-one baseline projection of $1.62 million, within a range of $1.40 million to $1.83 million. You should apply your chain's specific ramp curve, broken out by format and by new versus existing markets if your fleet size allows. For brands whose units open hot and then settle, as Ollie's and Cheesecake Factory describe, the year-one adjustment works in reverse.
Step 5: Take out what it pulls from your own stores
The chain operates Store 22 seven miles away. About 15% of the households in Store 22's trade area fall directly within the candidate site's drive-time polygon. A cannibalization run indicates the new store will siphon roughly $180,000 annually from Store 22. Deducting that volume leaves net new sales to the brand of approximately $1.72 million at maturity, spanning a range of $1.47 million to $1.97 million.
That net figure is what your CFO and investment committee need to evaluate. Opening a store that rings $1.90 million while cannibalizing $180,000 from an existing lease is fundamentally a $1.72 million capital allocation choice.
Step 6: Turn the forecast into a rent number
Every forecast serves an underwriting goal: establishing what you can afford to pay on the lease. If your brand sets a total occupancy threshold of 10% of revenue (base rent, common area maintenance, taxes and insurance), structure your deal terms against the conservative end of the mature range rather than the midpoint. Applying 10% to the $1.65 million conservative threshold yields a maximum occupancy budget of $165,000 annually, or $22 per square foot gross across the 7,500-square-foot space.
If the landlord asks $28 per square foot base plus $6 in common area charges, taxes and insurance, your total occupancy hits $255,000 annually. That represents roughly 13% of your midpoint projection and over 15% of your downside case. Instead of telling the leasing broker the space feels overpriced, you have the exact numbers and revenue requirements to anchor your counteroffer. Real estate teams with a model running in their platform can do this math the same day as the site visit and get the LOI out faster. When the deal goes to committee, the range, the analogs and the rent math are the backbone of defending the forecast.
Check the model on stores you already opened
Every step in this process relies on the assumption that your model truly understands what makes your stores work. Before relying on an algorithm to underwrite capital on raw space, evaluate how well it would have forecasted the fleet you have already built.
First, Discovery: learn what actually drives your stores
Validating a forecast begins well before running predictive equations. An initial Discovery phase analyzes your historical store network to resolve three core operational questions: who your core customer is, which variables actually move sales across your fleet, and where your actual customer trade area begins and ends. The answers often surprise the team. A frozen dessert brand we worked with believed stores that sold a higher share of pints did better overall. When the data was tested, pint mix didn't predict revenue, and the model was built on the variables that did.
Then, backtest on stores the model never saw
Once the model learns your fleet, hold part of it back. The cleanest version is an out-of-time backtest. Train your model strictly on locations opened before a specific calendar date, then task it with forecasting every subsequent opening without giving it access to post-opening revenue. This mirrors the exact operational position of underwriting a prospective site: evaluating an unopened address with zero register history.
Review your backtest output using this sequence:
- Rank order first. Confirm whether the model roughly sorts the held-back stores the way they actually performed; that is most of what a real estate team needs. If it ranks your weakest location near the top, stop there.
- Then the miss in dollars. Track the absolute variance across each forecast and identify systematic bias. If your predictions consistently run hot in new markets, the model is telling you something about new markets. Similarly, keep the regression-to-the-mean trap in mind: verify whether your model routinely over-projects volume on your best-looking sites.
- Then the outliers. Every retail chain has an outlier store that outperformed or underperformed its underwriting. Inspect what the data model caught and what it missed. That review is where your team's field experience refines the model's parameters.
Your backtest variance also sets the baseline confidence band for all subsequent site evaluations. If your holdout tests missed by a certain amount, a forecast on a new site should not claim a tighter band than that. We recommend applying this exact backtest requirement to any external vendor under consideration, as detailed in our guide to choosing site selection software.
What it takes to forecast your own stores
Building a forecast model trained on your own sales requires adequate training volume: typically 40 or more mature operating stores with stabilized sales histories. With a smaller portfolio, multi-variable statistical models tend to overfit and memorize individual site quirks rather than isolating true fleet patterns.
At GrowthFactor, custom revenue forecasting is handled through Labs, built collaboratively alongside your team through an established sequence.
Discovery runs for 30 days. It analyzes your sales history to define your core customer profiles, isolate the specific variables driving top-line sales, and measure your actual trade areas. Discovery is normally $5,000 and free for brands operating 40 or more stores. A mutual NDA is signed upfront. Your sales data is never pooled or shared with any other retailer, and each customer's model is trained only on that customer's own stores.
The model is built and fine-tuned by Labs data scientists with your team over a few months of back and forth, focusing on the specific drivers validated during Discovery.
The backtest runs the calibrated model against holdout stores it has never analyzed, so your team reads the results before anyone uses it on a lease.
Live sites follow. Once deployed, your custom model operates directly inside the platform, generating complete site forecasts in under 10 seconds. Drop a pin on a candidate parcel to review your revenue range, and toggle inputs on and off to see what moves the projection. If you already have a model built elsewhere, the platform can load it side-by-side with your Labs model to compare outputs on the same parcel.
GrowthFactor customers report forecast error roughly half the industry norm (GrowthFactor customer survey, January 2026). TNT Fireworks opened 153 locations in six months, with every location meeting or exceeding its revenue projection. "Our business is very seasonal, and we have to manage evaluating hundreds of locations in a short time frame, while picking the best ones. GrowthFactor makes doing that simple," said Carson Anderson, TNT's Managing Director. You can explore more on our customer stories page.
If your portfolio currently sits below the 40-store threshold, rely on foundational site selection practices rather than complex algorithmic models. Match candidate sites against your best analogs, establish clear underwriting criteria using plain-language Custom Evaluators, model local cannibalization, and spend time walking the corridor. Keep detailed records of your underwriting assumptions for every new opening, ensuring you have clean, structured data ready to train a model once your fleet reaches scale. Small teams can still move fast on the platform: Lil Sweet Treat went from 2 locations to 8 in a year, every one at or above underwriting, with two founders and no analysts.
To review a live forecast run against your target development markets, book a demo. If your fleet operates 40 or more locations, start with Discovery.
Frequently asked questions
What is retail sales forecasting?
Retail sales forecasting estimates what a retailer will sell over a future period, and it covers two different jobs. Demand forecasting projects units by product, store and week from each store's own sales history, so the right inventory gets ordered. Location forecasting estimates what a store at a specific address will sell in a year, often before it opens, so a real estate team can decide whether to sign the lease and at what rent.
How do you forecast sales for a new retail store?
Start with the stores you already run. Find the analog stores whose trade areas look most like the candidate's, adjust for what the candidate doesn't share with them (fewer households, a closer competitor, a stronger anchor), apply your ramp from year one to maturity, and subtract the sales the new store will pull from your nearby stores. Report the result as a range, and check the model on stores it never saw before you use it on a lease.
What's the difference between retail demand forecasting and new store sales forecasting?
Demand forecasting extends a store's own sales history forward by SKU and week, and a miss shows up as overstock or empty shelves that get fixed next week. New store sales forecasting has no history at the site, so it borrows from your other stores and the trade area, and a miss shows up as a lease you carry for years. The inputs differ too: prices, promotions and seasonality for demand, and households, traffic, competitors and analog stores for a location.
What methods are used to forecast a new store's sales?
Four, often in combination: analog stores (use the performance of your most similar existing stores), gravity models such as the Huff model (estimate the share of trade-area spending a store captures given its size, distance and competition), regression on your existing stores (relate sales to trade-area variables), and machine learning models that find non-linear patterns across many variables. Each needs enough of your own stores to learn from, and each should be tested on stores it didn't see.
How many stores do you need to build a sales forecast model?
Generally 40 or more mature stores with revenue data for a custom model trained on your own sales. The more variables a model uses, the more stores it needs to learn from instead of memorizing. Brands with fewer stores get more from analog comparisons, written site criteria, cannibalization checks and a site visit, while they build the store history a model needs later.
How accurate is a new store sales forecast?
It depends on the brand, the model and how similar the new site is to the stores the model learned from, so a vendor's headline accuracy figure tells you little. The useful measure is how the model did on your own stores that it never saw: train it on stores opened before a cutoff, forecast the ones opened after, and compare both the rank order and the dollar miss. That miss is also the honest width of the range on every forecast that follows.
How long does a new store take to reach mature sales?
It varies by format. Chipotle says a new restaurant can take around 36 months to ramp up sales and profitability, while Ollie's says its new stores open with higher first-year sales than its average existing store. Use your own stores' ramp curve, split by format and market type if you have enough openings.
What is retail forecasting software?
Most products sold as retail forecasting software, such as RELEX, o9 and Oracle, are demand planning tools that forecast inventory by SKU. Software that forecasts what a store location will sell is usually sold as site selection software or a consulting engagement, from vendors such as GrowthFactor, Audiense (formerly Buxton), Kalibrate, SiteZeus and Esri. Ask any of them to forecast stores you already opened, without seeing their sales, before you trust a forecast on a new site.
