Data and methodology

MarketGround combines public population, demographic, employment, price and business data to analyze the areas you select. This page explains the sources, geographic allocation methods, population models, economy indexes and validation checks behind the results.

Sources

How an area becomes numbers

Census demographic estimates are published for fixed geographic areas. MarketGround allocates block-group values to a hexagonal grid of roughly 0.1 square kilometer cells, using 2020 Census block populations as weights. The selected shape is intersected with that grid, with partial cells weighted by their geographic overlap.

Counts are aggregated across the selected area. Medians are reconstructed from the underlying distribution brackets, rather than added or averaged across areas. The interface and exports include notes for measures presented as approximations.

Benchmarks compare an area with the city, county and metro selected from its interior point, and with the nation. An index of 100 equals the comparison value. For example, 142 is 42 percent higher. Medians and rates are compared directly. Counts with a defined denominator are compared as shares, such as rented homes as a share of occupied housing units. Counts without a denominator appear as values without an index. The results identify areas extending beyond their selected county or metro benchmark.

Combined demographic characteristics use the Census Bureau's published joint tables. These cover sex by age, tenure and household income by age of householder, households with children under 18, and educational attainment by sex. Each estimate is allocated to the selected area. The seven displayed age groups combine the Bureau's 23 age bands, and median age is reconstructed from the age distribution.

Margins of error

Every figure drawn from the American Community Survey is an estimate from a sample, and the Census Bureau publishes a 90 percent margin of error beside each one. MarketGround carries that margin through every calculation and shows it beside the figure as a range, such as 16,249 ± 859 households. Nine times in ten, a complete count would fall inside the range.

For a count, each block group's margin is scaled by the share of the block group the area captures, and the scaled margins are combined as the square root of the sum of their squares. This is the Bureau's own approximation. It ignores the correlation between neighboring block groups, so it tends to understate the range of a large area, and the product labels it approximate. It can also run wider than a margin the Bureau publishes directly. The Bureau estimates a tract from its own survey responses, while an area here is built from its block groups, so a tract drawn as a shape reads a wider range than the tract's published one. For census tract 1 in Dallas County, population reads ± 747 here against the Bureau's ± 554. Ratios such as people per household widen further, because their formula is looser still. Block groups whose estimate is zero still publish a margin, and adding every one of them would widen the range without adding evidence, so, as the Bureau advises, only the largest of those margins is counted, once, for each figure.

We measured that rule on the 2024 release. On a 2 km circle around downtown Dallas (27 block groups), population, households, housing units and Hispanic residents keep exactly the same range, 25 of 121 counts narrow by more than 10 percent, the empty home value brackets narrow from ±50 to ±16, and the figures faded for low reliability fall from 38 to 36. On a 3 km circle around Marfa in West Texas (3 block groups), 26 of 121 counts narrow by more than 10 percent and at most by 37 percent, and 82 figures stay faded either way, because an area that small is uncertain whatever the arithmetic. On a 15 km circle around downtown Dallas (720 block groups), 17 counts narrow by more than 10 percent, and the five figures that were faded before the rule are all reliable after it.

A percentage, such as renters among occupied homes, uses the Bureau's formula for a proportion, which accounts for the uncertainty in both the count and its total. When that formula cannot be evaluated, which happens when the total is much less certain than the count, the Bureau's ratio formula is used instead, as the Bureau instructs. Averages such as people per household and per capita income use the ratio formula. A percentage's range is given in percentage points, so 34% ± 3 pts runs from 31 to 37 percent.

A median rebuilt from brackets has no published margin. Its range comes from the Bureau's design factor method, which takes the standard error of a 50 percent share of the distribution, scales it by the design factor the Bureau publishes for that characteristic and state, and reads the distribution that far either side of the middle. Compared with the Bureau's published margins for every Texas block group in the 2024 release, the range fell within a factor of two of the published margin for 84 percent of block groups for household income, 85 percent for median age and 79 percent for home value. It tends to run slightly wide for income and home value and slightly narrow for age. Where the range would reach an open-ended top bracket, or no design factor is loaded, the page says the range is not computed. A median that can only be blended from published block-group medians, such as median gross rent, has no honest range and says so. The Bureau has said the 2024 release carries the last design factors it will publish, so later releases use those.

What a ring adds to the rings inside it is the difference between two overlapping areas, and combining the two rings' ranges would count the people they share twice. The range of a ring's increment is computed for the band itself instead, from the share of each block group that lies between the two boundaries.

A benchmark index takes its range from the analyzed area's margin alone. The comparison city, county, metro and nation carry far smaller margins at their scale, and theirs is ignored. When the index's range includes 100, the page says the area is indistinguishable from the comparison.

A figure whose margin is more than half its estimate, a coefficient of variation above about 30 percent, is faded and marked low reliability in the interface, reports and exports, and the combined tables follow the same rule. A zero estimate with any margin is marked as well. The number is still shown, but it should not be quoted without its range. A small area can have many such figures, which is what survey data looks like at that scale rather than an error.

Day and night population

Population history uses two sources shown as separate series. LandScan USA from Oak Ridge National Laboratory supplies modeled day and night population for 2016 through 2021. A Census-based series is also available from 2020. Its resident population starts with Census block counts and is updated using the Bureau's annual county and city population estimates, which set each county's total and each city's or township's share of it, and ACS information, which sets the split within a city. Growth is placed by city and not within it, so a newly built edge of a growing suburb can read low while the city as a whole is right. Daytime population is calculated from residents and commuting flows, adjusted for work from home and calibrated against overlapping LandScan results.

LandScan nighttime population describes where people are present overnight, including visitors. The Census-based nighttime series describes residents. The chart keeps these series separate and shows their overlapping years. For example, a 1.5-kilometer area in downtown Houston has 21,185 residents in the 2020 Census, a Census-based estimate of 20,790 for 2025 and a LandScan nighttime estimate of 30,038 for 2021.

The daytime model is calibrated against LandScan in overlapping years. Calibration incorporates modeled activity associated with shoppers, diners, patients and students in addition to commuting patterns. The result is presented as an estimate for the selected area.

Jobs by industry come from the Census Bureau's LEHD Origin-Destination Employment Statistics, or LODES. Workplace-block job counts are allocated to the analysis grid by geographic overlap. The measure counts covered jobs at their workplace locations and is reported separately from population. Alaska uses its last available year of 2016 and Michigan uses 2021 for later periods, as indicated beside those results.

Two years of LandScan, 2019 and 2020, are identical releases, so the chart draws the repeat as a hollow point rather than present it as a second observation. The 2020 Census count is drawn as a square, because it is an enumeration rather than an estimate. On the map, the population layer for 2016 to 2021 is LandScan and for later years it is Census-based.

Travel areas

Travel areas use OpenStreetMap roads and published agency transit timetables. Transit calculations use a selected departure day and time. When a walking area is returned, it is identified as walking in the results.

Timetables are loaded for agencies in thirty-three of the forty largest metros, namely New York, Los Angeles, Chicago, Dallas-Fort Worth, Houston, Atlanta, Washington, Philadelphia, Miami, Phoenix, Boston, Riverside-San Bernardino, San Francisco-Oakland, Detroit, Seattle, Minneapolis-St. Paul, Tampa, San Diego, Denver, Baltimore, Charlotte, St. Louis, Portland, Austin, Pittsburgh, Las Vegas, Cincinnati, Columbus, San Jose, Nashville, Virginia Beach-Norfolk, Providence and Milwaukee. Coverage inside a metro follows the feeds an agency publishes in a form we can fetch and load, so it is not always complete. San Francisco has BART, Caltrain, Golden Gate, AC Transit and SamTrans but not yet Muni. New York has the subway, the Queens and Staten Island buses, the Long Island Rail Road, Metro-North and NJ Transit, but not yet the buses of the Bronx, Brooklyn and Manhattan. Cincinnati is served on the Kentucky side only. Orlando, San Antonio, Sacramento, Kansas City, Indianapolis, Cleveland and Jacksonville have no timetable loaded, and a travel area there says so rather than returning a walking shape.

Standard driving areas use posted or inferred speeds, turn restrictions and one-way streets under free-flow assumptions. Typical traffic calculations use Mapbox routing for a selected departure time and are identified in the results. Walking uses a speed of 4.5 kilometers per hour on the footway and street network. That speed also applies to the walking portions of transit trips, which include scheduled waiting time.

Transit access

Transit access gives two shares of an area's residents. The first is the share living within 400 meters of any transit stop, the second the share within 400 meters of a frequent one. The distance is a straight line from the center of each 90-meter grid cell to the stop, about a five-minute walk, and not a path along the walking network, so a river or a freeway between a home and a stop does not lengthen it. A cell counts as near when its center is within the distance. The residents are the Census-based night population, and because the shares come from that grid rather than from the survey they carry no margin of error.

A stop is frequent when it has four or more departures an hour between 10 am and 3 pm on a weekday, a bus every fifteen minutes or better. The weekday is the busiest of the first two weeks in the agency's current timetable, so a holiday Monday does not stand in for normal service, and departures where the timetable lets no one board are left out. A station is counted once however many platforms it has, carrying every line that calls at any of them. Where an agency's timetable does not group its platforms into stations, each platform counts as its own stop. Only stops that some route serves are counted, and stops of an agency whose timetable has lapsed are left out and named. An area with no current timetable within five kilometers reads "No timetable loaded for this area" rather than zero, since that is a gap in our coverage and not a fact about the area.

County momentum

County momentum is the growth of the counties an area sits in, from the Census Bureau's annual population estimates. Each county holding at least a tenth of the area's residents gets a line, up to three, largest first. The line gives the county's net migration in the newest full year, split into moves from elsewhere in the United States and from abroad, births against deaths, and net migration per 1,000 residents beside the same rate for the state and the nation. Rates use the average of the two July estimates, which is the Bureau's own convention. The figures describe the whole county and never the area itself, because the Bureau publishes the components of change for nothing smaller. The 2020 estimate covers April to July only, so it is never shown as a year's growth, and the small chart beside each line runs from 2021.

On the Pro and Team plans each county line also carries its movers from IRS tax returns. The IRS matches each year's returns to the year before by address, so a return filed in a different county from the year before counts as a household that moved, with the adjusted gross income on it. The line gives the households and income moving in and out, the net income per resident, the average income of households moving in, moving out and staying, and the counties sending and receiving the most income. The IRS does not publish a flow between two counties under 20 returns, so those count only in the totals, and a county where every flow is that small has no figures. The IRS improved its matching with the 2022 and 2023 filing years, so these are not compared with earlier years.

Economy indexes

Each economy index is the same figure computed for an area and for the nation, and 100 times their ratio, so the nation reads 100 and 112 is 12 percent above it. The metro and county columns on the Economy tab are their own indexes against the nation, computed by the same code from the same sources, so every column sits on one scale. Where the inputs carry margins of error the index gets a range, and an index whose range includes 100 is drawn dashed and described as level with the nation. The indexes are modeled from published statistics applied to the area. They are not observed spending or behavior. Social vulnerability is the one exception to the rebasing. It is shown as CDC's own percentile of US census tracts, with the nation's percentile beside it, because that is the scale CDC publishes and the one its users know.

The indexes appear on the Economy tab and in the report's Economy part, with the parts behind each one, and beside every other figure in All Data and the CSV export, for each location and for the combined area. The tab and the report name the price parity year and the week of the mortgage rate they used.

Purchasing power is median household income divided by the area's price level, the Bureau of Economic Analysis Regional Price Parity over 100. A county in a metropolitan area takes its metro's parity and any other county takes its state's nonmetropolitan parity, which BEA defines to include micropolitan counties. Metro membership follows the July 2023 delineation BEA uses. An area's price level is the parity of each county it covers, weighted by the residents there, from the newest parity year at or before the ACS year. If fewer than 99 percent of an area's residents have a price level, the index is withheld rather than computed from the rest. Its range comes from the margin of median household income. A parity is one figure for a whole metro, so within a metro the index follows income alone. It overstates purchasing power in a neighborhood dearer than its metro, such as Manhattan inside the New York metro, and understates it in a cheaper one.

Affordability follows the National Association of Realtors' method for owning: median family income against the income needed to buy the median home with 20 percent down, a 30-year fixed loan and a quarter of income for the payment. It adds two renter measures, twelve months of median rent against median household income and the share of renters paying 30 percent or more of income, each inverted so that a smaller share reads as more affordable. The index is the equal-weight geometric mean of the three. The payment is proportional to the loan at a given rate, so the mortgage rate cancels out of the index against the nation; it is used only for the share of the needed income a median family has, at Freddie Mac's newest weekly rate. Median family income and median rent are blends of published medians with no margin, so the range reflects the home value, household income and renter share margins only.

Affluence stands in for wealth, which the American Community Survey does not measure. It is the equal-weight geometric mean of three measures, each against the nation. They are median home value, the share of households with interest, dividends or net rental income, and the share of households with income of $200,000 or more. All three are required, so a small area without a median home value gets no affluence index rather than one built from two parts. A share measured at zero counts as a part, and inside the mean any part below 1 counts as 1. Its range comes from the three parts' margins, treated as independent. The owner share of occupied homes and the share of households with retirement income are shown beside it for context and left out of the mean, because they follow the age and tenure of an area's households more than its money. A downtown of young renters with high incomes would otherwise read like a much poorer neighborhood.

Household spending applies the Bureau of Labor Statistics' Consumer Expenditure Survey to the area's incomes. The survey publishes mean yearly spending for nine ranges of income before taxes, and each of the area's ACS household income brackets takes the mean of the range it falls in. The one bracket that straddles a range, $60,000 to $75,000, takes two thirds of one and a third of the next. The index is total spending per household against the nation, and each category beside it (food at home and away, housing, apparel, transportation, health care, entertainment, personal care and education) is its own index. It describes what the area's mix of incomes would spend at national habits, not what anyone was seen to spend, so it is shown only as an index and never as dollars. Its range comes from the income brackets' margins. The survey's own sampling error is the same on both sides of the comparison, so it is left out.

Discretionary income is what a typical household has left after taxes, housing and essentials, in the nation's prices. It starts from the area's own aggregate household income. Taxes come from PolicyEngine, the open-source successor to the tax model BLS used until it stopped publishing income after taxes with its 2024 tables. For every state, PolicyEngine computes the federal income, state income and payroll taxes on the survey's average household in each income range, and for the highest range also at the average return the IRS reports above $500,000 and above $1 million, with its mix of capital gains, business income and deductions. The area keeps the share of income those taxes leave, weighted by its residents in each state and read at its own average income above $200,000. From what is left come the area's rent and owner costs, which include utilities, and the survey's food at home, transportation and health care at the area's price levels for goods and services. The Census Bureau withholds the rent or owner cost total for about a sixth of block groups, so an area's published totals are scaled to all of its homes, and the index is withheld where fewer than half of an area's renter or owner homes have one. The remainder is divided by the area's price level and reads zero where housing and essentials take everything after taxes.

Social vulnerability is the CDC/ATSDR Social Vulnerability Index. CDC ranks every US census tract on sixteen census measures, such as poverty, crowding, age and vehicle access, and sums the rankings within four themes and overall. Three of the themes are shown beside the overall percentile, namely socioeconomic status, household characteristics, and housing type and transportation. CDC's fourth, racial and ethnic minority status, is not shown. The overall percentile is CDC's own as published, and CDC counts minority status as one of its sixteen measures within it. For an area, each tract's summed score is weighted by the residents the area captures in it, and the result is placed among all US tracts by CDC's own percentile rule, so a shape that is exactly one tract reads that tract's published percentile. Higher means more vulnerable. The 778 tracts CDC could not score in 2022, 503 of them with no residents, are left out, and the index is withheld where they hold more than half of an area's residents. The nation's figure is the same weighted score over every US tract. CDC publishes no margin for its scores, so the index has no range. The themes shown and their names are CDC's, and the index describes need and exposure rather than anything about the people who live there one by one.

Job mix diversity is how evenly the jobs located in an area spread across the twenty industry sectors of the Census Bureau's LEHD Origin-Destination Employment Statistics, the same workplace counts behind the jobs figures, so they are jobs where people work, not residents. The measure is Shannon evenness, the entropy of the area's sector shares over the entropy of twenty equal shares, which reads 1 when every sector holds the same share and 0 when one sector holds every job, indexed against the nation's. The nation read 0.904 in 2023, so an area spread perfectly evenly reads about 111. Beside it sit the Hachman index, one over the sum of each sector's share times its location quotient against the nation, which reads 1 for a mix identical to the nation's and falls toward 0 as the mix departs from it, and the largest sector with its share. Both are context and do not enter the index, because likeness to the nation is a different question from evenness. Midland County, Texas is spread nearly as evenly as the nation at 94 while its Hachman index is 0.05, because oil and gas holds 28 percent of its jobs against 0.4 percent nationally. LODES publishes no margin, so the index has no range. It is withheld below 1,000 jobs, where one employer moves the mix, and an area touching Alaska or Michigan, whose last published LODES years are 2016 and 2021, says so.

Growth momentum is the equal-weight geometric mean of three yearly growth rates, each indexed against the nation's over the same years: residents, jobs located in the area, and real income per person. Residents come from our Census-derived night surface, which the Bureau's county and sub-county estimates anchor year by year, read on the same cells the analysis uses and never spliced onto LandScan, which ends in 2021 and counts ambient population rather than residents. Jobs are LODES workplace counts on the same cells. Real income per person is BEA's per capita personal income for the counties the area covers, weighted by the residents in each, over the PCE price index, so a dollar in each year buys the same thing; it is a county figure applied to the area, like the price level. Every window starts in 2021, the first year of the residents surface, and ends at the newest year its source publishes, so the three windows differ in length and each rate is annualized. A growth rate can be negative, so a part is indexed as a growth factor, 100 times the area's yearly factor over the nation's, and an area adding 2 percent a year beside a nation adding half a percent reads about 101.5. The spread around 100 is therefore a difference in yearly growth, and a reading of 103 is a fast-growing place. The index needs residents and income and takes jobs when they are present. It is withheld below 500 residents at the start of the window, the jobs part below 1,000 jobs, and the income part where BEA publishes income for the counties of fewer than 99 percent of the area's residents. An area touching Alaska or Michigan, whose LODES is carried forward from an older year, has its jobs part left out and says so. BEA folds Virginia's independent cities into combination areas with their counties and Kalawao into Maui, and each member reads its combination's figure; Connecticut's planning regions, which BEA began publishing with 2024 alone, read the state's. None of the three sources publishes a margin, so the index has no range.

Home prices sit beside growth momentum as context and are not one of its parts. They are the yearly growth in the Federal Housing Finance Agency's annual all-transactions house price index from 2021, the year every growth momentum part starts, to FHFA's newest year. Each block group takes its census tract's growth where FHFA publishes the tract in both years, else the growth of the ZIP codes it lies in, weighted by the homes in each, else its county's, and an area's growth is the mean over its block groups weighted by the owner-occupied homes it captures in each. The panel says how many of those homes each level priced. FHFA leaves out a place with too few sales in a year, and Manhattan has no series with both years, so where fewer than half of an area's homes are priced the figure is withheld. The index is nominal and built from mortgages Fannie Mae and Freddie Mac bought or guaranteed, so jumbo loans and cash sales are not in it.

Mortgage credit access asks how readily people in an area can get a mortgage to buy a home. It reads the home purchase applications lenders report under the Home Mortgage Disclosure Act, for a principal residence on a first lien, closed end, on one to four site-built units, pooled over the three newest years. Three shares are each set against the nation's: the share of decided applications approved, the share not denied for credit history, and the share of loans below Regulation Z's higher-priced thresholds (1.5 points over the average prime offer rate, 2.5 above the conforming limit). The index is their equal-weight geometric mean. Each part is a share of good outcomes, so the spread around 100 is narrow, and the rates themselves are shown beside the parts with the nation's. Over 2023 to 2025 the nation denied 8.7 percent of such applications, 1.4 percent for credit history, and 3.3 percent of its loans were higher-priced; the Bronx denied 21.7 percent and Marin County 6.3. Small lenders may leave out denial reasons and rate spreads, so each rate is measured over the records that report them. A census tract's applications are given to an area in proportion to the share of the tract's 2020 residents it holds, and the index is withheld below 100 decided applications. Manufactured homes are left out of every part, because their lending is a different market: nationally in 2025 they were denied at 55.9 percent against 8.7 for site-built homes, and half their loans were higher-priced. Their share of the area's applications is shown beside the index, as is the share of loans insured or guaranteed by FHA, VA or USDA. The index describes the area as a whole. It is never broken out by the applicants' race, ethnicity or sex, and those fields are not read. HMDA counts every application, so there is no sampling margin.

Tract designations are not indexes. They give the share of an area's residents living in census tracts that a federal program designates, beside the nation's share. A branch in a low- or moderate-income tract counts toward a bank's Community Reinvestment Act lending test, and Opportunity Zones and New Markets Tax Credit eligible tracts carry tax incentives for investment. Residents are counted by the tract they live in, from the 2020 census. Community Reinvestment Act income levels are the FFIEC's census file for the year, which sets each tract's median family income against its metro's, or outside metros its state's nonmetropolitan median: low below 50 percent, moderate 50 to 80, middle 80 to 120 and upper above that, with distressed or underserved nonmetropolitan middle-income tracts flagged as the regulators designate them. The FFIEC cannot class about one populated tract in fifty, mostly college dormitories and military bases, so an area's four levels can add to less than 100 percent. New Markets Tax Credit eligibility is the CDFI Fund's 2016-2020 ACS file: a tract qualifies on a poverty rate of 20 percent or more, a median family income at or below 80 percent of its benchmark, or as a high-migration rural tract, and the Fund's severe and deep distress tiers are shown with it. Nationally 37.9 percent of residents live in an eligible tract. The 2018 Opportunity Zones run to 31 December 2026 and were drawn on 2010 tracts, so each 2020 tract takes the share of its land inside a 2010 zone, from the Census Bureau's tract relationship file, and an area counts that share of the tract's residents; overlaps under 1 percent are boundary redrawing and are left out. Nationally 9.6 percent of residents live in one. The new zones, which take effect on 1 January 2027, will be added when Treasury publishes the certified list. Both CDFI Fund files and the relationship file write Connecticut under its former counties, and those tracts are matched to the planning-region tracts by their unchanged tract codes.

Banking counts the full-service bank branches in the FDIC's Summary of Deposits and the credit union sites serving members in NCUA's branch file inside the area, as of June, and gives branches per 10,000 residents beside the nation's. Deposits are booked where an account sits, not where its owner lives, and a few head offices book deposits from across the country: the 216 offices holding more than $5 billion each hold 46 percent of the nation's $19 trillion. Deposits per resident and each bank's share of the area's deposits therefore count only offices holding under $2 billion, and the concentration figure is the Herfindahl index regulators use to screen bank mergers. Credit unions report deposits for the whole credit union, so they count as branches but not in deposits. Distance to a branch is measured from the center of each resolution-8 cell (about a third of a square mile) where residents live to the nearest bank branch or credit union site, weighted by the residents in each, for the area and for the whole country. NCUA gives credit union addresses and no coordinates, so they are geocoded with the Census Bureau's geocoder; about one site in ten matches no address and is placed at its ZIP code's center, which can put a credit union near the edge of a small area on the wrong side of it.

Add-on datasets

An add-on dataset holds figures published for a census geography, a county or a ZIP, rather than measured inside a drawn area. We read it through the area's block groups. Each block group's share of its county's homes, people or land is known from the decennial blocks, so a count is split by the share of the county the area captures and a rate is the weighted mean over the homes captured, never a sum. The panel says how many source areas the figure came from and what share the largest holds. When one county holds nine tenths or more, the first line says the figures are that county's applied to the area, because that is what they are. An area outside the dataset's coverage reads that the dataset does not cover it, never zero. The first such dataset is Florida registered vessels by county and length class from the Florida Department of Highway Safety and Motor Vehicles, 2019 onward, counted at the owner's registration address.

A scored dataset ranks every county on two scores. Each score is a weighted mean of its factors, where a factor is the county's national percentile on one measured rate, taken over the factors the county has with the weights renormalized to them, and then ranked again to 0 to 100. The panel shows each factor's percentile and the rate behind it, and a coverage figure says how much of the score's weight had data. A factor a county cannot have, such as visits to a Corps of Engineers lake where there is none, is missing rather than zero and drops out of the mean. The boating market scores read registrations, documented vessels, the boat trade and storage, ramps and trailer parking, water and shoreline, marinas, Corps lake visits, seasonal homes, the owner pool within a ninety minute drive and summer vessel traffic from the Coast Guard's shore receivers.

Traffic counts

Traffic counts are annual average daily traffic reported by state departments of transportation through FHWA's Highway Performance Monitoring System, both directions summed. A road within 100 meters of an area counts as bordering it, and we report up to three bordering roads by volume, one per signed route, with the year the count was taken where the state dates it and the release year marked as an estimate where it does not. When no counted arterial borders an area we name the nearest within three kilometers and say how far away it is. Coverage is arterials and highways; local streets are not counted, counts lag a year or two, and quality varies by state. Street names come from the road network at the point where the counted road passes the area.

Through the week, each counted road's annual average is spread over 168 hours by the pattern of FHWA's continuous-count stations (the Travel Monitoring Analysis System), which record every hour of every day. A station's year becomes 168 shares that sum to one, balanced by day of week so a month with more Tuesdays kept than Sundays does not tilt it; stations that kept fewer than twenty days for every day of the week across nine months feed only the group averages. A road takes the median pattern of the stations in its state on roads of the same class group, urban or rural, with the national pattern as the fallback, and a station's own pattern is blended in only when it sits within five kilometers on the same signed route. Every hourly figure is the road's AADT times a share, so the week always sums back to seven days of the count, and the sentence beside each figure names the counters it came from.

When the place itself is busy is a model, not a count. The residents inside an area, the jobs inside it and its open shopping and food-and-drink places, Overture's own two groups, (with transit stops within 400 meters) are each spread over the hours of a weekday, a Saturday and a Sunday by curves from the American Time Use Survey: the share of people at home and awake, the share of the employed at a workplace, and the share at a store, restaurant or other place of leisure at each hour. The job and place terms carry weights fitted on New York City's bi-annual sidewalk counts, by comparing each site's morning to evening and Saturday to weekday ratios with the model's over the same hours. That fit is used only when it explains enough and keeps both weights within reason; on New York's retail corridors it did neither, so the weights are held at one and the sentence beside the grid says so. The result is drawn with the busiest hour at 100 and every other hour as a share of it, and it is always captioned as a pattern of when the area is likely busier or quieter, never as how many people are present.

What opened and what is no longer listed

Each monthly Overture Maps release is captured and compared with the ones before it, state by state. A place first seen after we started watching counts as opened. A place absent from two consecutive releases counts as no longer listed, dated to the first release it was absent from; a single absent release is treated as noise until a second confirms it. Overture does not say whether a place closed, moved or was merged, so we never call a disappearance a closure, and a closing date is shown only where a matched Foursquare record reports one. When a place vanishes and a place with the same brand or a near-identical name appears within 50 meters in the same release, the pair is treated as one place under a new identifier and is excluded from both counts; the share of disappearances explained this way is published with each release in our validation output. Rates are per hundred open places, and the comparison with the metro uses the same rule over the whole metropolitan area. A window longer than our captured history reads since the first capture and says so.

Businesses on the map

The Retail layer draws places as points from zoom 10, the zoom the map opens at. Where a tile holds more than 8,000 open places it draws a deterministic sample of them until zoom 14, nested so a place shown at one zoom stays shown at the next, and every open place from zoom 14. A brand filter draws every location of the chain from zoom 10.

Site criteria

A criteria sheet is a list of thresholds a client applies to a site: residents within a seven-minute drive, household income against the metro median, competitors within two miles, an anchor within 400 meters, growth since a base year. Each line names its own geography, because published sheets mix them, and each site is measured once per geography with the same engine as the rest of the product. The grid then shows four states. Pass and fail are read against the threshold. Close means the measured figure sits within its margin of error of the threshold, the margin described above for a count, a percentage or a median, and five percent for a modeled figure or a median whose range is not computed, because a block-group estimate should not be printed as a hard red when the difference is inside the survey's own uncertainty. Not measured means an input is absent, such as no counted road within the distance asked, and is never a fail.

The starter templates are starting points, not our measurements. Where a franchisor publishes its site criteria the template quotes them and names the source; where none is published the thresholds are industry conventions and the template says so. Applied to the estate they were written from, published criteria describe fewer sites than one would expect: on generous readings of one value salon chain's four preferred-location criteria, 45 percent of its own 4,303 salons meet all four at once. A screen directs attention; it does not settle a decision.

Site comparison stress test

A stress test asks whether a site ranking survives a change in the reasonable assumptions. Each site is measured under every geography chosen (the saved shape, rings from the site's center, travel times from it) and, when a criteria sheet is attached, judged with the thresholds as written, tightened by twenty percent and loosened by twenty percent. Within a scenario the sites are ordered by the lines they pass, then by the measures they lead on, then by the ranking measure; nothing is weighted, and every rank can be traced to the figures beside it. A change of leader between two neighboring scenarios is reported as a flip and named with the figures that moved; when the deciding figures sit within each other's margin of error or five percent, the pair is called level rather than a change. What one ring adds to the ring inside it is a subtraction of sums between nested radii only. A drive time and a radius are not nested, so they are shown side by side and never differenced. Across the Mellow Mushroom estate measured for a September 2026 study, a ten-minute drive reached a median 1.72 times the residents of the three-mile ring in 1.79 times the area, so the two shapes usually disagree, and the stress test says by how much. The result describes how stable a ranking is under the assumptions chosen. It says nothing about the probability that a business succeeds at any site.

Sales forecast

A sales forecast is a fit to figures you supply, one per existing store, against what the product measures around each store on one trade area: every catalog variable, the daytime and jobs figures, the derived densities, any brand or category counts you name, the busiest counted road, and any other numeric columns from your upload. The outcome is taken in logs and must be above zero. Two models are fitted. A ridge regression on standardized features, with its penalty chosen by leave-one-out, is fitted on every estate of fifteen or more usable stores; a gradient-boosted model of shallow trees is fitted beside it on estates of sixty or more, with its number of rounds chosen by early stopping on a holdout inside the training fold. Both are judged on stores they were not fitted on, holding out one metro at a time where the estate spans four or more metros and a seeded five-fold split otherwise, which the report says flatters the model when nearby stores resemble each other. The served prediction comes from whichever model did better on that held-out error, and a tie within two points of R² goes to the simpler one. Every prediction carries a range from the 10th to the 90th percentile of the held-out errors, its largest drivers, and a flag when a feature lies outside the range the estate spans or the candidate's distance from the estate as a whole is beyond what the estate itself contains. Under fifty stores the report warns that the error is a rough guide. A forecast explains the part of a store's figure that its trade area explains; the operator, the building and the lease explain the rest, and the report shows how much was left unexplained rather than hiding it.

Lookalike search

A lookalike search takes a site that works and finds the neighborhoods in a metro that read like it. Every res-8 hex in the country, about three quarters of a square kilometer, carries a profile of eighteen raw dimensions computed the way an analysis of a drawn shape is: who lives there (residents per square kilometer, median age, the share aged 25 to 44, people per household, median household income, the shares of households above $100,000 and below $35,000, bachelor's degree or higher, renters, median home value, work from home), who is there by day (the daytime to night-time ratio, jobs per square kilometer, the office share and the retail and food share of jobs) and what is there (businesses per thousand residents, shops and restaurants per square kilometer, and openings and closures per hundred businesses over twelve months). The three medians are rebuilt from the Census brackets apportioned to the hex, as every median in the product is, and the daytime figure carries the same calibration the analysis page shows. The reference is the same profile over the cells a saved location or a drawn shape covers; a shape above fifty square kilometers averages to nowhere and is replaced by a one-mile ring around its center, which the panel says. Each dimension is centered on its national median and divided by its interquartile range, after a log for the skewed densities and dollar figures, so a reference profiled in one metro means the same thing in another. The distance between two profiles is the root of the weighted mean squared difference over the dimensions both have, with the weights you set from zero to three, and the match is the share of that weight within half an interquartile range. Hexes with fewer than two hundred residents and fewer than a hundred jobs are never candidates, the reference and the two rings of hexes around it are excluded, and no result is chosen within two rings of another, so twenty results are twenty neighborhoods. Every result shows all eighteen values beside the reference's, with the closest dimensions and the widest gaps in words. Race, ethnicity and household composition are not dimensions and cannot be added as weights, because a tool that ranks neighborhoods on them is a liability for the brokers who use it whatever the intent. The business dimensions cover the Overture listing alone. A result is a starting point for an analysis, not a score, which is why the pin opens the ordinary analysis around it.

Decision history

A decision recorded on a project freezes what was on the table at the time: the candidate sites and how many lines of the criteria sheet each passed, the sheet's stated lines and version, and the stress test's sentences. The project moves on; the record keeps its context. An outcome (a lease signed, an opening, a loss in negotiation, a withdrawal) is a separate, later, optional entry on the same record. Nothing scores a decision by its outcome, because a rejected site may have failed a lease despite an attractive market, and a signed lease says nothing about the analysis that preceded it. Where an organization has switched on its consent, outcomes may be counted into aggregate figures by outcome kind and concept, as counts only, and never as sites, clients or reasons.

Report narrative

A report can open with a few paragraphs of prose. The draft is written by a language model from the project's own figures, the same summary the analysis tabs show plus the trend, the places roll-up, the screener grid and the stress test's sentences, under instructions that forbid any number not in that input. Before the draft is shown, every number in it is matched against the input within the rounding written, so "about 41,200" is accepted for 41,183 and "$80,000" is refused for $71,000. A sentence with an unmatched number is removed and the count of removed sentences is shown. The check is a filter, not a proof, because a sentence can be wrong without a number in it, so nothing prints until a person has read the draft, edited it and saved it, and the report says who did and which data release the draft was written against.

Competition and market share

For a chosen kind of business, a category from Overture's taxonomy with everything under it or a brand matched exactly as listed, the product counts the open places inside a shape and within a ring around it, and divides the shape's residents by the competitors inside it, indexed against the same figure across the containing metro, where 100 equals the metro. Those two figures assume nothing beyond the place data and the project's own competitor pins. Unless a project switches them off, the competitor pins its members placed count inside the shape and within the ring like listed places, and each replaces the listing it was made from, or a listing of the same business within 75 meters of it, so a store is never counted twice. The result says how many competitors are pins.

The market share estimate is a Huff gravity model. Each cell of the shape's population grid gives its residents to the stores in proportion to travel time raised to a negative power, the decay, which defaults to 2 and is shown beside 1.5 and 2.5 so the range is visible. Travel times come from our own routing on the free-flow road network; when it is unavailable the model falls back to straight-line distance at 30 km/h and says so. Every store is treated as equally attractive because the data holds no floor area, sales or footfall, so the result is a measure of position, not performance. The share is of nearby residents by travel time alone. It is never revenue, footfall or a forecast, and the assumptions print beside the number on the report. Across a project's own locations the same model reports the residents two locations both claim, which is the model's view of cannibalization beside the plain geometric overlap of their trade areas.

Co-tenancy profiles

For every brand with at least twenty listed stores nationally, keyed by its name with punctuation and spacing removed, the co-tenancy profile is the set of brands and categories found within 400 meters, a five-minute walk in a straight line, of its existing stores, excluding the brand's own other stores. Each neighbor carries the share of the brand's stores that have it, the base rate of that neighbor around branded stores in general, the lift as their ratio, and a weight equal to the share times the log of the lift, so a neighbor that is both common and distinctive counts most. A neighbor enters the profile at a share of at least ten percent and a lift above 1.2, at most sixty per brand.

A site's co-tenancy match is the weight of the profile's neighbors listed within the radius of the site divided by the whole profile's weight, shown with the neighbors present and missing so the number can be checked. For a vacancy the same arithmetic scores every profiled brand and ranks them, split into brands already within the radius, brands present elsewhere in the metro, and brands not yet in the metro. Profiles are rebuilt after each places capture, and each brand carries how much of its top twenty neighbors survived from the previous build. It is neighbors, not visitors. Nothing here observes who shops where, and a match says only that the businesses a brand usually sits beside are already present.

Chain footprints

A brand's footprint is the count of its listed stores in the newest places capture, by state and by metro, with the residents and households of each area divided by that count. Brands are keyed by their name with punctuation and spacing removed, so a brand listed under two spellings appears twice, and only brands with twenty or more listed stores have a footprint; the largest five hundred get a public page. Metros are Census metropolitan statistical areas, and a metro of 250,000 or more residents with no store listed is shown first, ordered by population, because that absence is the question most readers bring. Residents and households are the American Community Survey five-year estimates summed over the block groups inside each area. Newly listed and no longer listed count the changes recorded between captures and name the first capture, because the history begins in August 2026 and a short history is not a trend. A footprint is a count of doors, never a measure of how well any store does, and every page says so in its first line.

Both co-tenancy profiles and footprints count storefronts only. The places capture lists a Western Union counter inside a grocery store, an Amazon Locker, a bitcoin or bank ATM, a propane or key kiosk and an electric vehicle charger exactly as it lists a shop, and counted as stores they would make Western Union the largest brand in the country and a common neighbor of every coffee shop. So a place filed as an ATM, a kiosk, a package locker or a charging station is left out, and so are the money transfer, coin, key and propane counters and the grocery sushi counters that operate inside other businesses. A bank keeps its branches and loses only its ATMs.

What we check

Validation compares loaded Census blocks with published state counts and checks that grid allocation preserves block-group population totals within computational tolerance. Reconstructed medians are compared with published estimates and their margins of error. Modeled population totals are checked against the Census population estimates used for calibration. Data must pass the required checks before publication.

Counties and census tracts are assembled from their component block groups. Other named boundaries are intersected with Census blocks for areas up to 5,000 square kilometers. Larger areas use the fractional hexagonal grid scaled to the selected area. The same methods are used when calculating regional benchmarks.

The economy indexes compute the nation through the same functions as an area, and the tests hold the nation to exactly 100. A price parity load is refused unless BEA's files put the nation at 100 and the newest year has a parity for every county in the 50 states and DC. Each year from 2016 to 2024 covers all 3,144. When the parities were first loaded, the resident-weighted price level of the whole country came to 99.3, within a point of BEA's 100, which confirms that counties were matched to the right metro or state portion. The income source tables behind affluence cover every block group with a household count in each year loaded.

The spending table is refused unless its nine income ranges, weighted by the survey's own household counts, reproduce its all-household column. Income and most spending lines agree within a dollar, and the lines BLS draws partly from its diary survey within one percent. The tax model was checked against hand calculations from the 2024 tax schedules, and every state without a wage income tax returns none. Run under 2023 law on the survey's households, it reproduces the share of income BLS published as left after taxes for 2023 within 0.6 points for incomes from $50,000 to $200,000, within 1.7 points above that and within 3 points below, where refundable credits run slightly rich. On the IRS's average returns its federal rates come within 0.6, 0.1 and 0.9 points of the rates the IRS reports for incomes of $200,000 to $500,000, $500,000 to $1 million and $1 million or more.

A Social Vulnerability Index load is refused unless CDC's percentile rule, applied to each tract's summed score, gives the percentile CDC published for it. On the 2022 release it does so exactly for all 83,342 scored tracts, overall and for every theme, and every tract with residents in our block groups is in CDC's file.

The twenty-sector jobs surface refuses any block whose sectors do not sum to its total, each state's cells must hold its blocks' jobs to within rounding before the state is kept, and the national rollup the nation's mix is summed from is refused until every state's cells match the jobs surface. The 2023 build holds 150,944,613 jobs, the jobs surface's total to the job, and the tests hold the nation's Hachman index against itself to 1.

The real income loader refuses a file that is not BEA's CAINC1 table, requires every county holding a block group to have a figure for BEA's newest year through its own row, its combination area or its state, and reads BEA's own United States row as the nation, whose population matches the Bureau's midyear estimates BEA cites. The growth engine refuses to index two series whose windows differ, and the tests hold the nation to 100 on every part and the composite.

The HMDA loader refuses a file that is not the FFIEC's loan-level layout and a year in which fewer than 99 percent of the decisions naming a tract outside Puerto Rico fall in our 2020 tracts. That check found that HMDA filed Connecticut's 2023 applications under the state's former counties, about 31,000 home purchase decisions, so those tracts are matched to the planning-region tracts by their unchanged tract codes, a match that agrees with the CT Data Collaborative's crosswalk for all 879 tracts it lists. The 2.2 percent of decisions with no tract are left out of the area and the nation alike.

Data periods used in the published examples

The location examples use American Community Survey five-year estimates ending in 2024, workplace employment data for 2023 and modeled daytime population for 2025. The business comparisons use the Overture Maps release of August 19, 2026. The case studies were calculated on September 12, 2026, and the research comparisons were calculated on September 13, 2026. Historical population sources and their periods are described above.

What we do not publish

This page documents the sources, overall methods and validation process. MarketGround retains its specific calibration factors, model parameters and allocation implementation as proprietary information.

Source attribution includes the US Census Bureau, the Bureau of Economic Analysis for Regional Price Parities, Freddie Mac for its Primary Mortgage Market Survey (retrieved through FRED at the Federal Reserve Bank of St. Louis), the Bureau of Labor Statistics for the Consumer Expenditure Survey, the IRS for its Statistics of Income, taxes computed with PolicyEngine, the Centers for Disease Control and Prevention and Agency for Toxic Substances and Disease Registry for the CDC/ATSDR Social Vulnerability Index (2022 database, US), LandScan USA from Oak Ridge National Laboratory, the Overture Maps Foundation, OpenStreetMap contributors, Protomaps, the transit agencies that publish the timetables behind transit travel areas, and the Bay Area 511 open data program for the Muni and SamTrans timetables. The associated terms include public domain Census, BEA, BLS and IRS data, the citation CDC/ATSDR asks for its index, CC BY 4.0 for LandScan USA, CDLA-Permissive 2.0 for Overture places, ODbL for OpenStreetMap, and each agency's or 511's own open data terms for timetables.