How Our Safety Map Works

A step-by-step look at how we turn millions of crime reports into the neighborhood safety map — from raw police data to the colors on your screen.

Cities

6M+

Crime incidents

84K+

Neighborhoods projected

The pipeline, step by step

Eight stages take us from raw police reports to the map you see.

Collect real crime reports

We gather millions of individual crime incidents directly from police departments and public data portals across 27 major US cities.

Sources include city open-data portals, the Crime Open Database (CODE), and direct police department feeds
Each incident includes what happened, when, and where — down to latitude and longitude
Our database currently holds 6+ million incidents spanning 2018 through 2024

Standardize into 16 crime types

Different cities describe crimes differently. We map every incident to one of 16 standard categories using the FBI's NIBRS classification system.

Violent crimes: Homicide, Physical Assault, Robbery, Sexual Assault, Kidnapping, Weapons
Property crimes: Residential Burglary, Commercial Burglary, Theft from Vehicle, Shoplifting, Other Theft, Vehicle Theft, Vandalism
Disorder crimes: Drug Activity, Public Disorder

This means an "ASSAULT, AGGRAVATED: DOMESTIC VIOLENCE" from one city and a "Felony Assault" from another both land in the same category — making apples-to-apples comparisons possible.

Correct for reporting differences

Not every city reports crime the same way. Some under-report certain categories. We detect and adjust for these gaps so one city's low numbers don't create a false sense of safety.

We identify "anchor cities" with consistent, thorough reporting and use them as a baseline
For each crime type, we measure how much a city's reporting deviates from what demographics would predict
We also handle COVID-era distortions: shoplifting dropped in 2020 because stores closed, not because crime decreased — so we exclude or adjust those years

Understand every neighborhood

For each of the 84,000+ census tracts in the US, we assemble a detailed profile of what makes that neighborhood tick — far beyond just crime history.

Demographics & economics: income levels, poverty rate, unemployment, education, age distribution
Built environment: number of bars, transit stops, retail stores, parks, gas stations — extracted from OpenStreetMap
Housing: home values, vacancy rates, homeownership, building age
Spatial context: what's happening in neighboring tracts (crime doesn't stop at boundary lines)

Train projection models

Using neighborhoods where we have real crime data, we train machine learning models that learn the relationship between a neighborhood's characteristics and its crime patterns.

We build 16 separate models — one for each crime type — because the factors that drive theft are different from those that drive assault
Each model uses gradient boosting (LightGBM), a proven technique that learns complex patterns without overfitting
Models are validated by training on some cities and testing on others to ensure projections generalize to new areas

Smooth across boundaries

Raw projections can have sharp jumps between adjacent tracts. We smooth them using physical barriers — a highway, river, or rail line between two neighborhoods means crime patterns are less likely to spill across.

We extract barrier data (highways, waterways, rail lines, elevation changes) from OpenStreetMap
Neighborhoods separated by barriers are weighted less in smoothing; adjacent neighborhoods with no barriers are weighted more
This produces a map that reflects how crime actually flows through physical space

Track trends over time

For neighborhoods with multi-year crime data, we calculate whether crime is trending up, down, or holding steady — shown as Improving, Stable, or Deteriorating on the map.

We fit a trend line through yearly crime densities from 2018 to 2024
A neighborhood is "improving" if crime density dropped more than 5% per year, "deteriorating" if it rose more than 5%, and "stable" otherwise
COVID-affected years are excluded from trend calculations to avoid misleading signals

Color the map

Finally, we convert projected crime density into the green-to-red color scale you see on the map. Colors are relative — they show how a neighborhood compares to all other neighborhoods nationwide.

We compute national percentile thresholds: the 50th, 75th, 90th, and 95th percentile of crime density across all tracts
Green means below the national median. Yellow is moderate. Orange is elevated. Red is among the highest in the country
Thresholds are computed separately for each crime type, so rare crimes (like homicide) and common crimes (like theft) each show meaningful geographic variation

Reading the map colors

Colors are comparative, not absolute. They show how a neighborhood ranks against all other neighborhoods in the country.

Lower crimeNational medianHigher crime

Green

Below the national median for this crime type. Lower than most neighborhoods — but not zero crime.

Yellow

Around the national median. Typical crime levels — where most neighborhoods fall.

Orange

Above the 75th percentile. Elevated compared to most neighborhoods.

Red

Above the 90th percentile. Among the highest crime areas nationally.

Common questions

Where does the data come from?

All crime data comes from official public sources — city police department open-data portals and the Crime Open Database (CODE), an academic dataset of harmonized incident records. Neighborhood characteristics come from the US Census Bureau (American Community Survey) and OpenStreetMap. We do not use private or proprietary data.

Why do some neighborhoods have no reported incidents?

Some local police departments do not publish individual crime records through open-data portals. In these areas, map colors are based on projections from neighboring areas and demographic patterns. When we have real incident data, we show it.

How accurate are the projections?

Our models explain roughly 45% of the variation in crime density across neighborhoods — which is strong for this type of prediction. However, projections are estimates, not certainties. They're most reliable in areas near cities where we have real data, and less certain in rural or remote areas. Use the map as one input among many, not as a definitive safety rating.

What do the colors mean?

Colors are relative to all other neighborhoods in the country. Green means crime density is below the national median — not that there's zero crime. Red means crime density is in the top 5% nationally. Most neighborhoods fall in the yellow-to-light-orange range.

Why don't you just show a safety score?

Safety means different things to different people. A single number hides the nuance — a neighborhood might have high property crime but very low violent crime, or vice versa. We let you explore each dimension separately and form your own assessment. The 16 crime category filters on the map let you focus on what matters most to you.

How often is this updated?

Crime incident data is refreshed as new data becomes available from source agencies, typically spanning 2018 through the most recent available year. Census and demographic data is from the 2024 American Community Survey. We re-run the full projection pipeline whenever we add new cities or data sources.

Explore the map

See these projections in action. Browse any neighborhood in the US and explore safety data across 16 crime categories.

Open the map

Our full data pipeline is open source. All crime data comes from official public sources. We do not collect user data, sell information, or accept funding from the real estate industry.