UrbanRisk — Spatio-Temporal Crime Risk & Day-by-Day Validation Map

Grid Safety Score (0–100) Baseline

0–40 (Hotspot F) 55–70 (Caution C) 85–100 (Safe A+)

Actual Crime in Top 20% Predicted Hotspot (HIT)

Actual Crime Outside Top 20% Hotspot (MISS)

Click any grid cell or street on the map to inspect its Safety Score & Hawkes risk breakdown.

Selected Location Safety Score Click map

Grade -

Select a location on the map

Click any micro-grid cell or street coordinate to compute its calibrated Safety Score & next-24h Poisson crime probability.

Next-24h Crime Prob

--%

Expected Crimes / Day

Historical Incidents

Violent Crime Share

--%

Personal Safety (Violent)

-- Grade -

-- violent / day

Property & Auto Theft

-- Grade -

-- theft / day

Ambient Pedestrian Exposure (H5): -- (Baseline)

Safety Score by Time of Day

Self-Exciting Hawkes Intensity (λ̂) Point Process

Decomposes predicted daily rate into chronic baseline density, recent near-repeat self-excitation (7d/14d), and spatial spillover from adjacent blocks:

Expected Crime Type Distribution (Leakage-Free)

Rolling Out-of-Time Daily Forecast Validation Strictly Causal (< t → t)

To validate whether we can predict crime locations day-by-day without future leakage, for each day t we use only crimes from days < t to rank all city grid cells, then check where actual crimes occurred on day t.

Mean Out-of-Time ROC-AUC

Mean PAI @ 10% Area

Crimes in Top 10% Cells

--%

Crimes in Top 20% Cells

--%

PEI @ 10% (vs Oracle Ceiling)

--%

PEI @ 20% (vs Oracle Ceiling)

--%

Inspect Specific Validation Day on Map

Daily Backtest Log (Click a Day to Plot on Map)

Date	Crimes	Hit @ 10%	Hit @ 20%	PAI	ROC-AUC

Step-by-Step Optimization & Verified Gains Product Roadmap

Updated resultados.xlsx: 2016 vs. 2025 vs. 2026

Your 5 Master's algorithms (Naive Bayes, KNN, GMM, Random Forest, XGBoost/GB) executed across all 5 thesis subsets on last year's (2025) and 2026 Chicago data:

Dataset Subset	NB	KNN	GMM	RF	XGB

Empirical Proof: Post-Incident Target Leakage resultados.xlsx Audit

In resultados.xlsx, Random Forest & XGBoost achieved 80% accuracy on THEFT vs BATTERY. Why? Because Domestic and Location Description (which are only known after the crime is reported) accounted for 58.0% of tree importance!

With Leaky Features

79.3%

Pre-Incident Only (Clean)

66.0%

Leakage Inflation

+13.3%

Majority Baseline (Theft)

54.1%

5 Flawed Assumptions in Original Project & How We Fixed Them

B2B Commercial Risk & Underwriting API License Key Auth

Sell high-precision risk scores directly to P&C Insurance Carriers (Actuarial hazard multipliers & loss ratios) and Commercial Real Estate (CRE) developers (Retail & office site selection, parking risk, and ambient foot-traffic exposure).

Select Active License Key

Latitude

Longitude

Asset / Property Use

Underwriting Risk Scorecard

Actuarial Loss Multiplier

1.00x

Property Safety Score

Hazard Tier

Foot-Traffic Density

Direct cURL Code for Underwriting Engine:

Tactical Police Patrol Allocation Engine Counterfactual Simulator

Simulates dynamic automotive fleet positioning for Chicago CPD & NYPD. Compares traditional fixed beat centroids against Hawkes Spatio-Temporal Hazard Staging, benchmarked on held-out 911 calls.

Police Department Agency / Metro

Police Shift

Priority Focus

Available Automotive Patrol Fleet: 25 Squad Cars