How accurate is AtlasTAP’s site selection?

Each model is rebuilt at past dates and scored on what Houston venues went on to do. Some inputs, such as population and parcels, exist only as they stand today. We publish the numbers whether they pass or fail, and we do not sell a ranking as validated until it passes.

How to read these numbers
Spearman rank correlation
From -1 to 1. 0 means our order says nothing about the outcome; 1 means it matches exactly.
AUC
From 0 to 1. 0.5 is a coin flip; 1 means every flagged venue closed before every unflagged one.
Share inside the range
From 0% to 100%. A P25 to P75 range should contain about half of outcomes, so the target is 50%.
Lift
A ratio. 1 is no better than the baseline named beside it; 2 is twice as good.

Do the cells we rank highest get the best new venues?

Site scorer (ranked res-8 candidates per concept), site-model-1.1, 8 origins, computed 2026-10-21

Passes its gate

Scored against: Mean monthly receipts in months 13 to 24 after the origin of venues first reporting after the origin, per concept and res-8 cell, over the concept's median.

The most recent checkable origin, 2025-01-01, uses the holdout window: months 13 to 20, receipts through September 2026. This row is separate from the eight-origin gate.

Top 10% precision
3.4%
Base rate
1.0%
Lift (ratio)
3.405
95% bootstrap interval
2.845 to 4.052
Spearman of the learned rank vs realised receipts, by origin
2021-01-010.317
2021-07-010.312
2022-01-010.269
2022-07-010.305
2023-01-010.354
2023-07-010.341
2024-01-010.338
2024-07-010.316

Baseline (composite (sports_bar)) at each origin: 2021-01 0.431, 2021-07 0.331, 2022-01 0.044, 2022-07 -0.103, 2023-01 -0.176, 2023-07 0.013, 2024-01 0.283, 2024-07 0.239

GateGot
pooled Spearman of learned score vs realised receiptsneeds ≥ 0.250.316 (passed)
bootstrap 90% CI lower bound of pooled Spearmanneeds > 0.10.285 (passed)
origins with positive Spearmanneeds ≥ 48 (passed)
pooled top-vs-bottom quintile liftneeds ≥ 1.53.109 (passed)
  • Openings are proxied by first receipts month; a re-permitted existing venue looks like an opening.
  • The 300 ft school buffer is carried as unverified (Tex. Alco. Bev. Code 109.33 text not fetched at build time).
  • Competition counts every NAICS 722 location with a cell that had started and had not closed by the origin (6957 placed).
  • 50 percent receipts range: walk-forward margin 1.019522 log1p around q50, mean leave-one-origin-out coverage 0.499 against 0.45 to 0.55.

Does the receipts range for a new venue contain what it went on to do?

Analog engine, catchment feature space, catchment-1.0, 4 origins, computed 2026-10-19

Fails its gate

Scored against: Mean monthly receipts in months 13 to 24 after the origin of venues first reporting after the origin.

Share inside the P25 to P75 range (target 50%), by origin
2023-01-0155.4%
2023-07-0152.0%
2024-01-0149.1%
2024-07-0145.9%
GateGot
log-MAE lift over global median, every originneeds ≥ 0.2-0.410 (failed)
published P25 to P75 coverageneeds 43% to 57%27.9% (failed)
walk-forward calibrated P25 to P75 coverageneeds 43% to 57%50.6% (passed)
log-MAE lift over global median, level-corrected estimate, every originneeds ≥ 0.20.043 (failed)
log-MAE lift over past-openings median, level-corrected estimate, every originneeds ≥ 0.20.029 (failed)
  • Catchment vectors come from today's hex metrics and contours; only the receipts side is as-of.
  • The calibrated band reads only outcomes whose window closed by each origin, but the estimates it scales are today's stored analog rows, so the estimate side carries the caveat above.

Do the venues we flag as at risk close more often?

Venue survival (landmark logistic) walk-forward, landmark-1.0 / closure-1.3, 4 origins, computed 2026-10-20

Fails its gate

Scored against: Closure within 24 months, labelled as of the horizon.

AUC of 24-month closure risk, by origin
2024-01-010.602
2024-07-010.605
GateGot
Landmark AUC of 24-month risk vs realised closure, every originneeds ≥ 0.650.602 (failed)
Calibrated Brier below prevalence baseline, share of originsneeds ≥ 100%0.0% (failed)
  • Neighbour, operator, violation and occupancy features are omitted: they were not keyed for history units.
  • The 3-month silence window and the successor window after the horizon are part of the outcome label, not the features.
  • Calibration is isotonic on the latest eligible landmark.
  • 2023-01-01: landmark model not fitted (not enough landmark events or features).

Does the gravity model predict a new venue's sales? (now a feature only)

Huff calibration (per concept), huff-0.2.0, 4 origins, computed 2026-10-19

Fails its gate

Scored against: Mean monthly receipts in months 13 to 24 after the origin of same-concept venues first reporting after the origin.

Log-error lift over the pool median, by origin
2023-01-01-0.146
2023-07-01-0.093
2024-01-01-0.076
2024-07-01-0.028

Baseline (pool median) at each origin: 2023-01 1.433, 2023-07 1.536, 2024-01 1.569, 2024-07 1.465

GateGot
log-error lift over pool median, every originneeds ≥ 0.2-0.146 (failed)
  • Travel times are today's contours; the road network is treated as fixed over the window.
  • Attractiveness is the sqrt of trailing receipts, so a new venue's attractiveness is unknown at the origin and set to the pool median.

Monthly ZIP forecasts, tested before we show them

Harris County only. Each forecast is tested on past months it never saw, and the results are listed here whether it passed or not.

Can we forecast building permits in a ZIP?

Passed its test

Closer than last year's pattern in 79% of ZIP checks

654 of 832 ZIP checks. A ZIP check is one ZIP at one test date, and ties are left out. The forecast range, built to hold about 8 in 10 months, held 83% (3,755 of 4,515 ZIP-months).

What we tested: Permits issued per month in 105 Houston ZIPs. At 8 past dates (2022-12 to 2026-02) we forecast the next six months using only what was known then, and compared with what happened. The yardstick is repeating the same month last year.

Evidence dated 2026-10-14.

Can we forecast recorded crime in a ZIP?

Passed its test

Closer than last year's pattern in 70% of ZIP checks

480 of 685 ZIP checks. A ZIP check is one ZIP at one test date, and ties are left out. The forecast range, built to hold about 8 in 10 months, held 78% (3,204 of 4,110 ZIP-months).

What we tested: Houston Police recorded incidents per month in 137 Houston ZIPs. At 5 past dates (2024-12 to 2025-12) we forecast the next six months using only what was known then, and compared with what happened. The yardstick is repeating the same month last year.

Evidence dated 2026-10-14.

Can we forecast new liquor licenses in a ZIP?

Passed its test

Closer than last year's pattern in 65% of ZIP checks

700 of 1,070 ZIP checks. A ZIP check is one ZIP at one test date, and ties are left out. The forecast range, built to hold about 8 in 10 months, held 86% (5,414 of 6,292 ZIP-months).

What we tested: Liquor licenses first issued per month in 121 Harris County ZIPs. At 9 past dates (2024-03 to 2026-03) we forecast the next six months using only what was known then, and compared with what happened. The yardstick is repeating the same month last year.

Evidence dated 2026-10-14.

Can we forecast license issues across every Harris County ZIP?

Not proven, switched off

Not shown: the test could not prove it

0 of 157 ZIPs could be shown to beat last year's pattern. A fair test needs dated copies of the source data as they stood at each past date, and those copies do not exist yet, so no ZIP could be scored. Showing the forecast without that proof would be a guess.

What we tested: All 157 Harris County ZIPs stayed in the count, with none dropped. The plan was 33 monthly dates (2023-01 to 2025-09), each looking up to 12 months ahead. This forecast stays switched off until the test can be run on real history.

Evidence dated 2026-10-10.

Rank your own venues

Enter venues you already operate and their sales. We place each one in its neighbourhood cell, read where our sports-bar model ranks that cell, and show how closely the model's order matches yours.