Explainability

How we grade properties

A strict rubric with bounded weights. AI is only used for the location adjustment.

Overview

We grade facilities with a deterministic, additive rubric. Scores start at 50, then earn or lose points from build year, access model, climate coverage, unit count, and location tagging. Location gets a bounded AI adjustment, while every other input is deterministic. Each driver is capped so no single input can dominate. Every score has a receipt with drivers and rubric versioning.

Location module (separate from quality)

Nearby ≠ comp

Zip is for filtering. True comps rely on precise location and trade area context: lat/lng, geohash for nearby queries, market/submarket tags, and normalized address. “Nearby” does not automatically mean “comparable.” Submarket fit + quality + unit-type match make a comp, not just distance.

Location signals

Normalized address, lat/lng, geohash, market/submarket tags for clustering.

AI location adjustment

AI provides a bounded location + market-fit delta with an explanation saved to “Why this value.”

Context-aware

“Nearby ≠ comp.” Access model, climate, and unit mix must align for a true comp set.

Quality Index

A single, legible Quality Index (0–100) with rubric versioning (rubricVersion: v1.1). Scores are deterministic, derived from strict inputs, and never inflated by narrative text. Location AI adds a bounded adjustment only.

Example Quality IndexrubricVersion: v1.1
78/100
Every score has a receipt: we store the drivers and weights behind the Index so you can defend decisions and audit changes.

Category scores (0–100 each)

Five categories roll up into the Quality Index. Weights are bounded and versioned for consistency across facilities.

Product Quality82/100

Build year + climate coverage with strict deductions for missing inputs.

Convenience76/100

Access model emphasis (drive-up vs indoor) with bounded impact.

Security/Trust74/100

Proxy from build year + indoor access (no marketing claims).

Unit Mix Fit70/100

Unit count bands only; missing unit count is penalized.

Location Quality78/100

Lat/lng + submarket tagging plus bounded AI market-fit adjustment.

Unit-type specific adjustments

The rubric is strict on missing data. If critical fields are absent, we apply automatic deductions.

Missing build year
-12 pts
Missing lat/lng
-10 pts
Access model unknown
-8 pts

Comp Score (beyond distance)

Comp Score blends distance + submarket fit + quality delta + unit-type match + access model match + climate match. Distance alone is not enough.

1.2 miComp score 89

Example Comp Explanation

Why comp

  • Same submarket tag
  • Climate/indoor match
  • Drive-up mix within 5%

Why not

  • Slightly newer build date
  • Access hours differ (24/7 vs 9-9)

Estimated sell rate (Est. Sell)

Modeled

Est. Sell is a deterministic model derived from scraped web rates plus facility context. It is designed to estimate likely sell-through pricing, not to replace observed web rates.

Model formula
baseWeb = scraped web rate
upliftRaw = sizeBase + configAdj + marketAdj + reliabilityAdj
uplift = clamp(upliftRaw, 0.06, 0.22)
estSell = baseWeb × (1 + uplift)
estSellPerSqft = estSell ÷ sqft (when sqft is known)
What drives the estimate
  • Unit size bucket (smaller units typically carry higher sell-through uplift).
  • Unit configuration (climate and indoor/outdoor adjustments).
  • Market context (Quality Index and bounded location AI delta).
  • Reliability context (freshness of scrape and confidence score).
Why we trust this output
  • Deterministic: the same inputs always produce the same estimate.
  • Bounded: uplift is capped between 6% and 22% to avoid unrealistic jumps.
  • Signal-based: combines product, location, and data-quality inputs instead of a fixed multiplier.
  • Transparent: web rate remains visible, and Est. Sell is always labeled as modeled.

Estimated occupancy

Hybrid model

Estimated occupancy blends listing-based occupancy evidence with a deterministic reliability-scaled model. It is directional and not PMS truth unless PMS is explicitly connected and shown.

Model formula
listingRange = listing-based occupancy from availability + velocity
modelPoint = fundamentals + pricing + promos + competition + local modifiers
modelRange = clamp(modelPoint ± halfWidth)
listingWeight = f(coverageScore, freshnessScore, volatilityScore)
finalRange = (listingWeight × listingRange) + ((1-listingWeight) × modelRange)
What drives the estimate
  • Competition and overbuild: supply pressure balanced against local demand density with conditional penalties.
  • Seasonality and local shocks: snowbirds, universities, military PCS, lease-ups, turnover, disasters, employer shocks, remodeling.
  • Promos and pricing position: interpreted as demand pressure with anti-double-count safeguards.
  • Facility fundamentals: quality, access, climate mix, unit count, and age priors.
Why we trust this output
  • Deterministic: same inputs always yield the same output.
  • Bounded: point/range are clamped with per-domain budgets and top-3 seasonality control.
  • Uncertainty-first: stale or thin coverage lowers listing weight and widens range width.
  • Transparent: blend math, active modifiers, applied caps, and quality scores are shown in explainability views.

FAQ

No. Distance is one factor. Comp Score blends submarket fit, quality delta, unit-type match, access model, and climate alignment.