Architecture

Two-Tier Stacked Ensemble

Domain isolation prevents cross-contamination. The same stacking architecture class behind Netflix Prize and Airbnb's market pricing engine — applied to neighborhood-level prediction.

1

Domain Models (Layer 0)

Seven models train independently on separate proprietary signal domains. Each generates held-out out-of-fold predictions passed as meta-features to Layer 1. Cross-domain feature injection is forbidden — it violates stacking isolation.

2

Meta-Learner (Layer 1)

Stacks all domain signals into gentrification, growth, and structural decline probabilities. Two variants: lagged (full signal history, data ceiling 2024) and no-lag (near real-time signals only, ceiling 2026).

3

Temporal Causal Ordering

Different signal types lead neighborhood change by different margins. Domain lookback windows are calibrated to this causal cascade — each signal layer anchored at its validated lead time, with zero overlap between feature and label windows.

4

Leakage Prevention Framework

Home value velocity, income velocity, and proxy features excluded from classification features. Three leakage types enforced: temporal (feature overlaps label window), definitional (feature encodes label formula), and proxy (r > 0.7 with excluded outcomes).

Layer 0 · Domain Models
α anticipatoryforward-looking investment signals
β communitylocal conditions & service quality
γ structuralhousing stability indicators
δ demographicpopulation composition trajectory
ε regressioncontinuous appreciation targets
ζ regionalcounty-level macro context
η areaaffordability & area risk
↓   OOF predictions as meta-features   ↓
Layer 1 · Meta-Learner
generalizedall signals → 3 outcomes (ceiling 2024)
no_lagnear real-time only (ceiling 2026)
↓   probability scores   ↓
Outputs
gentrificationfragile tracts → appreciation above MSA
growthstable tracts → sustained appreciation
declinedemographic erosion signal
Validation
methodwalk-forward expanding window CV
test folds2022, 2023, 2024
excludedCOVID years 2020–2021
horizonsT+1 through T+5
Causal Ordering

Temporal Causal Cascade

Signal types lead neighborhood change by different margins. The ensemble is calibrated to this ordering — not flat feature averaging.

Earliest Signal
Investment Intent
2–3 years ahead
Construction activity and forward investment signals. First indicators that neighborhood trajectory is shifting before any market repricing is visible.
→
Intermediate Signal
Market Conditions
1–2 years ahead
Local vacancy, blight, and service quality trends. These amplify or attenuate the early investment signal and sharpen the trajectory forecast.
→
Concurrent Signal
Demographics
Concurrent / 1yr lag
Population composition, income, and labor force dynamics. Demographic trajectories confirm direction and magnitude of neighborhood change.

Each domain model is anchored at its validated lead time — zero overlap between feature and label windows.

Validation Design

Why walk-forward matters

The design choices that make these AUC numbers usable for capital decisions — not just benchmark artifacts.

DESIGN CHOICE 1
Expanding-window CV, not random splits
Train on all prior years, test on the next year — in chronological order. Models cannot access future data during training, matching the exact conditions of live deployment.
DESIGN CHOICE 2
COVID years excluded
2020 and 2021 excluded from label years. Pandemic-era collection disruptions, stimulus distortions, and eviction moratoriums produce signal breaks that are exogenous shocks, not structural change.
DESIGN CHOICE 3
Scaler fit per fold, not global
Feature scaling fitted on training data only, per fold. Global scaling leaks future distribution statistics into training — a subtle source of inflation common in published benchmarks.
DESIGN CHOICE 4
Stratified eligibility
Gentrification model trains on fragile tracts only. Growth model trains on non-fragile tracts only. Mixing populations conflates distinct signals and inflates apparent AUC by 25+ points.

What you get from this design

✓ AUC numbers that match what you'll see in production — not inflated by random-split artifacts
✓ Per-fold breakdown available — not just mean AUC, so you can see year-to-year stability
✓ Three leakage types explicitly prevented: temporal, definitional, and proxy correlation
✓ Fairness-calibrated: disparate impact below 1.4 threshold across protected demographic groups
✓ Drift baseline established — signals actively monitored for distribution shift post-deployment

Curious how the signal layers perform in your market?

We'll walk through the domain-level accuracy breakdown for your target metro areas.

Request a Demo