Validation

Model Performance

True out-of-sample accuracy on held-out future periods — the model never sees tomorrow's data until tomorrow.

T+1 Horizon  ·  2022–2024 Walk-Forward Test Folds
85.2%
Pre-Appreciation Signal
AUC · fragile tracts → T+1
68.6%
Growth Timing Signal
AUC · stable tracts → T+1
1–5 Years
Prediction Horizon
T+1 through T+5 labels
Walk-Forward Validation
Fold 1
Learns 2013 — 2021
→
2022 ?
Fold 2
Learns 2013 — 2022
→
2023 ?
Fold 3
Learns 2013 — 2023
→
2024 ?
2020–2021 excluded: pandemic-era data collection disruptions and stimulus distortions invalidate signal continuity in those years.
Every metric reported is a genuine blind prediction — the model never peeks at future data.
The industry problem

Most predictive models are benchmarked by randomly mixing past and future data. This lets the model effectively "peek" at answers — inflating apparent accuracy to 90%+ that collapses in production.

Our approach

Walk-forward validation forces the model to make genuine forward predictions. It only sees the past and must predict what happens next — the exact conditions faced in deployment.

85.2% AUC on data the model has never seen

Those are the numbers you can rely on for capital allocation decisions — not backtest artifacts.

Signal Architecture

Key Signal Layers

Five independent signal domains, each trained in isolation and stacked by the meta-learner.

Signal Layer Coverage What It Captures
Construction Activity 43 cities  ·  154K tract-years Investment intent 2–3 years ahead of market repricing
Local Market Conditions 43 cities Vacancy, blight, and local service quality trends
Demographic Trajectory 67K tracts Population composition, income, and labor force dynamics
Regional Economics 120 counties County-level macro context and economic cycle position
Affordable Housing Risk National Subsidy program expiry and area-level affordability pressure
Government Data Foundation
Walk-Forward CV
Fairness Calibrated · DI < 1.4
T+1 to T+5 Horizons
9 Active Models
COVID Excluded

See the signals for your portfolio

We'll prepare a metro-specific accuracy breakdown relevant to your investment thesis.

Request a Demo