Market intelligence

Don't just find a home.
Find the return.

Signal Alpha predicts which segments of the New York market are mispriced relative to where demand is actually heading — and projects risk-adjusted returns before the closed comps catch up.

— the thesis

Closed comps are a lagging indicator.

By the time a segment shows up in the closed-sale data, the return is already priced in. Every buyer working from public comps is, by definition, late.

Signal Alpha reads demand as it forms — from behavioral signals no portal collects — and forecasts segment-level performance ahead of the market. The output isn't a listing recommendation; it's a ranked map of where the next leg of appreciation is most likely to land, with the risk and liquidity that come with it.

— how it works

A four-stage pipeline from raw signal to ranked return.

Two classes of data flow in. Derived features are computed per segment. An ensemble of supervised and time-series models produces a risk-adjusted return score with confidence intervals and per-prediction attribution. Output ships as scorecards and an API.

— architecture · fig. 1

Signal in, ranked return out.

Simplified architecture. Public-record ingestion is operational from day one; proprietary telemetry compounds with usage. SHAP-style feature attribution accompanies each segment score so every prediction can be explained, not just consumed.

— 01 · ingestion

Public + proprietary.

Two classes of data, fused at the segment level. The public layer is competitive parity. The proprietary layer is the wedge.

Public
  • ACRISdeeds, mortgages, transfer history
  • DOBpermits, violations, work-type density
  • 311 complaintsper-building and per-block density
  • RLS · StreetEasyclosed and active comps
  • NYC Open Datazoning and rezoning actions
  • Census · ACSdemographic shifts
  • MTA ridershipper-station weekly trends
  • Tax rollsassessed value trajectories
Proprietary
  • Swipe telemetryswipe · dwell · skip
  • Visit notesground-truth observations
  • Building sentimentscored per-building
  • Off-market sourcingseller-intent signals
  • Search-to-tour conversionby segment
— 02 · features

Derived per segment.

Raw inputs become structured features. Each segment carries a moving panel of metrics that feed both supervised and time-series models.

Standard metrics
  • Absorption rateunits cleared per period
  • Price velocitylist-to-close drift
  • Supply / demand imbalanceactive inventory vs. tour volume
  • Days-on-marketfull distribution, not just median
  • Concession-adjusted effective pricenet of credits, buy-downs, included furnishings
Proprietary metric
  • Demand Intent Indexcomposite of swipe telemetry, dwell time, search-to-tour conversion, and visit-note sentiment. The signal that leads closed comps.
— 03 · modeling

Ensemble + leading indicator.

No single model carries the prediction. The output is an ensemble vote with a forecast horizon, a confidence interval, and a feature-attribution trace.

What runs
  • Gradient-boosted ensemblesXGBoost / LightGBM for segment-level price and appreciation
  • Time-series forecastingabsorption and demand trajectory
  • Demand leading-indicator modelswipe-level intent empirically leads closed comps — a forecast window measured in weeks
  • Geospatial clusteringsegments defined dynamically by neighborhood × building type × price band, not fixed zip codes
  • Risk-adjusted return scoreprojected appreciation discounted by predicted volatility, carrying cost, and expected time-on-market
  • SHAP-style attributionevery segment score ships with the why, not just the number
— 04 · output

Ranked + explainable.

The output is built to be acted on by an operator who has to defend the call. No black box, no single number without context.

Surface
  • Segment scorecardsranked, with confidence intervals
  • Opportunity alertstriggered when leading indicators diverge from closed comps
  • Driver breakdowntop features moving each score, per-prediction
  • Dashboard + APIread access, programmatic queries, exports
Built for review
  • Every score is auditableinputs, model version, attribution trace, timestamp
— the moat

v1 is useful. The flywheel makes it unreplicable.

Public-data v1 is already a sharper read than what most NY operators run on. The proprietary demand layer is what makes the system get better the more it gets used — a compounding dataset no portal can replicate.

The asymmetry is structural. Listing portals are seller-side businesses. Their revenue comes from agents and brokerages paying to surface inventory. They have no operational reason to collect demand-side ground truth — and if they tried, their seller customers would object to the asymmetry it would create. Demand telemetry is not a feature they are choosing not to build. It is data their business model prevents them from having.

Signal is buyer-side from inception. Every user interaction produces signal. Every visit produces ground truth. Every swipe sharpens the Demand Intent Index. The graph cannot be replicated retroactively — competitors arriving later face a dataset deficit that grows daily.

The flywheel. More users → more demand telemetry → better predictions → better outcomes for users → more users. The dataset is the moat. The moat compounds.
— honest framing

Residential real estate is illiquid and transaction-heavy. Signal Alpha forecasts relative segment opportunity and risk-adjusted ranges — not guaranteed returns. Every score ships with a confidence interval and a driver breakdown so the limits of the prediction are visible, not hidden. Past performance does not guarantee future outcomes. The system is a decision aid for sophisticated operators, not an automated underwriter.

See where the next leg lands.

Signal Alpha is one of sixteen interoperating subsystems behind Signal Homes. The investor brief covers the rest.