ChrionML® Research StandardVersion 1.0 · October 2026
ChrionML® CML logoChrionML® AI LabsResearch & Insights
Methods and reproducibility

One protocol. Local model selection.

ChrionML® uses a consistent research workflow to transform published market histories into quality-checked datasets, out-of-sample model comparisons and decision-ready findings.

Research lead: Olu (Tim) DaramolaCurrent protocol: ZIP-level condo/co-op forecastingEvaluation: Expanding-window backtesting

Research goal

To evaluate whether a transparent, repeatable forecasting protocol can produce comparable short-horizon evidence across heterogeneous local housing markets while allowing the best-performing model to differ by city.

Data and sources

The underlying historical observations are derived from Zillow Home Value Index (ZHVI) market pages. The relevant ZIP-level condo/co-op series is selected by geography and property type. Zillow’s default headline value may represent an all-homes series and is therefore not assumed to match the series under study.

ChrionML® independently compiles the monthly observations used in the analysis, verifies the date sequence, checks for missing or duplicated periods, and preserves source links in each publication.

Research workflow

Qualify the question

Define geography, property type, forecast horizon and comparison measures.

Locate the source series

Select the corresponding Zillow ZHVI view and document its public URL.

Construct the dataset

Compile monthly values into a structured city-level time series.

Perform quality assurance

Check sample size, chronology, missingness, duplicates and the latest observation.

Run chronological backtests

Evaluate every candidate using only information available before each prediction date.

Select locally

Choose the lowest-error model independently for each market.

Quantify uncertainty

Use rolling residual behavior to form an empirical prediction range.

Publish transparently

Separate source facts, modeled results, interpretation and limitations.

Candidate models

The current protocol compares five interpretable forecasting candidates: Last Value, 12-month Seasonal Naive, Drift, 36-month Linear Trend and Holt Damped Trend. A more complex candidate is selected only when it improves out-of-sample performance.

Evaluation design

Each completed city study contains 80 monthly observations. The first 36 observations establish the initial training window; the remaining 44 months support expanding-window, one-step-ahead backtests. Mean absolute error (MAE) is the primary selection measure because it remains interpretable in the same dollar units as the index.

Anti-leakage rule. No observation from the prediction month or any later month is available to the model when producing a historical backtest forecast.

Limitations

  • ZHVI is a market index, not an appraisal of an individual property.
  • A ZIP-level condo/co-op series can conceal meaningful variation across neighborhoods and buildings.
  • Short-horizon statistical forecasts do not directly model interest rates, inventory, regulation or local economic shocks.
  • The lowest historical MAE does not guarantee the smallest future error.
  • Comparisons describe the selected measures and should not be read as universal city rankings.

Data and code availability

Each publication links to the public Zillow market pages used to identify the underlying series and to the corresponding ChrionML® box score. Research artifacts are versioned through the public GitHub repository as they are prepared for release.

Research archive →GitHub repository →