A residential value forecast has to account for differences between properties, changes over time, and conditions shared across a neighborhood. It also has to communicate how much confidence to place in the result.
Worldcastr produces a distribution of possible future values. The interface summarizes that distribution with lower, central, and higher estimates at each forecast year. This article explains what those estimates represent and how we evaluate them.
From property records to forecast paths
The published forecasting system combines county assessment records with demographic and economic information. Those inputs come from different sources, so they must be aligned by place and time before they can be used together.
The model represents a trend and generates possible trajectories around it. Working with trajectories allows the system to model relationships across forecast years. Shared spatial information also represents dynamics that nearby properties may have in common.
The central estimate is the median of the modeled outcomes for a particular year. The lower and higher estimates mark the boundaries of the middle 80 percent. Actual outcomes can fall outside that interval.
Evaluation includes accuracy and coverage
We evaluate forecasts from held-out origin years: points in the past from which the model must forecast forward. This lets us compare its predictions with outcomes that were not available at the forecast’s starting date.
Point accuracy measures how close the central estimate came to the outcome. Interval coverage asks how often outcomes fell inside the predicted band. We also compare the model with a persistence baseline that repeats the last known value.
Those checks answer different questions. A model can have a reasonably accurate center while its range is too narrow. Results also vary by geography and horizon, which is why the methodology reports more than one aggregate score.
What to keep in mind
The property data includes assessed values, which can lag or differ from transaction prices. A forecast trained on those records should be interpreted in that context. Less complete local data and changes in economic conditions can also affect performance.
The boundary of a multi-year band is not necessarily one continuous scenario. The lower estimate is calculated separately from the distribution at each year; different modeled trajectories can contribute those values.
For a practical reading, begin with a specific geography and year. Compare all three estimates, then consult the evaluation and source information relevant to that forecast. The full methodology includes the architecture, diagnostics, results, and known limitations.