Methodology

What goes into every Sialim forecast, and where the approach can break.

The target

We forecast the Zillow Home Value Index (ZHVI) for each of 894 US metros: all home types, middle price tier (33rd–67th percentile), smoothed and seasonally adjusted, monthly. ZHVI measures the typical home value in a market, not the median sale price, which makes it less noisy and better suited to forecasting.

The model

Each metro gets its own statistical model fit on up to 26 years of monthly history. We fit two versions per metro: a price-only model, and a version augmented with macro conditions. The macro version uses the 30-year mortgage rate (Freddie Mac), national unemployment, national wage growth, and metro-level for-sale inventory change. Each regressor is lagged one month, so a forecast for July 2026 only uses data that existed before that month began. No look-ahead.

Three horizons

Each metro's model produces three forecasts from one fit: next month, quarter-end, and one year ahead. Beyond the first month, macro regressors are held at their latest observed value, so the far horizons lean mostly on each metro's trend. Uncertainty bands widen with the horizon; treat the year-ahead number as a direction and magnitude estimate, not a target.

Two model classes, one referee

Each metro is forecast by competing proprietary approaches. One class models each metro's own history in isolation. Another learns across all metros jointly, drawing on leading indicators that move before prices do: days on market, the share of listings with a price cut, sales volume, Zillow's Market Heat Index, observed rents, and each metro's price-to-rent balance versus its own five-year norm. Learning across markets lets small metros borrow strength from patterns observed in roughly 200,000 metro-months of history.

Both models are scored at the same backtest origin, 13 months back: each predicts the following month, quarter, and year from identical information, and whichever came closer to what actually happened wins the right to publish that metro's forecast at that horizon.

Model selection and backtesting

Within each approach, variants with and without macro conditions are compared on a six-month holdout, and the better one advances. The backtest error shown on each metro page is that winner's average miss over the window. Confidence ranges are estimated from the distribution of outcomes, not from symmetric assumptions.

Regime labels

The cooling / stabilizing / recovering / overheating / neutral badge is a rule-based label. Where Zillow publishes them, it is driven by market-condition data that moves before prices do: the Market Heat Index and its three-month change, the change in the share of listings with a price cut, and the change in days on market, combined with price momentum. Metros without heat data fall back to momentum-only rules. The opportunity map's sell-side pressure component uses price-cut share for the same reason, with inventory as the fallback.

Limitations, stated plainly

The backtest window is a single six-month period, which understates how wrong the model can be when conditions shift. ZHVI is itself a smoothed index, so month-to-month errors look small compared to raw sale prices. Confidence bands come from the model and tend to be too narrow during rate shocks or policy changes. National unemployment and wages stand in for metro-level labor data, so a local plant closure or boom will not show up until it moves prices. Use these forecasts as one input among several, not as a decision.

Data sources

Home values: Zillow Research ZHVI (© Zillow). Mortgage rates: Freddie Mac Primary Mortgage Market Survey via FRED. Unemployment and wages: US Bureau of Labor Statistics via FRED. Inventory: Zillow for-sale inventory, metro level. Forecasts refresh monthly when Zillow publishes new data.