Zestimate just got a makeover – it’s smarter and sees better. Incorporating ideas from the Zillow Prize, it’s now more accurate and “half of all Zestimates fall within 2% of the home’s eventual sale price.” That accuracy comes with a caveat. It is for Zestimates of for-sale properties. For homes that are not listed for sale, the error is 7.7%. Nevertheless, for home buyers, Zestimates give a good, usable idea of current house values.
But what about future house values? The Zestimate forecast predicts the change in the Zestimate over the next 12 months based on past Zestimate values and the Zillow Home Value Forecast. The Zillow Home Value Forecast forecasts the change in the Zillow Home Value Index (ZHVI) over the next 12 months. And all this boils down to: improved forecasts of the Zillow Home Value Index power better (Z)estimates of future house values.
Improving Zillow Home Value Index forecasts
The HousingIQ Housing Market Vitality Indicators (HMVI) gauge the impact of local economic conditions on future house prices for over 400 metro areas. Incorporating the HMVI into a model for forecasting the ZHVI improves both statistical model quality and forecast accuracy. The table below summarizes the results. In all but seven metro areas, there is improvement on both counts — statistical quality and forecast accuracy.
Impact classification combines the Model Quality and Forecast Accuracy improvements into a single metric.
The N and % Metros columns identify the size of the subset of metros that are contained in each row.
Model Quality is the median improvement (decrease) in BIC (Bayesian Information Criteria) between a model that does not incorporate HMVI and one that does. Setting aside technicalities, an improvement greater than 10 is considered as Very Strong and a value between 6 and 10 as Strong evidence in favor of the model that incorporates HMVI. Values greater than 2 indicate Positive evidence.
Forecast Accuracy is the median reduction (decrease) in MAE (mean absolute error) between a model that does not incorporate HMVI and one that does. The error of a forecast is the difference between the predicted and actual values. Ex. A prediction of 10 for an actual value of 12 yields an error of 12 – 10 = 2. The greater the reduction in absolute error, the closer the predicted value to the actual value.
Impact translates the Forecast Accuracy improvement into dollars. Using the median home price of $226,800 and a median national error of 7.61% (weighted average error of both off- and on-market homes), we calculate the dollar value of the improvement in forecast accuracy. Ex. The 4% decrease in error for the ‘C’ Impact classification brings the forecasted price $691 closer to the actual price.
The crosstabs below shows the distribution of the BIC Improvement and % MAE Reduction. It reinforces the possibility to gain substantial improvements in statistical model quality and reduce forecast error by incorporating HMVI. The cells are color coded to track the Impact classification above.
The BIC Improvement categories summarize the statistical evidence in favor of incorporating HMVI as a predictor. Very Strong corresponds to a greater than 10 point improvement (decrease) in BIC; Strong to a 6 – 10 point improvement; and Positive to a 2 – 6 point improvement. Improvements less than 2 points offer no evidence and negative improvements identify a detrimental impact as is the case for 1.1% of the metros in this analysis.
The % MAE Reduction categories reflect the practical significance of the error reduction. An Excellent reduction corresponds to greater than 40% decrease in error and Good to a 10 – 40% decrease. Error reduction less than 10% is Incremental insofar as it likely needs something additional to be valuable.
Bottom line: Forecasts of Zillow Home Value Index are improved by incorporating Housing Market Vitality Indicators. In over 75% of the metros, the forecast improvement is statistically significant and has practical value. Such improved metro forecasts can significantly enhance (Z)estimates of future house values. Download the data underlying this analysis.
Contact us to explore how HMVI can be blended with Automated Valuation Model (AVM) systems and other house price forecast engines to produce better predictions of house values.
We used the ZHVI Single Family Homes and HMVI Situation data through May 2019 for 370 metro areas. Metro areas comprised of Metropolitan Divisions were mapped to the largest constituent Metropolitan Division as HousingIQ tracks the more granular Metropolitan Divisions.
The ZHVI data was transformed to a y-o-y change. All data were confirmed to be stationary and pre-whitened to account for auto-correlation. To justify use of HMVI Situation, Granger causality between HMVI Situation and ZHVI was tested. In 82% of the cases there was Granger causality at the 5% significance level with a median lag order of 10 months.
Two ARIMA models were evaluated for each metro area. A univariate model and another model that included lagged HMVI Situation as an exogenous input. The two models were compared using BIC for model quality and MAE for accuracy. The results were expressed such that positive values correspond to an improvement over the univariate model.