Based on our historical model of new home sales, Zillow expects sales of new, single-family homes to fall by 3.1 percent to a seasonally adjusted annual rate (SAAR) of 488,000 units in September. Our structural model suggests a slightly smaller decline, down 1.8 percent to 495,000 units (SAAR). (For the differences between the models, see here.)
The historical model has the highest accuracy, but the structural model provides unique insights into the market forces driving new home sales. As with any economic model, our results are sensitive to our assumptions. Our structural model requires that we estimate several key input variables as much as three months ahead, including:
The baseline model assumes no change in the homeowner vacancy rate since June. The homeowner vacancy rate increased sharply during the recession, but has since fallen and is approaching its pre-crisis levels. The vacancy rate increased during the second half of last year, but declined between January and June of this year. Inventory data, which are closely correlated to vacancies, suggest that vacancies could increase, but strong existing home sales data for September point in the opposite direction. In light of these mixed signals from incoming data, holding the homeowner vacancy rate constant at its June level is a safe assumption.
The baseline model assumes that the homeownership rate increased slightly over the summer, from 64.8 percent in June to 65.1 percent in September. A downward trend in the homeownership rate since the peak of the housing bubble has substantially held back housing turnover in the post-crisis period, but the homeownership rate began to increase modestly in the second quarter of this year.
This recent trend, combined with the recent uptick in new and existing home sales, and our skepticism that the homeowner vacancy rate has dropped dramatically, suggest that the homeownership rate has likely increased since June. Higher sales could coincide with a lower homeownership rate if the vacancy rate were to fall, but given the signal from inventory data that the vacancy rate has at least been flat, higher sales should indicate upward movement in the homeownership rate.
If we assume that the homeownership rate increased even more strongly, to 65.3 percent—in line with a best-fit ARIMA(2,1,4) function on the homeownership series for December 2004 through June 2014—then new home sales would have totaled 519,000 in September, up 3.1 percent from August. By contrast, if we assume the homeownership rate remained constant at its June level, then new home sales would have totaled 458,000 in September, down 9.1 percent from August.
In the baseline model, we assume median family income grew 3.2 percent in September, in line with the change in the Consumer Price Index for services, as discussed in the assumptions underlying our forecast for September existing home sales.
The baseline model considers two separate changes in the median sales price of new homes. Sales prices for new homes have been trending upward since the end of the recession, but fell over the summer—from $286,000 in June to $276,000 in August. If we extrapolate this recent trend forward, then the median sales prices of new homes would have continued declining to $279,000 in September.
If we find a best-fit ARIMA function for the series—an ARIMA(2,1,3) process, in this case—then the median sales price of new homes would have increased slightly to $282,000. We also assume a more modest increase to $279,000.
In the former case, where the median sales price of new homes rose to $279,000, new home sales would have totaled 492,000 in September, down 2.4 percent from August. If the median sales price rose more strongly to $282,000, then September new home sales would have totaled 495,000, down 1.8 percent from August.
We forecast the percent of loans in foreclosure three months ahead, as these data are only currently available through June. In this case, the best fit ARIMA model is an ARIMA(3,1,4), which produces an estimate of 2.39 percent of loans in foreclosure, down from 2.49 percent in June 2014. Although this variable has larger effect on the model, this is a reasonable estimate given that loans in default have declined steadily since 2011, with a three-month average change of -0.15 percentage point.
Finally, we forecast the number of households three months ahead using a best-fit ARIMA model, estimated over the period April 1976 through June 2014—in this case, an ARIMA(0,2,1) process. This produces an estimate of 115.297 million households in September 2014, up 0.2 percent from 115.097 million in June.