This post is a summary of a three-part series on now-casting the MBA Weekly Mortgage Applications Index. To read more detail, see parts one, two and three.

Authoritative macroeconomic data are typically published with a (sometimes substantial) lag, leaving policymakers and consumers to base real-time decisions about the present and future on past information. However, over the past two decades, consumers have increasingly looked to the internet for advice and information when contemplating major purchases or investments. A growing body of research literature recognizes that, when aggregated, these individual revelations of consumer intent contain valuable information about the state of the economy.  They can be a particularly useful input for estimating economic data in advance of their publication—an exercise that is commonly referred to as “now-casting”, the application of statistical methods commonly used to predict the future to assess the present.[1]

Shopping for a home has become one of the most common purchases that consumers research online. According to the National Association of Realtors’ (NAR) 2013 Profile of Home Buyers and Sellers, 92 percent of home buyers use the internet in their home search.[2] The share of home buyers who used the internet to buy a mortgage was much lower, but still, about one in ten home buyers searched or applied for a mortgage online with first-time and young home buyers doing so more frequently. Housing and mortgage data would appear to be particularly well-suited to forecasts using internet searches.

A number of prior studies have attempted to now-cast changes in home prices, mortgage refinancing applications, and other housing-related time series using Google queries.[3],[4] Typically, these now-casts rely on lagged values of the dependent variable, an index capturing the number of searches for a specific word or phrase, and other explanatory variables such as exogenous market data. In most cases, prior values of the variable being estimated account for the largest portion of the explanatory power, while the index of search terms yields modest but significant improvements.

Loan requests on the Zillow Mortgage Marketplace (ZMM), a free online marketplace that connects thousands of mortgage borrowers and lenders each day, could provide similar predictive power for mortgage market time series—in particular, the Mortgage Bankers Association’s (MBA) Weekly Mortgage Applications Index. Of course, data tracking the number of searches on any one particular marketplace (even if a market leader) suffers from its own liabilities—the vagaries of market share can introduce trends that tell us little about underlying economic activity and minor changes to website design can create otherwise unexplainable discontinuities in consumer behavior. However, with due skepticism, we demonstrate how data on Zillow loan request provide valuable, real-time insight into the state of mortgage markets.

In this three-part series, we test the added value of Zillow loan requests as an input to now-casting the MBA Weekly Mortgage Applications Index.

• The first analysis compares the forecast reliability of several models, concluding that the Zillow data contain substantial predictive information to estimate the MBA indexes.
• The second analysis compares how these models perform using shorter time series of Zillow loan requests as the explanatory variable, finding that partial-week Zillow data still contain a strong advance signal about the direction of the MBA indexes.
• The third analysis compares forecasts using Zillow loan requests to forecasts using the Google Trends search index, demonstrating that now-casts using the Zillow data outperform similar models using Google data.

As transactional and internet-based data become more widely available, there are growing efforts to apply this information to gain insight into the state of the economy. There have been both notable successes, and highly-visible failures. The real world is messy and real-time data typically do not have the convenient statistical properties of carefully designed surveys or rigorously collected accounts. But they do tell us something. In several related analyses, we have looked into how internet searches can provide real-time insight into the state of mortgage markets, and can be used to forecast the MBA Weekly Mortgage Application Index.

A relatively simple model—which uses a lagged value of the index, the contemporaneous weekly percent change in loan requests on the Zillow Mortgage Marketplace (ZMM), and banking holiday dummies—performs reasonably well in “now-casting” the index. Like all models, however, it is not perfect. The best performing of the models tends to miss the actual level of the index by just under 6 percent on average for our sample covering 147 weeks; the worst performing of the models misses the actual level by just under 9 percent on average.

Before internet search data became widely available, policymakers and markets often relied on idiosyncratic indicators as their preferred real-time guides to the state of the economy. For instance, former Federal Reserve Chairman Alan Greenspan is reported to have closely followed scrap steel prices and men’s underwear sales, believing the two to lead other indicators of industrial activity and consumer sentiment respectively. In this light, internet searches can be viewed as a modern analogue to the real-time rules-of-thumb that decision-makers have always relied on to guide their decisions about the state of the economy.

[1] See for example Hyunyoung Choi and Hal Varian, “Predicting the Present with Google Trends,” University of California, Berkeley, December 2011.

[2] National Association of Realtors, 2013 Profile of Home Buyers and Sellers, Washington, DC, November 2013.

[3] See for example, Lynn Wu and Erik Brynjolfsson, “The Future of Prediction: How Google Searches Foreshadow Housing Prices and Sales,” Technical report, Massachusetts Institute of Technology, 2010; Rajendra Kulkarni, Kingsley E. Haynes, Roger R. Stough, and Jean H.P. Paelinck, “Forecasting Housing Prices with Google Econometrics,” George Mason University School of Public Policy, Research Paper No. 2009-10, November 11, 2009; and Rebecca Hellerstein and Menno Middeldorp, “Forecasting with Internet Search Data,” Federal Reserve Bank of New York, Liberty Street Economics, January 4, 2012.

[4] Google Trends publishes indexes of the relative importance of a various search terms by geography and month. To our knowledge, the second and third leading internet search sites—Bing and Ask—do not publish similar information.