Predicting New Home Sales with Google Trends

Main Question:

New Home Sales and Recessions

Monthly new home metrics (the number of permits, starts, completions, sales, etc.) are largely considered useful economic indicators. How well does Google Trends search data for new homes correlate to actual data? Can Google Trends data serve as a leading economic indicator sooner than standard monthly reports?

Data

Monthly single-family home sales data are available from the U.S. Census Bureau in cooperation with the Department of Housing and Urban Development. Data for this research project includes seasonally un-adjusted observations from January 2004 through February 2018.
Google Trends seasonally un-adjusted data are gathered from January 2004 through February 2018. The main search variable: ‘new homes for sale’.

Literature Review

(Choi and Varian, 2011), provide a summary of how Google Trends data are aggregated and reported, demonstrating several applications (motor vehicle sales, unemployment benefits, consumer confidence), and results that suggest Google search queries improve predictive modeling performance.
(Brynjolfsson and Wu, 2015), apply Choi and Varian (2011) framework in detail to new home sales, determining, “that a one-percentage point increase in search frequency about real estate agents is associated with selling an additional 3,520 future quarterly housing sales in the average US state”.
(Meese and Rogoff, 1983), provides an evaluation of the out-of-sample forecasting accuracy for exchange rate models, serving as the framework for this study.

Project Components

The study occurs in three stages: (1) examining and transforming data for stationarity, (2) modeling various time horizons following the Meese and Rogoff (1983) methodology, and (3) evaluating predictive modeling results via Root Mean Squared Error (RMSE).

Initial Time Series Plotting

Checking for stationarity

## 
##  Augmented Dickey-Fuller Test
## 
## data:  NewSalesFRED.TS
## Dickey-Fuller = -1.4544, Lag order = 5, p-value = 0.8041
## alternative hypothesis: stationary

H0: There is a unit root for the series, non-stationary / Ha: There is no unit root for the series, stationary / With p-value greater than 0.05, we cannot reject the null hypothesis and the series is non-stationary

## 
##  Augmented Dickey-Fuller Test
## 
## data:  NewHomesGOOGLE.TS
## Dickey-Fuller = -2.5231, Lag order = 5, p-value = 0.3579
## alternative hypothesis: stationary

H0: There is a unit root for the series, non-stationary / Ha: There is no unit root for the series, stationary / With p-value greater than 0.05, we cannot reject the null hypothesis and the series is non-stationary

Seasonal differencing

Checking again for stationarity

## 
##  Augmented Dickey-Fuller Test
## 
## data:  NewSalesFRED.TSD
## Dickey-Fuller = -7.4592, Lag order = 5, p-value = 0.01
## alternative hypothesis: stationary

H0: There is a unit root for the series, non-stationary / Ha: There is no unit root for the series, stationary / With a p-value less than 0.05, wereject the null hypothesis in favor of the alternative hypothesis that the series is stationary.

## 
##  Augmented Dickey-Fuller Test
## 
## data:  NewHomesGOOGLE.TSD
## Dickey-Fuller = -6.8616, Lag order = 5, p-value = 0.01
## alternative hypothesis: stationary

H0: There is a unit root for the series, non-stationary / Ha: There is no unit root for the series, stationary / With a p-value less than 0.05, we reject the null hypothesis in favor of the alternative hypothesis that the series is stationary.