Forecasting Recession; An Attempt to Identify Predictors

I attempted to identify variable (s) that has predictive power for economic growth/decline. Real Gross Domestic Product (hereafter referred to as RGDP) for the United States is the object of forecast. Plot of this observation can be found in Annex 1—Figure 1. This variable is reportedly a perfect manifestation of an economy’s vitality[1]

Pool of possible predictor variables include:

a.      Real Final Sales of Domestic Product (hereafter referred to as Sales):  During recessions, evidence suggests, that consumer confidence declines. Consumers tend to spend less, not only on durable goods but on non-durable goods[2]. This data set measures percent change in quarterly sales of domestic product compared to one year ago and is seasonally adjusted. Plot of this observation can be found in Annex 1—Figure 4.

b.     10 year treasury constant maturity minus federal funds rate (hereafter referred to as treasury): The slope of the yield curve is the difference between long-term and short-term interest rates. Dataset in this exercise measures this difference for 10 years against three months. Average spread, reportedly, becomes negative before recessions[3]. Plot of this observation can be found in Annex 1—Figure 3.

c.      Average Hourly Worked (hereafter referred to as hourly worked): Intuitively, this variable should have positive relationship with that of that of recession. As economies contract and businesses slow down, employees are either separated or are employed for lesser amount of time. Plot of this observation can be found in Annex 1—Figure 2.

The intention was to check whether or not these variables signal slowing economic activity. If so, which one has the strongest predictive power. Can combination of all or some of them make a more reliable predictor pool?


In the exercised, I utilized data on the United States’ macro and microeconomic variables from 1970 up to 2011. Using R, initially datasets were smoothened and tested. Projections were based on Vector Autoregression. Vector Autoregression allows for forecasting with multiple variables. The idea is to see if the predictor variables can explain variances in the Real Gross Domestic Product and if past dynamics of the forecast object holds explanatory power. Once the model is specified, association and direction of association of each predictor variable with the forecast object is being tested, using the granger causality test.

          Using suitable Vector Autoregressive Models, an actual forecasting was undertaken as a recursive, one-to-four step ahead forecast approach. Then, each model’s superiority was assessed by looking at their mean squared error, whether or not one encompasses the other, and whether forecasts are unbiased and optimal or not.

Model Selection

Four models were specified for this exercise. These models are;

a)     A Trivariate Model that includes RGDP, treasury, and sales;

b)     Bivariate 1 using RGDP and sales;

c)      Bivariate 2 using RGDP and treasury;

d)     Bivariate 3 using hourly worked.

Estimates of the multivariate models made it evident that past lags of neither of the endogenous variables (sales, treasury, and hourly worked) could explain future variations in RGDP. To confirm this, here, a recursive forecasting for one-four step ahead was undertaken.

     The dataset was divided into two disproportionate parts: original sample size consists of 26 observations-up to 1997. Estimates of this was then used to project an out of sample forecast up to 2011. 
     It was observed that all the four models were quite away from trajectory of RGDP. There were stark differences between projections of the models and levels of RGDP growth. 


This exercise, while, strengthens current findings of relative predictability of economic growth/trough using the yield curve/spread, it also indicates that average hourly worked and real sales of gross domestic product have no predictive power—whatsoever. 

Note: This post contains select sections of an elaborate exercise. To obtain the document in full (that includes R codes, too), contact me via .

[1] Glenn D. Rudebusch, John C. Williams, 2007. “Forecasting Recessions: The Puzzle of the Enduring Power of the Yield Curve”.
[2] Ivaylo D. Petev, lSQ-CReSt & luIgI PIStafeRRI, StanfoRD . 2012. “Consumption in the Great Recession”
[3] Arturo Estrella and Mary R. Trubi. 2006. “The Yield Curve as a Leading Indicator: Some Practical Issues”, Federal Reserve Bank of New York, Volume 12, Number 5 July/August 2006.


Post a Comment

My Twitter Feed