Do credit booms predict US recessions?

• Date: 01 September 2020
• Published date: 01 September 2020
• Author: Marius M. Mihai
• DOI: http://doi.org/10.1002/for.2662

Received: 5 September 2019 Accepted: 12 January 2020

DOI: 10.1002/for.2662

RESEARCH ARTICLE

Do credit booms predict US recessions?

Marius M. Mihai

Department of Economics, CUNY

Graduate Center, NewYork, New York

Correspondence

Marius M. Mihai, Department of

Economics, CUNY Graduate Center, 365

5th Avenue, New York,NY 10016.

Email: mmihai@gradcenter.cuny.edu

Abstract

This paper investigates the role of bank credit in predicting US recessions since

the 1960s in the context of a bivariate probit model. A set of results emerge. First,

credit booms are shown to have strong positive effects in predicting declines in

the business cycle at horizons ranging from 6 to 9 months. Second, I propose

to isolate the effect of credit booms by identifying the contribution of excess

bank liquidity alongside a housing factor in the downturn of each cycle. Third,

the out-of-sample performance of the model is tested on the most recent

credit-driven recession: the Great Recession of 2008. The model performs better

than a more parsimonious version where we restrict the effect of credit booms

on the business cycle in the system to be zero.

KEYWORDS

business cycles, credit growth, factor models

1INTRODUCTION

An important area of research in empirical macroeco-

nomics is concerned with understanding and predicting

the business cycle. This is of special importance to house-

holds, corporations, and governments to ensure the right

measures are taken in the event of an unexpected reces-

sion. Business cycle turning points are usually associated

with increases in unemployment and large declines in

output and investment. Early detection models of such

severe events are of utmost significance and can help

place the economy in a favorable position to guard to

some extent against some of the negativeeffects mentioned

above. In that regard, there has been a considerable

number of academic papers in the past two decades or so

that took various approaches to predicting US recessions.

The majority of work on this topic involved the use of

univariate probit models to determine macroeconomic

and financial variables that are significant in predicting

recessions. For example, Estrella and Mishkin (1996)

found that interest rate spread (10-year T-bond less

3-month T-bill) along with the 3-month Treasury bill are

useful predictors of turning points in the business cycle.

In another paper, Estrella and Mishkin (1998) extended

their analysis and tested the predictive power of the stock

market and other money supply indicators in addition to

the term spread. They found that stock prices are useful at

a horizon of one to three quarters. Later on, Wright (2006)

showed that after including the term spread controlling

for the federal funds rate gave a better in-sample fit and a

better performance out-of-sample. Kauppi and Saikkonen

(2008) introduced a dynamic component into the model

and found that lagging the probability of a recession vari-

able could yield superior results compared to a more classic

static probit model. In more recent work, Liu and Moench

(2016) revisited the classic leading indicators previously

proposed in other papers and reassessed the performance

of these models through receiving operating characteristic

(ROC) curves.1

The last decade has brought a plethora of macroeco-

nomic and financial data available for researchers to work

on and design more sophisticated modeling approaches.

One good example in this sense is the free data set from

the Federal Reserve Bank of St. Louis (FRED). As a result,

it became possible to shift from estimating simple probits

with just leading macro and financial indicators to more

1The list of other important papers on this topic also includes Chauvet

and Potter (2005), Dueker (1997), Hamilton (2011), Kim and Nelson

(1998), S. Ng (2014), and E. C. Y.Ng (2012).

Journal of Forecasting. 2020;39:887–910. wileyonlinelibrary.com/journal/for © 2020 John Wiley & Sons, Ltd. 887

MIHAI

complex models that include static or dynamic factors built

from these larger data sets. The purpose of these newer

techniques was to induce models to capture more macroe-

conomic information that might otherwise have been

omitted. One of the first papers to introduce this approach

was Chen, Iqbal, and Lai (2011), who estimated US

business cycle turning points by making use of common

factors and showed in real-time simulations the advan-

tages of such methods. A similar paper is Fossati (2015),

which also builds dynamic factors that summarize interest

rate, output and stock market variables from FRED and

shows that factor-augmented probit models have a better

fit compared to the more parsimonious models with

only leading indicators. Taking a comparable econometric

approach, Christiansen, Eriksen, and Møller (2014) found

that sentiment variables offer a higher predictive power for

US recessions in addition to other important macro factors.

Bayesian approaches have been adopted by Fornaro

(2016), who used a methodology with shrinkage in the

parameters to collapse all available macro data and

estimate an augmented probit model that does well com-

pared to a simpler specification.

In light of the most recent global financial crisis, another

branch of empirical macroeconomics has started to ques-

tion the effects of financial deepening and excess credit

on the health of the world economy. This triggered

economists to look specifically at the role of bank credit

in business cycles and financial crises. Mendoza and

Terrones (2012) identified credit booms in a panel of 61

emerging and industrial economies over the 1960–2012

period and found that recessions and financial crises had

a higher probability of occurring as a result of a pro-

longed credit expansion. More evidence of how booms

influence recessions is presented in Jordà, Schularick, and

Taylor (2013), who showed that credit-intensive expan-

sions tend to be followed by deeper recessions and sluggish

recoveries. Within the credit boom literature, researchers

have also analyzed the importance of housing markets

alongside bank credit. Mian and Sufi (2011) showed that

borrowing against the increase in home equity by home-

owners led to a high number of defaults between 2006 and

2008. Anundsen, Gerdrup, Hansen, and Kragh-Sørensen

(2016) proved empirically in a panel of 16 OECD countries

how important house prices and credit are in the like-

lihood of a financial crisis. More recently, Ponka (2017)

estimated a static probit model augmented with macro

factors and various credit variables toshow the importance

of credit spreads in forecasting recessions.

What I propose to do in this paper is to build on these

two strands of literature and estimate a model that tests

the usefulness of credit booms in predicting US business

cycles. While there has been a considerable amount of

evidence to provide motivation as to why it is important to

monitor credit booms in both emerging and industrialized

countries, the role of credit in forecasting US business

cycles has not been analyzed extensively. My contribution

is different since I propose to study credit booms and

recessions jointly in a bivariate-autoregressive probit sys-

tem. These types of models have been initially proposed by

Kauppi and Saikkonen (2008) and extended to a bivariate

case by Nyberg (2014). Modeling business and credit cycles

together allows for a more in-depth analysis of their con-

temporaneous relationship, and also takes into account

any interdependence that could be useful in predicting

crises. My results suggest that credit booms are a signif-

icant predictor of recessions in the USA at the 6- and

9-month horizon; overall, the bivariate model gives a bet-

ter in-sample fit compared to a more parsimonious model

where we do not consider the effect of credit booms or the

housing market. Furthermore, the bivariate model does

a better job at forecasting out-of-sample the most recent

credit-driven recession than its univariate counterpart.

Another interesting exercise was to compute the partial

effects that quantify the contribution of credit booms on

US recessions. Thus I was able to trace the impact of excess

credit on economic downturns over time and identify

the recessions where credit booms played a more influ-

ential role. All models point towards the fact that credit

booms alongside the housing market became more valu-

able indicators over the past two recessions in the USA.

The model suggests that, on average, a credit boom occur-

ring 6 months prior increases the probability of recession

by about three percentage points.

The rest of the paper is organized as follows. Section 2

discusses the identification of recessions and credit booms

in the USA. Section 3 establishes the estimation and

interpretation of macro factors. In Section 4 I introduce

the autoregressive bivariate model and parameter estima-

tion. Section 5 discusses the in-sample and out-of-sample

results. Finally, Section 6 concludes.

2IDENTIFICATION OF US

RECESSIONS AND CREDIT BOOMS

Identification of credit booms has been thoroughly dis-

cussed by Mendoza and Terrones (2012) and Schularick

and Taylor (2012). Other relevant papers in this area are

Dell'Ariccia, Igan, and Laeven (2012), Dell'Ariccia, Igan,

Laeven, and Tong (2014), and Crowe, Dell'Ariccia, Igan,

and Rabanal (2013). Most researchers would agree that

booms are periods of extraordinary credit growth, or devi-

ations from a long-term trend. This definition is closely

dependent on the method applied to detrend the data,

which in most cases was assumed to be a two- or one-sided

HP filter. Since credit booms have been for the most part

studied in the context of panel of industrial or emerging

888