Forecasting recessions: the importance of the financial cycle

Abstract

Financial cycles can be important drivers of real activity, but there is scant evidence about how well they signal recession risks. We address this question, using a range of financial cycle measures. Unlike most papers, ours assesses forecasting performance not just for the United States but also for a panel of advanced and emerging market economies. We find that financial cycle measures have significant forecasting power both in and out of sample, even for a three-year horizon. Moreover, they outperform the term spread - the most widely used indicator in the literature - in nearly all specifications. These results are robust to different recession specifications.

Introduction

Predicting recessions has been a long-standing quest for market practitioners, policymakers and academics alike. As early as the first volume of The American Economic Review, Irving Fisher (1911) looked at developments in the “nation's bank book” to forecast the likelihood of a contraction in the United States.² Following Estrella and Mishkin (1998), a consensus has emerged that the inverted yield curve, i.e. long-term bond yields below short-term interest rates, is among the best signals of impending recessions, if not the best one.

The notion that finance matters for the real economy has regained significant attention, especially after the Great Financial Crisis.³ Some authors even argue that all recessions in the United States since 1985 had financial origins (Ng and Wright (2013)).⁴ Despite this, research exploring how financial developments affect recession risks, i.e. the likelihood that a recession will occur in the near future, and whether they outperform the term spread as recession signal is scant. We fill this gap.

We explore in particular the importance of the financial cycle for predicting recession risks. This is not only of interest academically, but also from a policy perspective as the concept of the financial cycle is underpinning financial stability analyses in many policy institutions.⁵

In the spirit of Minsky (1982) and Kindleberger (2000), the term “financial cycle” refers to the self-reinforcing interactions between asset prices, risk, risk taking, and financing constraints (Borio (2014)).⁶ To measure the financial cycle empirically, the literature has emphasised the role of joint, medium-term fluctuations - typically in the range of 10-20 years – in credit and asset prices as important interactions between credit, asset prices and the real economy play out at these frequencies along two dimensions. First, medium-term cycles of credit and asset prices are especially highly correlated with the medium-term GDP cycle (e.g. Runstler and Vlekke (2018) or de Winter et al. (2017)). Second, peaks in the (medium-term) financial cycle coincide with considerable financial stress or banking crises (e.g. Drehmann et al. (2012)), which in turn usher in deep and protracted recessions (e.g. Jordà et al. (2013)).

In this paper, we look at two financial cycle measures specifically. First, the “composite” financial cycle proxy as in Drehmann et al. (2012). This combines medium-term cycles in real credit, the credit-to-GDP ratio and real residential property prices. Second, the debt service ratio (DSR), defined as interest payments plus amortisation divided by GDP. Juselius and Drehmann (20202) show that the financial cycle can be described by the joint behaviour of leverage and the DSR. The DSR, in turn, is the key link to the real economy as it has a large negative impact on GDP growth (see also Drehmann et al. (2018)).

By focussing on the financial cycle, we expand on the few papers in the literature that assess the predictive power of individual variables that proxy some aspects of the financial cycle. For instance, the literature has found balance sheet conditions, property prices, credit growth or corporate credit spreads to be important (e.g. Ng (2012), Liu and Mönch (2016), Ponka (2017), Guender (2018) or Hwang (2019)).

As our aim is to better understand how financial cycles affect recession risks, we do not to try maximise overall forecast performance. We run simple panel probit models and do not do a horse race across all possible explanatory variables.⁷ But it is clear from the literature that using more elaborate methods such as time varying coefficient models, Bayesian model averaging or machine learning would boost forecast performance even further (e.g. Berge (2015) or Hwang (2019)).

We assess predictive performance of the financial cycle both in and out of sample. We rely on the area under the receiver operating characteristic (ROC) curve (AUC) to judge forecast performance (Berge and Jordà (2011)). Throughout, we compare the performance of the financial cycle with that of the term spread in light of the importance of the term spread as a benchmark even though its performance has shown to change over time (e.g. Chauvet and Potter (2002) or Hwang (2019)).

And in contrast to much of the literature, we assess the signalling power of financial cycle measures and the yield curve not only for the United States but also for an unbalanced panel of 16 advanced economies and nine emerging market economies (EMEs) from 1985 to 2017. This makes our results significantly more general. For instance, of the above-mentioned papers that look at variables related to the financial cycle, only Guender (2018) goes beyond the United States and considers a panel of European countries. And even the literature on the information content of the yield curve barely looks at EMEs; the exceptions are Mehl (2009), who examines the effect of the yield curve on growth; Öztürk and Pereira (2013), whose panel analysis includes EMEs; and several country-specific studies (e.g. Grabowski (2009) for Poland, Zulkhibri and Rani (2016) for Malaysia, and Paweenawat (2017) for Thailand)).

We find that financial cycle measures are very useful for gauging recession risks. They perform generally better than the term spread. When the competing variables are considered on a standalone basis, AUCs for financial cycle measures are statistically significantly higher and are significant even for a three-year horizon, for which the term spread is uninformative. When financial cycle measures and the term spread are included jointly in a probit model, they all remain statistically significant up to a two-year horizon. But combining information from the spread and financial cycle measures improves only marginally, and not significantly, AUCs and other evaluation metrics relative to specifications that include just financial cycle measures.

The high forecasting performance of the financial cycle measures also applies to the out-of-sample tests. Here, we carry out two exercises to assess the indicators’ predictive content in (quasi) real time. We examine first the effect of real-time data⁸ and fixed model parameters estimated using our first 10 years of data. We then assess the combined effect of real-time data and model parameters estimated recursively. In both exercises, the financial cycle measures retain statistically significant forecasting power.

The results apply to both advanced economies and EMEs. That said, the forecast performance of both the term spread and the financial cycle is generally weaker in EMEs, possibly reflecting much more volatile business cycles there. Across countries, a reason for the disappointing performance of the term spread of is that for several economies the variable is “contaminated” by credit risk premia. As a result, in some episodes, the yield curve steepened rather than flattened ahead of recessions, including in some periphery countries ahead of the 2011–12 euro sovereign debt crisis. By contrast, the financial cycle measures are immune to this problem.⁹ Consistently with this explanation, the term spread retains forecasting information for the United States as shown by the literature.

The performance of financial cycle proxies is robust regardless of whether we forecast the likelihood that the economy will be in a recession at a specific point in time several periods ahead – the literature standard – or whether the business cycle will turn within the next few periods. Following much of the literature (e.g. Rudebusch and Williams (2009)), in the main text we focus on the first approach – in our case one, two or three years ahead. As an alternative, we estimate the risk that the business cycle will turn from boom to bust within the next one, two or three years. This second type of exercise is much closer to the spirit of Irving Fisher, who had observed in 1911: “The exact date of such crisis (recession in his context), of course, it would be foolish to predict, but it if it occurs it would seem likely to occur between, say 1913 and 1916” (Fisher (1911), p 304).

The rest of the paper is structured as follows. In the second section, we briefly introduce the notion of the financial cycle and document how the nature of the business cycle, and its link with the financial cycle, have changed in the past 50 years. In the third section, we explain our methodology. In the fourth, we evaluate the performance of financial cycle proxies and compare it with that of the term spread based on full-sample information, i.e. ex post. In the fifth, we consider out-of-sample exercises, seeking to mimic the information policymakers have when assessing risks in real time, i.e. ex ante. In the sixth, we consider a different definition of recession risk as robustness check. Then we conclude.