The Global Financial Cycle and Growth Risk
Global, not local, financial conditions determine growth risk
We know that the tidal whipsaw of the global financial cycle lifts all boats when it comes in. And on its way out, it leaves strugglers on all shores. Nations, even and especially those at the technological frontier, are not meaningfully insulated from the global financial cycle. To the contrary, advanced economies are more exposed to fluctuations in the risk appetite of global banks than emerging markets. Financial institutions headquartered in New York sit at the center of global financial intermediation. We know from work at the Bank of International Settlements that global banks transmit financial conditions from the center to the periphery. Financial conditions ease worldwide when they do so at the heart of the system. It has been shown that the strength of the dollar closely tracks global financial conditions. When the dollar strengthens, global financial conditions get tighter. When the dollar weakens, global financial conditions ease.
Given all these results, I was surprised to find the strategy pursued by Adrian — arguably the father of macrofinance — and coworkers in “The Term Structure of Growth-at-Risk,” an IMF working paper published in 2018. They use the same technology as the one we replicated two days ago. The idea, as before, is that tighter financial conditions increase the probability that future growth in real GDP will disappoint even if does not affect the baseline scenario. We have shown that this result is robust to using forward-looking estimates of growth instead of using contemporaneous rate of growth as the baseline. The strategy in “The Term Structure of Growth-at-Risk” is to do this for a bunch of advanced economies and emerging markets.
Tobias Adrian was kind enough to share their AE panel dataset with me. They have constructed country-specific measures of financial conditions (FCI) using risk spreads et cetera. They carry out quantile regressions of RGDP growth rates on FCI, credit growth, and dummy for financial booms. They define Growth-at-Risk (GaR) as the lowest 5th quantile of future GDP growth distribution. They bring heavy artillery to the task, but the basic idea is that the probability of poor growth performance in the future is modeled as a function of contemporaneous national financial conditions.
I don’t think we should also condition on credit aggregates and suchlike if we are interested in isolating the effect of financial conditions on growth risk. The reason is that they are causally downstream from FCI. That is, credit aggregates mediate between FCI and growth risk. Controlling for them removes the causal channel that works through them, thereby underestimating the effect of FCI on growth risk. As we have learned from Judea Pearl and others working in causation, one is not free to choose covariates. Identification is only possible if and when we condition on the right covariates — a function of the causal diagram. So, in what follows, we’ll stick to the predictive relationship between FCI and growth risk alone.
For reasons that will become clear shortly, instead of running quantile regressions, we construct an indicator for poor performance (lower than the 5th country-specific percentile) and run logistic regressions instead. That is, we directly model the probability of poor performance as a function of financial conditions. We first construct the global factor as the cross-country mean of country-specific FCIs by quarter. This will be our main explanatory feature. The response will always be the logit probability of poor performance defined as growth at rates lower than the 5th country-specific percentile. We then estimate the following five models.
(1) Global Model: The response is a linear function of the global factor. The coefficients are assumed to be the same for all advanced economies.
(2) Pooled Model: The response is a linear function of country-specific FCI. The coefficients are assumed to be the same for all advanced economies.
(3) Glocal Model: The response is a linear function of the global factor and country-specific FCI. The coefficients are assumed to be the same for all advanced economies.
(4) Unpooled Model: The response is a linear function of country-specific FCI. The coefficients are allowed to be different for each advanced economy.
(5) Hierarchical Model: We stratify by quarter. The response is a linear function of country-specific FCI, but we admit random coefficients. The random coefficients are allowed to vary by quarter as a linear function of the global factor. That is, we have the hierarchical model as below, where we stratify by quarter indexed by j, the response Y(i,j) is the logit probability of country i seeing poor performance in quarter j, X(i, j) is the lagged country-specific value of FCI, and W(j) is the lagged value of the global factor.
This hierarchical model can be estimated from the combined equation as a mixed effects model:
The beauty of Gelman’s hierarchical approach is that the parameters of the second level model can be read off the fixed effects of the mixed effects estimator. The model (5) is essentially a compromise between the pooled model (2) and the unpooled model (4). Instead of treating the countries as the same or entirely different in the sense the coefficients are allowed to vary across countries in an unrestrained manner, it allows us to “borrow” information between countries. The Glocal Model (3) is just there to test the hypothesis that country-specific FCI contains any additional predictive information beyond the global factor. The other models are nested within the hierarchical model and live inside it as special cases. We shall see that this general formulation allows us to establish that only the global factor matters. The country-specific financial conditions contain no additional predictive information about growth risk beyond that already contained in the global factor. There is a single global financial cycle.
The global model (1) shows a very significant positive predictive relationship between the probability of poor performance in the next quarter and contemporaneous global financial conditions, as captured by mean FCI in this quarter. The parameter estimate is large (b = 1.25) and highly significant (t = 9.6).
The pooled model (2) shows that contemporary country-specific FCI positively predict the probability of poor performance in the next quarter. But the slope coefficient is smaller (b = 0.84) even if still highly significant (t = 8.3).
The glocal model (3) tests whether country-specific FCI contain any predictive information beyond that contained in the global factor. The answer is an unambiguous no. Country-specific FCI falls into insignificance (t = 0.1) once we include the global factor in the model. But the global factor remains highly significant (t = 5.6) and just as large (b = 1.24).
In the unpooled model (4), all country fixed-effects vanish (P > 0.21). With the exception of Italy, country-specific slopes for FCI are significant at the 5 percent level. We can see that higher values of FCI, ie tighter local financial conditions, predict a higher probability of poor performance.
Finally, we have the hierarchical model (5). In order to interpret the estimated coefficients of this model, we must refer back to the hierarchical formulation. The intercept is gamma(0,0), the intercept of the random intercept, which is significant (t = -18.9). The slope of the global factor (mfci) is gamma(0,1), the slope of the random intercept, which is very large (b = 1.65) and highly significant (t = 5.2). The slope for country-specific FCI is gamma(1,0), the intercept of the random slope, which vanishes (t = 0.0). The interaction term is gamma(1,1), the slope of the random slope, which also vanishes (t = -1.0). This shows that local financial conditions contain no more information than that already contained in the global factor. What the parameter estimates of the very general hierarchical model show is that the global model is right — growth risk is a function of the global financial cycle. Note that the slope coefficient is larger than that identified by the purely fixed-effects global model (b = 1.65 vs b = 1.25). The interpretation is that the straight-up fixed-effects panel regression underestimates the effect of tighter global financial conditions on growth risk.
We have shown that global financial conditions accounts for all the variation in growth risk for the advanced economies. The interpretation is that global financial intermediation is a single, unitary structure and not balkanized along national borders. All advanced economies face growth risk when global financial conditions tighten. There is simply no evidence to suggest that local financial conditions contain any predictive information beyond that contained in the global factor. The next figure displays the global financial cycle as proxied by the mean FCI of advanced economies.
Many thanks to Tobias for sharing their data and ideas.
Postscript. Here’s the probability of an awful next quarter conditional on global financial conditions (Global Model).