Macroprudential policy stance assessment: the case of Croatia

^* The author states that the views expressed in this paper do not represent the views of the Bank of England. This paper was written at time the author was employed at Croatian National Bank. The author would like to thank two reviewers for their helpful comments and advice.

¹ However, there is still no consensus on this, as found in Arslan and Upper (2017).

² Besides the papers mentioned below, see, e.g. Škrinjarić (2023d, 2024).

³ Also, financial crises are costly. Reinhart and Rogoff (2009) estimate that crisis episodes are related to significant increases in government spending, as government debt increases on average by 86% for three years following a banking crisis. Laeven and Valencia (2012) estimated that the cumulative cost of banking crises is about 23% of GDP during the first four years of their duration, and fiscal costs amount to about 6.8% of GDP (Laeven and Valencia, 2013). Jordá, Schularick and Taylor (2013) found that financial crises are costlier than other recessions, as after five years, the real GDP per capita is lower by 5% compared to other “normal” recessions. Recoveries from financial crises are slower when compared to other types of crises, as found in Kannan, Scott and Terrones (2013). Other relevant findings about the costs of financial crises can be found in Koh et al. (2020), Jordá, Schularick and Taylor (2012), Claessens, Kose and Terrones (2012), and Papell and Prudan (2011). To avoid the costs caused by financial crises, systemic risk should be reduced, in which endeavour macroprudential policy plays an important role.

⁴ Other interesting and important variables have been analysed, such as inflation-at-risk (López-Salido and Loria, 2021), bank capital-at-risk (Lang and Forletta, 2019; 2020), house-price-at-risk (Deghi et al., 2020), unemployment (Adams et al., 2020), or capital flows (Eguren-Martin et al., 2021; Gelos et al., 2022).

⁵ E.g., Plagborg-Møller et al. (2020) found significant country heterogeneity in their results. After a battery of forecasts and estimations made, the authors found a couple of significant mean growth predictors, and fewer still for the volatility of growth, alongside different signs of results, and great cross-country heterogeneity in the results, which prompted the authors to conclude that some caution needs to be taken when one tries to build a theoretical model on empirical results.

⁶ As an example, a credit-to-GNI (Gross National Income) gap and y-o-y GNI growth rates are used for the case of Ireland instead of GDP (Gross Domestic Product), as GNI is a better representation for this case, as well as ICSI (Irish Composite Stress Index) instead of CLIFS (Country-Level Index of Financial Stress) in O’Brien and Wosser (2021).

⁷ Besides the works that are examined below, it is worth mentioning other preceding research that links financial conditions and financial vulnerabilities to the real economy. A comprehensive overview is given in Boyarchenko, Favara and Schularick (2022) and Škrinjarić (2022a; 2023a).

⁸ For the US case, and 1980 for other countries.

⁹ As explained in Škrinjarić (2024), another example of different effects could be that when comparing the countercyclical buffer (CCyB) actions, one country could immediately introduce a value of 2% to take place in a given banking system. This country would get +1 value in the quarter when this action is taken. If the country does not change CCyB in the next few quarters, the value remains at +1. On the other hand, if another country decides to introduce 0.5% CCyB value, and in each subsequent quarter increase it by 0.5 p.p., it will get positive unit value every quarter and could have a cumulative indicator greater than the first country.

¹⁰ As it takes values {-2, -1, 0, 1, 2}.

¹¹ Others include the ESRB database for EU countries: [CrossRef], or Cerutti, Claessens and Laeven (2015). Some other sources are listed and commented in Alam et al. (2019), in appendix I, table 4.

¹² This number comes from counting the total number of all tools in the Excel sheet provided in Budnik and Kleibl (2018).

¹³ Introduction to quantile regression, alongside advantages to other approaches, such as being robust to outliers, heteroskedasticity, non-normality, etc., can be found in Koenker (2005), Davino, Furno and Vistocco (2013), or Koenker and Bassett (1978).

¹⁴ GDP deflator was used to deflate the original nominal GDP series.

¹⁵ See Škrinjarić and Bukovšak (2022) for Croatia’s best individual credit dynamics indicators.

¹⁶ See Škrinjarić (2022a; 2023b) for the composite indicator for Croatia.

¹⁷ SIC values resulted in the same ordering. As these are just ordinary regressions, the idea is to see the tradeoff between the explanatory power of the model versus the number of parameters included in the model. Information criteria give us this information.

¹⁸ I.e., we compare four models M_i, where i stands for how many lags of all variables on the right-hand side (RHS) of the ordered probit equation symbol were included. The explanatory variables included lagged value of the real growth itself, as it is usually put in GaR modelling, and the other variables included were: HIFS and YoY change of the credit-to-GDP ratio. E.g., M_3 means that all variables on the RHS were included with lags 1, 2 and 3 to regress the MPI dynamics on.

¹⁹ Although we are mostly interested in comparing the median and the GaR value, i.e. calculating distance to tail, we are showing all quantiles in graph 9 to get a better picture on the stability of the beta coefficients.

²⁰ We use MPI “shocks” from a variable that was defined with values being equal to +1 or -1 (before purging out effects of other variables), and as we are not using moving sums as some studies do, there is no reason to believe that a measure that takes +1 value in, e.g. 1Q 2019 should have effects 12Q ahead, as this is way overstretching the effects that macroprudential policy has. Using moving sums or similar transformations would include a greater autocorrelation, i.e. memory of the variable and thus it would make sense to look at a longer time horizon. But to observe effects 3 years in future is really overstretching and we don’t hear central banks saying that the macroprudential measures have significant effects on the GDP growth or in general the real economy so long in the future.

²¹ Here, we depict the median value of DTT as just a statistical value. It does not replace the theoretical “optimal” DTT value.

²² Some initial steps have been done by Vandenbussche, Vogel and Detragiache (2015) and followed by Eller et al. (2020).