South Korea Real GDP 12-Month Growth-at-Risk

Growth-at-Risk applied to Korean real GDP, with the headline being the 5th conditional percentile of the 12-month (four-quarter) ahead average annualized growth rate. The object of interest is the whole conditional growth distribution given current conditions and the current secular-growth state, and the question that matters is not where the centre of growth lands but how deep the downside can run when financial conditions tighten and where the centre settles as Korea's trend evolves. With $y_{t\to t+h}$ the average annualized real GDP growth rate from quarter $t$ to quarter $t+h$, this series is the $h=4$ case, and the headline is the following conditional 5th percentile expressed in percent:

where $x_t$ is the conditioning vector holding current growth, financial conditions, and the trend-growth state. The fan running from $q_{05}$ to $q_{95}$, the lower-tail expected shortfall, and the downside−upside asymmetry are provided as auxiliary series for reading which part of the distribution current conditions are acting on. A lower headline means a deeper conditional downside, and this gauge describes the conditional distribution rather than issuing a point forecast.

The conditional quantiles come from per-$\tau$ quantile regression (Koenker and Bassett 1978), and the procedure is as follows.

(1) Conditional quantile regression. For each $\tau$ the conditional quantile is set as an affine function of the regressors.

where the conditioning vector is defined as:

The level control $g_t$ is current annualized quarter-on-quarter growth, $\text{KRFCI}^{\perp}$ is the macro-orthogonalized financial-conditions index, and $g^{f}_t$ is the one-sided trend-growth state from the HLW model (Holston, Laubach, and Williams 2023). $g^{f}_t$ is the quarter-end Kalman-filtered estimate rather than the two-sided RTS-smoothed estimate that the HLW step publishes as its headline, and because the filter at $t$ uses only data through $t$ it carries no look-ahead. Coefficients are estimated at quarterly frequency on the target's native cadence, with Newey-West HAC standard errors at lag $h-1$ to absorb the overlapping $h$-quarter targets.

(2) Conditional quantile evaluation. The estimated intercepts and slopes are evaluated at each quarter-end on the same-period regressor vector

to yield seven conditional quantiles per row, and within-row monotonicity across $\tau$ is enforced by the Chernozhukov, Fernández-Val, and Galichon (2010) rearrangement. The series is produced at quarterly cadence and is not a quarterly model projected onto a daily grid, so no publication-lag forward interpolation is applied.

(3) Expected shortfall and asymmetry. The expected-shortfall and asymmetry layers come from an Azzalini–Capitanio skew-t (Azzalini and Capitanio 2003) fitted to the rearranged conditional quantiles. At quarterly frequency the sample is only about eighty rows, so rather than fitting the shape parameters at quarter-end and forward-filling them, the full four-parameter $(\xi, \omega, \alpha, \nu)$ skew-t is fit directly per quarter by seeded multistart least squares against the closed-form skew-t quantile function. The expected shortfall is defined as $E[y\mid y < q_{05}]$ under the per-quarter skew-t with the threshold pinned to the QR-direct $q_{05}$, which makes $\text{ES}\le q_{05}$ hold by construction. Quarters where the skew-t fit fails to converge or $\nu$ pins at its upper bound are handled by a skew-normal fallback fit, and residual grid-interpolation error is absorbed by a hard clip enforcing $\text{ES}\le q_{05}$. The asymmetry is defined as

so that a positive value means the downside tail is fatter than the upside.

(4) Real-time updating. The coefficients are re-estimated on the full sample at each quarterly run, so recent quarters revise as the HLW filter incorporates later data, and the most recently published quarters are the least settled.

The conditional growth distribution responds to financial conditions asymmetrically. When conditions tighten the lower tail moves down sharply while the upper tail barely moves, so a mean forecast buries that asymmetry in the average. Applying quantile regression to the lower tail (Koenker and Bassett 1978) revives the at-risk signal that the centre of the distribution loses. The theoretical basis of this gauge is the Vulnerable Growth framework of Adrian, Boyarchenko, and Giannone (2019), which stands on the conditional-quantile time-series tradition that Engle and Manganelli (2004) opened with the CAViaR family.

The 12-month horizon is the reporting cadence that Growth-at-Risk surveillance practice has settled on. After the Vulnerable Growth framework took hold in practice, central-bank Financial Stability Reports converged on a four-quarter horizon because one year ahead aligns with the cadence on which countercyclical capital decisions are set (Adams, Adrian, Boyarchenko, and Giannone 2021), and the International Monetary Fund (2017) Global Financial Stability Report uses the same surveillance cadence for its Growth-at-Risk reporting.

The honest summary of how far independent evaluation supports this framework is mixed. Plagborg-Møller, Reichlin, Ricco, and Hasenzagl (2020) report that much of the apparent at-risk signal in U.S. data dissipates once specification and predictor choices are stress-tested, whereas Brownlees and Souza (2021) show that direct quantile regressions using financial-conditions regressors nonetheless beat the unconditional benchmark in genuine out-of-sample evaluation in countries with longer panels. Because the Korean panel is short, this model reads the calibration debate as a boundary on what can be claimed rather than as evidence that settles it.

Conditioning on the HLW filtered trend in addition to the cycle and the financial-conditions index makes the whole conditional distribution ride Korea's secular trend so the level question is not absorbed into a full-sample intercept, which changes how this gauge works and how it is read. At the median $g^{f}_t$ carries a coefficient near $+1$, so the conditional median follows Korea's trend rather than a full-sample average. The 12-month median descends from roughly 3.9 % over 2006–2010 to roughly 2.3 % in the 2020s, matching realized four-quarter-ahead growth and the HLW trend at the era level. The 12-month $\tau=0.05$ pinball ratio is 0.90 against an expanding-window historical-quantile benchmark, and the upper-tail ratio is 0.94. On the coverage-test frame of Christoffersen (1998) and the quantile-evaluation toolkit that Galbraith and van Norden (2019) lay out, PIT hit rates fall within about two percentage points of nominal across all seven $\tau$. The sample is short enough that a ratio slightly under one is a small-margin figure on a single panel rather than a statistically established result.

Two honest caveats remain. Korea's post-2007 quarterly panel has only about 120 observations and only two genuine contractions, the global financial crisis and COVID, so the slope of $\text{KRFCI}^{\perp}$ at $\tau=0.05$ stays imprecisely estimated and the cyclical signal in the deep tail is constrained. The recent $q_{05}$ sitting slightly positive rather than negative is not a defect but the correct calibration in a low-trend environment with calm financial conditions, supported by a full-sample PIT at $\tau=0.05$ of 4.0 % against a nominal 5 % and by only one of the 21 realized post-2020 observations falling below the conditional $q_{05}$. The 2008–09 episode is where the construction is most legible. KRFCI's standardized score broke $+10$ in late 2008 (Baba and Shim 2010). The trend regressor sat near 4.5 %, and the conditional $q_{05}$ went as deep as roughly $-3$ at conditioning dates in early 2007. These dates time-stamp the four-quarter windows running from late 2007 into the 2008 trough, and the realized contraction then landed in the conditional lower tail.

The headline $q_{05}$ reads as a current-conditions Value-at-Risk for real activity over the policy-relevant year-ahead horizon. It is a downside-risk measure sized to KRFCI and anchored to Korea's secular trend through $g^{f}_t$. Central banks reporting Growth-at-Risk on this horizon use it to track how the downside-risk distribution evolves with the cycle (Adams, Adrian, Boyarchenko, and Giannone 2021), and the International Monetary Fund (2017) Global Financial Stability Report writes the cross-country version of the same surveillance object. For the Bank of Korea the natural anchor is the countercyclical capital buffer review, whose decision cadence roughly matches this horizon, so the gauge slots into the existing review without asking for a new surveillance schedule.

The fan and the expected shortfall are where this construction repays its complexity over a single-point growth nowcast. A widening downside spread with a stable median is a regime signal that current conditions are pricing tail risk before the central tendency moves, and the lower-tail expected shortfall is the natural input for scenario-stress design at the 12-month horizon. The asymmetry channel shows the direction in which conditions are tilting the conditional distribution, so its sign and magnitude become primary surveillance variables alongside the level of $q_{05}$.

The trend caveat runs in the opposite direction from a fixed-mean Growth-at-Risk. Before trend conditioning, a slightly positive $q_{05}$ in a calm-conditions environment looked like an understatement of downside risk, but once the HLW trend enters the regression the same reading is the calibrated answer the gauge returns in a low-trend environment rather than a defect. The information in the fan stays in its width and shape, and pairing the gauge with KRFCI keeps the conditions read and the trend read separable. The practical conclusion that reading the two side by side beats either alone is the direction the calibration debate that Plagborg-Møller, Reichlin, Ricco, and Hasenzagl (2020) opened also points toward.

KRGAR12M is an estimated conditional quantile path of the form $\hat{q}_t(\theta)=x_t^{\top}\hat{\beta}(\theta)$ fit by quantile regression of growth on the orthogonalized financial-conditions index, so the path's persistence is inherited from its conditioning variables and is not a property of the growth signal itself. The object that carries forecast quality is therefore the violation-indicator sequence, not the path, and unit-root, structural-break, and level Ljung and Box (1978) batteries on the path were deliberately not run, since a coverage question is a property of the joint forecast-outcome distribution that a univariate path diagnostic cannot address (Christoffersen 1998; Engle and Manganelli 2004; Murphy and Winkler 1987). A bounded estimated quantile also cannot be an unbounded unit-root process, on which standard unit-root inference would over-reject toward spurious stationarity without shrinking in the sample size (Cavaliere 2005; Cavaliere and Xu 2014). The evidence appropriate to a quantile object is coverage backtesting at the two pinned extreme tails, the fifth and ninety-fifth conditional quantiles.

Coverage is read first from the reused validation panel, which is the in-sample probability-integral-transform hit rate over the estimation panel, the empirical frequency that the realized outcome falls at or below the fitted conditional quantile and so targets the quantile level itself, reading 0.062 at the fifth-percentile quantile and 0.929 at the ninety-fifth over 113 quarters, the upper figure being cumulative coverage near its ninety-five-percent target rather than a breach rate (Diebold, Gunther, and Tay 1998; Rosenblatt 1952).

The twelve-month horizon is built on overlapping quarters, so no size-correct unconditional-coverage test exists on its full sample, a literal Kupiec (1995) test and a Newey and West (1987) autocorrelation-robust coverage statistic having both been found materially oversized in a pre-registered Monte-Carlo size check recorded in the model specification, so a formal verdict is withheld and the full-sample record is reported descriptively, with five lower-tail exceedances at a rate of 0.067 and three upper-tail exceedances at 0.04 over 75 quarters.

The size-correct check on the offset-zero non-overlapping subsample is itself uninformative at this horizon, because the eighteen-observation subsample expects only 0.9 exceedances at the five-percent tail, so it records one lower-tail and zero upper-tail exceedance, 0.011 at p = 0.915 and 1.847 at p = 0.174, with too little power to deliver a verdict, and the exceedance rate ranges from zero to 0.111 at each tail across the four admissible offsets. The subsample check is exact under the hypothesis of correct conditional coverage, which renders the spaced violation indicators independent Bernoulli draws (Christoffersen 1998), and it does not claim size control for exceedances of a fixed unconditional quantile of a persistent process, where dependence can cross the non-overlapping window boundaries.

The Christoffersen (1998) independence and conditional-coverage tests are read against that overlap. The independence test rejects at the lower tail, 5.190 at p = 0.023, and is uninformative at the upper tail where no two exceedances are consecutive, while the joint conditional-coverage test fails to reject at both tails, 5.625 at p = 0.060 and 0.316 at p = 0.854. The lower-tail independence rejection is the mechanical clustering that overlapping multi-step targets induce, consistent with the withheld full-sample verdict at this horizon (Christoffersen 1998). Because only a handful of quarterly exceedances occur, too few for the full dynamic-quantile regression, the Engle and Manganelli (2004) test is run in its constant-only binomial-score form, and it fails to reject at both tails, 0.439 with p = 0.508 and 0.158 with p = 0.691, with the full instrument set deferred until the sample supports it. The cross-quantile monotonicity check confirms a non-crossing fan at this horizon (Chernozhukov, Fernandez-Val, and Galichon 2010).

Two limits are stated as such. The overlapping target induces serial dependence in the hit sequence and the non-overlapping subsample expects fewer than one exceedance, so no formal coverage verdict is drawn here and the descriptive rate and the reused validation coverage are the honest content (Christoffersen 1998; Diebold, Gunther, and Tay 1998; Rosenblatt 1952; Berkowitz 2001). The seventy-five quarterly observations also carry limited power against a correctly specified tail, so the descriptive record is weak evidence for calibration rather than strong evidence of it (Kupiec 1995). Because these are estimated quantiles whose parameter uncertainty is not folded into the reference distribution, the coverage reading is descriptive of how current conditions place the growth distribution rather than a validated tail forecast (Campbell 2005; Escanciano and Olmo 2010).