Growth-at-Risk (GaR) Overview

• Growth-at-Risk (GaR) is defined as the lower quantile of future GDP growth, typically at the 5th percentile, capturing potential economic downturns.
• It employs diverse econometric models including conditional quantile regression, time-varying parameter systems, and high-dimensional Bayesian techniques to forecast tail risk.
• Recent frameworks enhance GaR by addressing calibration issues and tail extrapolation using methods like conformal calibration and tail-index estimation.

Searching arXiv for recent GaR-related papers to ground the article in cited research. Growth-at-Risk (GaR) is the lower tail of the future output- or GDP-growth distribution, typically operationalized as a low quantile—often the $5^{\text{th}}$ percentile—at a specified forecast horizon. In predictive form, it is the $\tau$-quantile of future growth conditional on the current information set, $$\mathrm{GaR}_{\tau,t+h}=F^{-1}_{y_{t+h}\mid\mathcal F_t}(\tau)$$; in a structural-quantile formulation, it can be written as $\text{GaR}_h=q_h(\tau\mid d)\big|_{\tau=0.05,d=0}$, where the target remains an unconditional quantile even though controls are used for identification (Gächter et al., 2023, Wojciechowski, 2024). Recent GaR research has expanded the object from a single left-tail forecast into a broader tail-risk framework comprising structural impulse responses, time-varying predictive distributions, high-dimensional shrinkage and screening methods, mixed-frequency and textual predictors, multi-country nonlinear factor systems, and conformal calibration procedures (Kohns et al., 2020, Bogani et al., 2024, Clark et al., 2021).

1. Core definitions and statistical objects

In the canonical predictive formulation, GaR at level $\tau$ is the conditional $\tau$-quantile of future GDP growth. One common expression is

$$\mathrm{GaR}_{\tau,t+h} \;=\;\inf\{\,y:F_{y_{t+h}\mid\mathcal F_t}(y)\ge\tau\} \;=\;F^{-1}_{y_{t+h}\mid\mathcal F_t}(\tau)\,.$$

Equivalent formulations appear in applied quantile-regression settings as $$\mathrm{GaR}(\tau)\equiv Q_{GDP_{t+1}}(\tau\mid X_t)$$, with small $\tau$, such as $0.05$ or $\tau$0, interpreted as downside risk (Gächter et al., 2023, Kohns et al., 2020).

A distinct structural definition emphasizes potential outcomes and unconditional quantiles. In that setup,

$\tau$1

where $\tau$2 is the $\tau$3-quantile of the $\tau$4-period-ahead growth distribution under treatment level $\tau$5. The paper defining this object states that GaR is the pessimistic $\tau$6 scenario for growth, “unconditional” on controls but identified via controls (Wojciechowski, 2024). This distinction matters because it separates the substantive estimand—an unconditional tail outcome—from the role of covariates in identification.

Although GaR is conventionally associated with the left tail, several papers treat both tails jointly. “A tale of two tails” studies the lower $\tau$7 and upper $\tau$8 percentiles over 130 years and shows that the behavior of upside risk is historically contingent rather than fixed (Gächter et al., 2023). “Robust Econometrics for Growth-at-Risk” makes the symmetry explicit by defining extreme lower and upper conditional quantiles, as well as the lower-tail expectation

$\tau$9

and the upper-tail expectation

$$\mathrm{GaR}_{\tau,t+h}=F^{-1}_{y_{t+h}\mid\mathcal F_t}(\tau)$$0

These objects extend GaR from a quantile-based warning indicator to a tail-distribution framework that also tracks the severity of extreme downside and upside realizations (Adrian et al., 1 Aug 2025).

2. Econometric frameworks for estimating GaR

The baseline GaR model is conditional quantile regression. In its simplest form,

$$\mathrm{GaR}_{\tau,t+h}=F^{-1}_{y_{t+h}\mid\mathcal F_t}(\tau)$$1

and estimation at $$\mathrm{GaR}_{\tau,t+h}=F^{-1}_{y_{t+h}\mid\mathcal F_t}(\tau)$$2 yields GaR(5%). This formulation underlies a large share of the applied literature, including mixed-frequency and large-predictor-set extensions (Isler, 7 Nov 2025).

A major strand replaces static coefficients with time-varying parameter regressions with stochastic volatility. In the 130-year study, the predictive distribution of future GDP growth is generated by a TVP-SV state-space system in which financial stress, current growth, three-year average credit-to-GDP growth, and three-year average real house-price growth enter with drifting coefficients, while the innovation variance follows an AR(1) law in logs (Gächter et al., 2023). The objective is not merely to estimate a point quantile, but the entire evolving predictive distribution.

High-dimensional Bayesian quantile regression addresses the case in which the predictor dimension is large relative to sample size. The horseshoe-prior Bayesian quantile regression model uses an Asymmetric Laplace Likelihood, a global-local shrinkage hierarchy,

$$\mathrm{GaR}_{\tau,t+h}=F^{-1}_{y_{t+h}\mid\mathcal F_t}(\tau)$$3

with half-Cauchy priors on local and global scales, and a fast sampling algorithm based on data augmentation and block-slice sampling. In the U.S. GaR application, forecasts are produced from a grid of quantiles and then converted into a continuous predictive density via monotonicity enforcement and kernel smoothing (Kohns et al., 2020).

Nonparametric and nonlinear panel methods enlarge the information set further. The multi-country quantile factor BART model writes the $$\mathrm{GaR}_{\tau,t+h}=F^{-1}_{y_{t+h}\mid\mathcal F_t}(\tau)$$4-th conditional quantile as a convex combination of a linear component and a BART-based nonlinear function, augmented by a common latent factor with stochastic volatility and asymmetric Laplace errors. Cross-sectional information is pooled across 11 advanced economies, and the model allows the data to determine how nonlinear each country-quantile relationship is through the weight $$\mathrm{GaR}_{\tau,t+h}=F^{-1}_{y_{t+h}\mid\mathcal F_t}(\tau)$$5 (Clark et al., 2021).

Mixed-frequency methods incorporate high-frequency indicators into quarterly GaR forecasts. In the textual-sentiment application, weekly tone extracted from SEC filings enters a MIDAS-quantile regression through an unnormalized Almon lag polynomial, so that low-frequency GDP growth is forecast using a weighted distributed lag of weekly sentiment observations (Isler, 7 Nov 2025).

Framework	Target	Distinctive feature
TVP-SV	Predictive distribution of future GDP growth	Time-varying coefficients and stochastic volatility
Structural GQR-local-projections	Unconditional quantile response $$\mathrm{GaR}_{\tau,t+h}=F^{-1}_{y_{t+h}\mid\mathcal F_t}(\tau)$$6	Controls used for identification, not for conditioning
HS-BQR	Conditional quantiles and full density	Horseshoe prior and fast MCMC in high dimensions
MIDAS-QR	One-quarter-ahead conditional quantile	Weekly predictors enter via Almon lag weights
QF-BART	Multi-country conditional quantiles	BART nonlinearity plus latent factor with stochastic volatility
Conformalized quantiles	Calibrated lower-tail predictor	Finite-sample coverage under exchangeability

3. Structural interpretation and quantile impulse responses

A central methodological issue in GaR is whether the estimated object is a conditional tail forecast or a structural tail response to an exogenous shock. “A Structural Approach to Growth-at-Risk” formulates a structural quantile function through the potential-outcome representation

$$\mathrm{GaR}_{\tau,t+h}=F^{-1}_{y_{t+h}\mid\mathcal F_t}(\tau)$$7

and defines the quantile impulse response at quantile $$\mathrm{GaR}_{\tau,t+h}=F^{-1}_{y_{t+h}\mid\mathcal F_t}(\tau)$$8 as

$$\mathrm{GaR}_{\tau,t+h}=F^{-1}_{y_{t+h}\mid\mathcal F_t}(\tau)$$9

Under the linear-in-$\text{GaR}_h=q_h(\tau\mid d)\big|_{\tau=0.05,d=0}$0 approximation $\text{GaR}_h=q_h(\tau\mid d)\big|_{\tau=0.05,d=0}$1, the QIR is $\text{GaR}_h=q_h(\tau\mid d)\big|_{\tau=0.05,d=0}$2 (Wojciechowski, 2024).

The same paper sharply distinguishes this object from the conditional quantile impulse response

$\text{GaR}_h=q_h(\tau\mid d)\big|_{\tau=0.05,d=0}$3

where the quantile is conditional on both the treatment $\text{GaR}_h=q_h(\tau\mid d)\big|_{\tau=0.05,d=0}$4 and controls $\text{GaR}_h=q_h(\tau\mid d)\big|_{\tau=0.05,d=0}$5. Its identification strategy separates treatment from controls: $\text{GaR}_h=q_h(\tau\mid d)\big|_{\tau=0.05,d=0}$6 is the variable whose causal effect on an unconditional quantile is being traced, while $\text{GaR}_h=q_h(\tau\mid d)\big|_{\tau=0.05,d=0}$7 is included to justify a “treatment-as-if-random” assumption. The key assumptions are potential-outcome existence and monotonic structural quantile function, conditional independence $\text{GaR}_h=q_h(\tau\mid d)\big|_{\tau=0.05,d=0}$8, and rank similarity conditional on $\text{GaR}_h=q_h(\tau\mid d)\big|_{\tau=0.05,d=0}$9 (Wojciechowski, 2024).

Identification proceeds through Generalized Quantile Regression, which imposes two moment restrictions at each horizon $\tau$0 and quantile $\tau$1:

$\tau$2

and

$\tau$3

Under linearity, these yield a sample-moment system in $\tau$4, implemented by grid-searching over $\tau$5, choosing $\tau$6 to satisfy the unconditional quantile restriction, estimating a first-stage Logit/Probit for the threshold event on controls, and selecting $\tau$7 to minimize a quadratic form in the treatment-weighted moment. Confidence intervals are computed with a “block-of-blocks” bootstrap using blocks of length $\tau$8 and $\tau$9 replications (Wojciechowski, 2024).

The empirical design uses monthly U.S. data from 1985:01 to 2023:08, with $\tau$0. The dependent variable is cumulative log-growth in industrial production,

$\tau$1

chosen because GDP is quarterly. Treatments are z-scored changes in the Excess Bond Premium and z-scored $\tau$2. Controls comprise contemporaneous consumption growth, investment growth, IP growth, inflation, S&P 500 return, $\tau$3-year yield, $\tau$4Fed-funds, and two lags of all eight series (Wojciechowski, 2024).

The conceptual implication is narrow but important: adding controls in ordinary quantile-local-projection settings changes the estimand from an unconditional quantile $\tau$5 to the conditional quantile $\tau$6, whereas in the structural GQR framework controls affect identification without changing the interpretation of the quantile response. This is the main methodological fault line between structural and reduced-form GaR estimators (Wojciechowski, 2024).

4. Empirical regularities in downside and upside risk

The historical evidence indicates that both the level and the drivers of GaR are time-varying. Over 130 years of U.S. data, the lower and upper tails of the GDP-growth distribution were both volatile before the 1970s, while the apparent stability of upside risk documented in some shorter-sample studies is specific to the Great Moderation. The same study reports that the distribution of GDP growth narrowed significantly since the end of the Bretton Woods system, with especially pronounced declines in the cross-time variability of predicted tail bounds (Gächter et al., 2023).

Financial stress emerges as a consistently asymmetric driver. In the 130-year TVP-SV evidence, financial stress always has a negative and statistically significant effect on the $\tau$7 quantile but virtually no impact on the $\tau$8 quantile. Credit growth and house-price growth behave differently: both can broaden the predictive distribution, raising upside and downside risk during some boom episodes, while also contributing to lower GaR during downturns such as the Great Depression, the Great Financial Crisis, or sharp housing reversals (Gächter et al., 2023). This directly contradicts the view that GaR drivers are temporally invariant or purely one-sided.

The structural-quantile evidence implies a stronger left-tail response to financial shocks than earlier conditional-quantile estimates suggested. After a $\tau$9 standard deviation credit-risk shock, downside risk at $$\mathrm{GaR}_{\tau,t+h} \;=\;\inf\{\,y:F_{y_{t+h}\mid\mathcal F_t}(y)\ge\tau\} \;=\;F^{-1}_{y_{t+h}\mid\mathcal F_t}(\tau)\,.$$0 reaches a peak-to-trough cumulated loss of about $$\mathrm{GaR}_{\tau,t+h} \;=\;\inf\{\,y:F_{y_{t+h}\mid\mathcal F_t}(y)\ge\tau\} \;=\;F^{-1}_{y_{t+h}\mid\mathcal F_t}(\tau)\,.$$1 percentage points at 12 months, whereas the median loss is only $$\mathrm{GaR}_{\tau,t+h} \;=\;\inf\{\,y:F_{y_{t+h}\mid\mathcal F_t}(y)\ge\tau\} \;=\;F^{-1}_{y_{t+h}\mid\mathcal F_t}(\tau)\,.$$2 percentage points and the $$\mathrm{GaR}_{\tau,t+h} \;=\;\inf\{\,y:F_{y_{t+h}\mid\mathcal F_t}(y)\ge\tau\} \;=\;F^{-1}_{y_{t+h}\mid\mathcal F_t}(\tau)\,.$$3 response is close to zero or only mildly negative. A richer quantile plot shows that the $$\mathrm{GaR}_{\tau,t+h} \;=\;\inf\{\,y:F_{y_{t+h}\mid\mathcal F_t}(y)\ge\tau\} \;=\;F^{-1}_{y_{t+h}\mid\mathcal F_t}(\tau)\,.$$4–0.2 region is most sensitive, with $$\mathrm{GaR}_{\tau,t+h} \;=\;\inf\{\,y:F_{y_{t+h}\mid\mathcal F_t}(y)\ge\tau\} \;=\;F^{-1}_{y_{t+h}\mid\mathcal F_t}(\tau)\,.$$5 percentage points at one year versus $$\mathrm{GaR}_{\tau,t+h} \;=\;\inf\{\,y:F_{y_{t+h}\mid\mathcal F_t}(y)\ge\tau\} \;=\;F^{-1}_{y_{t+h}\mid\mathcal F_t}(\tau)\,.$$6 percentage points. Volatility-risk shocks generate almost identical quantile impulse responses at all $$\mathrm{GaR}_{\tau,t+h} \;=\;\inf\{\,y:F_{y_{t+h}\mid\mathcal F_t}(y)\ge\tau\} \;=\;F^{-1}_{y_{t+h}\mid\mathcal F_t}(\tau)\,.$$7, despite a low correlation of about $$\mathrm{GaR}_{\tau,t+h} \;=\;\inf\{\,y:F_{y_{t+h}\mid\mathcal F_t}(y)\ge\tau\} \;=\;F^{-1}_{y_{t+h}\mid\mathcal F_t}(\tau)\,.$$8 between the credit and volatility series (Wojciechowski, 2024).

Multi-country evidence points to systematic tail comovement. In the 11-country QF-BART application, the specification with estimated nonlinear weight $$\mathrm{GaR}_{\tau,t+h} \;=\;\inf\{\,y:F_{y_{t+h}\mid\mathcal F_t}(y)\ge\tau\} \;=\;F^{-1}_{y_{t+h}\mid\mathcal F_t}(\tau)\,.$$9 and pooling prior uniformly outperforms the ABG benchmark, with about $$\mathrm{GaR}(\tau)\equiv Q_{GDP_{t+1}}(\tau\mid X_t)$$0 lower overall CRPS, about $$\mathrm{GaR}(\tau)\equiv Q_{GDP_{t+1}}(\tau\mid X_t)$$1 in the left tail, and about $$\mathrm{GaR}(\tau)\equiv Q_{GDP_{t+1}}(\tau\mid X_t)$$2 in the right tail. Estimated $$\mathrm{GaR}(\tau)\equiv Q_{GDP_{t+1}}(\tau\mid X_t)$$3 is close to zero in the center of the distribution and close to one in the $$\mathrm{GaR}(\tau)\equiv Q_{GDP_{t+1}}(\tau\mid X_t)$$4 and $$\mathrm{GaR}(\tau)\equiv Q_{GDP_{t+1}}(\tau\mid X_t)$$5 tails, implying that nonlinearities matter primarily in the extremes. The common latent factor explains over $$\mathrm{GaR}(\tau)\equiv Q_{GDP_{t+1}}(\tau\mid X_t)$$6 of variance in the tails but only about $$\mathrm{GaR}(\tau)\equiv Q_{GDP_{t+1}}(\tau\mid X_t)$$7–$$\mathrm{GaR}(\tau)\equiv Q_{GDP_{t+1}}(\tau\mid X_t)$$8 in the center, and its realizations spike during major downturns such as 1990 Q1, 2008–09, and 2020 Q2 (Clark et al., 2021).

Taken together, these findings support a stable empirical hierarchy. Financial stress is repeatedly identified as a downside-risk variable; credit and housing are more two-sided and regime-dependent; and extreme quantiles react more strongly than medians or means. A plausible implication is that GaR should be interpreted as a state-dependent tail object rather than a constant transformation of average-growth forecasts.

5. High-dimensional predictors, machine learning, and alternative data

As predictor sets expand from a few financial conditions variables to hundreds of macro-financial series, variable selection becomes central to GaR estimation. One approach is Bayesian shrinkage. The horseshoe-prior Bayesian quantile regression application uses 229 macro-financial series from the McCracken–Ng FRED-QD database and evaluates forecasts over horizons $$\mathrm{GaR}(\tau)\equiv Q_{GDP_{t+1}}(\tau\mid X_t)$$9. Across quantile-specific and density-calibration criteria, the model delivers the best performance: only HS-BQR’s PIT CDFs pass the Kolmogorov–Smirnov test at $\tau$0 for all horizons, it attains the highest average log-scores for $\tau$1 and ties for $\tau$2, and it produces the largest pseudo-$\tau$3 gains especially for extreme $\tau$4 (Kohns et al., 2020).

A second approach is screening via quantile partial correlation. The variable-screening paper defines GaR as $\tau$5, estimates sample QPCs conditional on an active set, and proposes QPC-Forward and QPC-Confounder-Adjusted algorithms, followed by Extended BIC model selection. Under $\tau$6-mixing and a $\tau$7-min condition, it establishes uniform convergence and screening consistency for time-series data. In the empirical GaR application with $\tau$8 Fred-QD predictors over 1987Q3–2022Q3, financial-condition and credit aggregates appear frequently, but the single most-selected predictor is CLAIMSx; during 2020Q1–2022Q3, labor-market variables such as CLAIMSx and AWHMAN are selected in nearly all windows, while NFCI is almost never chosen once one controls for a larger set of specific indicators (Chen et al., 2024).

“Machine-learning Growth at Risk” extends the screening logic into a full decomposition framework. Its QPCR algorithm sequentially selects predictors using sample quantile partial correlations, chooses model size by an EBIC-type criterion, and forms GaR forecasts as $\tau$9. In pseudo-out-of-sample forecasting, QPCR achieves the lowest mean quantile-prediction error, $0.05$0, against penalized quantile regressions, quantile random forests, and a GARCH(1,1) benchmark. The repeatedly selected predictors fall into capacity-utilization, labour-market slack, housing, and financial-spread/volatility groups; the method also decomposes overall GaR into sectoral contributions and standardizes targeted sector-specific downside-risk indices. The paper reports that the targeted financial index correlates $0.05$1 with NFCI but only $0.05$2 with the labour index, indicating partial isolation of financial downside risk (Adrian et al., 31 May 2025).

Alternative data have also entered the GaR framework. In the corporate-text application, tone growth is built from positive and negative word frequencies in 10-K and 10-Q filings using the Loughran–McDonald dictionary, converted to year-over-year firm-level tone growth, and aggregated into a weekly market-cap-weighted tone index. The weekly series enters a one-quarter-ahead MIDAS-quantile regression for GDP growth. Out of sample, the tone-based model attains a pinball loss of $0.05$3 against $0.05$4 for the weekly NFCI model, corresponding to a Quantile Skill Score of $0.05$5; the Diebold–Mariano statistic is $0.05$6 with $0.05$7. The improvement remains positive in recession quarters, and robustness checks show that the gain increases as the text input is refined from all filings to a curated narrative sample, with QSS rising from $0.05$8 to $0.05$9 (Isler, 7 Nov 2025).

The broad pattern across these methods is consistent: once the predictor space is expanded, labor-market indicators, housing variables, credit spreads, and high-frequency sentiment all contribute information not captured by coarse aggregate financial-conditions indices alone.

6. Calibration, robustness, and methodological controversies

A persistent problem in GaR estimation is that tail quantile forecasts can be materially miscalibrated, especially at extremal levels. The conformalized-quantile approach addresses this by splitting the sample into training and calibration sets, computing nonconformity scores $\tau$00 on the calibration set, and adjusting the original lower-tail quantile estimate according to

$\tau$01

Under exchangeability, the resulting one-sided prediction set has finite-sample coverage control, and under a tie-breaking rule the miscoverage probability lies between $\tau$02 and $\tau$03 (Bogani et al., 2024).

In simulations with heavy tails, nonstationarity, and high dimension, conformalization substantially improves calibration for both quantile regression and quantile random forests. In the GaR application using quarterly U.S. GDP growth and NFCI, un-conformalized QR overshoots the diagonal in the left tail at $\tau$04, with actual coverage around $\tau$05 when $\tau$06; conformalized QR restores near-nominal lower-tail coverage. Reported MAE values also fall materially: for QR based on aggregate NFCI, MAE declines from $\tau$07 to $\tau$08 at $\tau$09 and from $\tau$10 to $\tau$11 at $\tau$12; for the PCA-disaggregated specification, CQR-QR reduces MAE from $\tau$13 to $\tau$14 at $\tau$15 and from $\tau$16 to $\tau$17 at $\tau$18 (Bogani et al., 2024).

Another controversy concerns tail extrapolation. “Robust Econometrics for Growth-at-Risk” argues that existing GaR procedures often impose a constant conditional Pareto exponent, either implicitly or through skew-$\tau$19 extrapolation. The paper models a covariate-dependent Pareto exponent $\tau$20, derives asymptotic normality for the resulting extreme-quantile estimator under regular variation and nonparametric CDF estimation, and proposes symmetric upper- and lower-tail algorithms based on tail-index regression plus kernel estimation of $\tau$21. In simulation, the proposed method yields lower RMSE, tighter interquartile ranges, and more reliable confidence-band coverage than the “Old” skew-$\tau$22 extrapolation method (Adrian et al., 1 Aug 2025).

Its empirical long-run application to U.S. data from 1893Q1 to 2016Q4 illustrates the calibration issue starkly. Over 488 out-of-sample quarters, the old method places only $\tau$23 of realizations below the forecast $\tau$24 percentile and $\tau$25 above the forecast $\tau$26 percentile; the new semiparametric method records $\tau$27 below and $\tau$28 above, close to nominal coverage. In counterfactual densities for next year’s GaR, the new method implies a $\tau$29 percentile of $\tau$30 in a “Normal year” and $\tau$31 in a “Crisis year,” while the old method gives $\tau$32 and $\tau$33, respectively (Adrian et al., 1 Aug 2025).

Three recurrent misconceptions are therefore addressed directly in the recent literature. First, a conditional-quantile slope need not identify an unconditional tail effect: adding controls in standard quantile regressions changes the estimand, whereas the structural GQR framework is designed precisely to avoid that slippage (Wojciechowski, 2024). Second, tail relations are not time-invariant: short samples centered on the Great Moderation can make upside risk appear artificially stable and can conceal reversals in the roles of credit and house prices (Gächter et al., 2023). Third, nominal quantile levels do not guarantee nominal coverage: both conformal calibration and robust tail-exponent estimation are motivated by systematic under- or over-coverage of standard procedures (Bogani et al., 2024, Adrian et al., 1 Aug 2025).

In current usage, GaR is best understood not as a single model class but as a family of tail-focused forecasting and identification strategies. The unifying object is the lower quantile of future growth; the main differences lie in whether the quantile is conditional or unconditional, static or time-varying, reduced-form or structural, low-dimensional or high-dimensional, and calibrated only asymptotically or also in finite samples.