Factor-Augmented Quantile Regressions (FA-QRs)

• FA-QRs are quantile regression frameworks that incorporate low-dimensional latent factors to capture distributional heterogeneity, tail dependence, and high-dimensional comovement.
• They include various constructions such as low-rank multitask models, time-series quantile autoregressions, and panel quantile factor models, each tailoring factor structure for distinct empirical scenarios.
• Estimation approaches range from nuclear-norm penalization and alternating minimization to Bayesian and convolution-smoothed methods, ensuring robust inference and enhanced predictive accuracy.

Factor-Augmented Quantile Regressions (FA-QRs) are quantile-regression frameworks in which conditional quantiles are driven by a low-dimensional latent factor structure. In the literature, this label covers several closely related constructions: multivariate quantile regressions whose coefficient matrix is low rank and therefore factorisable; time-series quantile autoregressions augmented with estimated factors from large predictor panels; panel quantile factor models in which both factors and loadings may vary with the quantile index; and characteristic-based or matrix-valued formulations that impose additional structure on the loadings. Across these variants, the central objective is to model distributional heterogeneity, tail dependence, and high-dimensional comovement without reducing the problem to conditional means (Chao et al., 2015).

1. Core formulations

A canonical time-series formulation augments a quantile autoregression with latent factors extracted from a large information set. Let $y_t$ denote the target variable, $V_t$ a vector of targeted macro variables, and $F_t$ a vector of latent factors. With $Z_t=(Y_t,V_t,F_t)$, the $\tau$-quantile forecast is built from linear specifications of the form

$y_{t+1}=\theta_{(m)}'Z_{t(m)}+\epsilon_{t+1},$

and an FA-QAR is simply a candidate model in which one or more components of $F_t$ appear in $Z_{t(m)}$ (Phella, 2020). A closely related forecasting specification writes the $h$-step-ahead conditional quantile as

$$q_{\tau}(y_{t+h}\mid y_t,F_t)=\mu(\tau,h)+\phi(\tau,h)y_t+\beta(\tau,h)'F_t,$$

with estimated factors treated as observed regressors in the quantile stage (Bellocca et al., 14 Jul 2025).

A second formulation embeds factor structure directly in a multivariate quantile-regression coefficient matrix. In the factorisable multitask model, for responses $V_t$0 and predictors $V_t$1,

$V_t$2

so the matrix of quantile coefficients $V_t$3 has rank $V_t$4. When $V_t$5, the model becomes a one-shot low-rank multivariate quantile regression in which many conditional quantiles share a small number of quantile-specific latent factors (Chao et al., 2015).

A third formulation is the quantile factor model for panel data. At a fixed quantile,

$V_t$6

or equivalently

$V_t$7

Here both factor loadings and factor realizations may depend on $V_t$8, so different parts of the conditional distribution can be driven by different latent structures (Chen et al., 2019).

Several extensions enrich the loading structure. In the single-index quantile characteristics-augmented factor model,

$V_t$9

so factor exposures vary nonlinearly with observed characteristics through one-dimensional indices (Xu et al., 24 Jun 2025). For matrix-valued observations $F_t$0, the matrix quantile factor model posits

$F_t$1

which is a structured FA-QR with row and column loadings constrained to a Kronecker form after vectorization (Kong et al., 2022). A more recent direct FAQR specification instead decomposes high-dimensional covariates as $F_t$2 and models

$F_t$3

so dense effects are carried by common factors and sparse effects by idiosyncratic components (Wei et al., 1 Aug 2025).

2. Estimation strategies

The low-rank multitask approach estimates $F_t$4 by nuclear-norm penalized quantile regression,

$F_t$5

where $F_t$6 is empirical check-loss risk and $F_t$7 is the nuclear norm. Because the objective is non-smooth and repeated SVDs are expensive, the paper smooths the quantile loss by Nesterov smoothing and solves the penalized problem with FISTA, using singular-value soft-thresholding in the proximal step. Approximate optimization error is tracked explicitly through a tolerance $F_t$8 satisfying

$F_t$9

(Chao et al., 2015).

Quantile factor models for large panels typically use alternating minimization. In Quantile Factor Analysis, for fixed $Z_t=(Y_t,V_t,F_t)$0 one minimizes the average check loss jointly over loadings and factors, but solves the problem iteratively: given current factors, update each loading by a cross-sectional quantile regression; given current loadings, update each factor by a time-series quantile regression; then renormalize to satisfy factor-model identification restrictions. This yields the iterative quantile regression algorithm that plays the same role for quantile factors that alternating least squares plays for PCA (Chen et al., 2019).

Characteristic-based models use a projection-first architecture. In Quantile-Projected PCA, each time $Z_t=(Y_t,V_t,F_t)$1 is associated with a sieve quantile regression of $Z_t=(Y_t,V_t,F_t)$2 on basis functions of characteristics $Z_t=(Y_t,V_t,F_t)$3. The fitted values $Z_t=(Y_t,V_t,F_t)$4 are stacked into a matrix $Z_t=(Y_t,V_t,F_t)$5, PCA is applied to $Z_t=(Y_t,V_t,F_t)$6 to recover factors, and the first-stage sieve coefficients are then regressed on the estimated factors to reconstruct the loading functions $Z_t=(Y_t,V_t,F_t)$7. This three-stage procedure is designed for panels with large $Z_t=(Y_t,V_t,F_t)$8, possibly small $Z_t=(Y_t,V_t,F_t)$9, and heavy-tailed idiosyncratic errors (Chen et al., 2023).

Composite and probabilistic approaches alter the factor-extraction criterion rather than the downstream quantile regression. The composite quantile factor model minimizes a sum of check losses across a grid of quantiles, keeping factors and loadings common across those quantiles but estimating multiple intercepts $\tau$0; the optimization proceeds by alternating updates for factors, loadings, and intercepts, with majorization-minimization inside each subproblem (Huang, 2023). Probabilistic quantile factor analysis and Bayesian quantile factor models instead use the asymmetric Laplace representation of quantile regression. The probabilistic version employs a variational Bayes approximation with sparse Bayesian learning priors on loadings, while the Bayesian version uses MCMC with a Normal-Exponential mixture representation of the asymmetric Laplace distribution to obtain conditional Gaussian updates for factors and loadings (Korobilis et al., 2022); (Gonçalves et al., 2020).

The direct high-dimensional FAQR model with dense and sparse effects smooths the check loss by convolution and estimates the combined parameter vector $\tau$1 using an $\tau$2-penalized objective based on estimated factors and idiosyncratic components. The explicit gradient and Hessian of the smoothed loss permit standard gradient-based optimization rather than subgradient-only methods (Wei et al., 1 Aug 2025).

3. Identification, rates, and inference

Identification in FA-QRs is rarely purely algebraic; it is imposed through normalization, rank restrictions, or structured penalties. In low-rank multitask quantile regression, the factorization $\tau$3 is recovered only up to singular-vector conventions, and a particular decomposition can be fixed by an SVD of the estimated coefficient matrix. In panel quantile factor models, the standard normalization is

$\tau$4

diagonal with ordered entries, while matrix quantile factor models impose orthogonality separately on row and column loading spaces (Chao et al., 2015); (Chen et al., 2019); (Kong et al., 2022).

The theoretical literature establishes both non-asymptotic and asymptotic guarantees. For the nuclear-norm estimator, if the optimization error is of order $\tau$5 or smaller, the approximate estimator is asymptotically as good as the exact optimizer; the resulting prediction and Frobenius-norm bounds match, up to logarithmic factors, the low-rank multivariate mean-regression rates familiar from matrix estimation (Chao et al., 2015). In Quantile Factor Analysis, the unsmoothed estimator satisfies

$\tau$6

with $\tau$7, and a smoothed version yields $\tau$8-asymptotic normality for loadings and $\tau$9-asymptotic normality for factors (Chen et al., 2019).

Matrix structure improves efficiency. For matrix-valued quantile factor models, the average Frobenius error converges at rate

$y_{t+1}=\theta_{(m)}'Z_{t(m)}+\epsilon_{t+1},$0

which is faster than the rate obtained by flattening the matrix into a vector factor model and ignoring the Kronecker loading restriction (Kong et al., 2022). Characteristic-based Quantile-Projected PCA is explicitly developed for heavy tails and even fixed $y_{t+1}=\theta_{(m)}'Z_{t(m)}+\epsilon_{t+1},$1; the factor estimates converge at rate $y_{t+1}=\theta_{(m)}'Z_{t(m)}+\epsilon_{t+1},$2, and under additional structure can attain $y_{t+1}=\theta_{(m)}'Z_{t(m)}+\epsilon_{t+1},$3, while the estimated loading functions admit asymptotic normality after appropriate normalization (Chen et al., 2023).

Recent work extends this theory to more complex loading maps. The single-index quantile factor model establishes asymptotic properties for latent factors, loading functions, and index parameters under a three-step sieve procedure, with $y_{t+1}=\theta_{(m)}'Z_{t(m)}+\epsilon_{t+1},$4-rate factor estimation and $y_{t+1}=\theta_{(m)}'Z_{t(m)}+\epsilon_{t+1},$5-rate asymptotic normality for the index parameters under the stated growth conditions (Xu et al., 24 Jun 2025). Robust quantile factor analysis addresses weak factors by requiring each factor to have a strong effect only near some unknown quantile level. Its inverse-density-weighted estimator achieves

$y_{t+1}=\theta_{(m)}'Z_{t(m)}+\epsilon_{t+1},$6

together with $y_{t+1}=\theta_{(m)}'Z_{t(m)}+\epsilon_{t+1},$7- and $y_{t+1}=\theta_{(m)}'Z_{t(m)}+\epsilon_{t+1},$8-asymptotic normality for factors and loadings, and consistent selectors for the total number of factors and for factors of any desired strength (Chen et al., 27 Jan 2025).

This body of results suggests that FA-QR inference depends critically on how factors are obtained. One-shot low-rank estimators, fixed-$y_{t+1}=\theta_{(m)}'Z_{t(m)}+\epsilon_{t+1},$9 quantile factor estimators, composite-quantile procedures, and weak-factor-robust estimators solve different identification problems and therefore support different asymptotic regimes.

4. Factor construction across quantiles

A central conceptual distinction in the FA-QR literature concerns whether factors are themselves quantile-specific. In fixed-$F_t$0 quantile factor models, $F_t$1 and $F_t$2 both vary with $F_t$3, so the lower tail, center, and upper tail may be governed by different latent spaces (Chen et al., 2019). The factorisable multitask model is similar in spirit: $F_t$4, $F_t$5, and even the number of factors $F_t$6 are allowed to change with $F_t$7, but the factors are functions of observed regressors rather than latent drivers of the joint distribution of $F_t$8 and $F_t$9 (Chao et al., 2015).

By contrast, some procedures estimate factors that are common across multiple quantiles. In the composite quantile factor model, the same $Z_{t(m)}$0 and $Z_{t(m)}$1 are estimated jointly from several quantile losses, while only the intercept terms $Z_{t(m)}$2 vary with $Z_{t(m)}$3. The stated interpretation is that the resulting factors estimate the mean factor model but are obtained using information across multiple quantiles, thereby adapting to skewness, kurtosis, and tail behavior in the panel (Huang, 2023). Robust quantile factor analysis takes a different route: factors are invariant in $Z_{t(m)}$4, but loadings are quantile-specific and strength is enforced only after integrating over quantiles, so factors can be arbitrarily weak in the mean or at most individual quantiles (Chen et al., 27 Jan 2025).

Probabilistic formulations change the inferential machinery rather than the fundamental decomposition. Probabilistic quantile factor analysis builds a quantile-specific latent factor model from an asymmetric Laplace likelihood, sparse Bayesian learning priors, and variational Bayes updates, and uses the resulting factors to construct “low,” “medium,” and “high” uncertainty indexes and “loose,” “median,” and “tight” financial conditions indexes (Korobilis et al., 2022). Bayesian quantile factor models use an analogous asymmetric-Laplace mixture representation with MCMC, retaining the factor-model interpretation while making the quantile restriction explicit in the likelihood (Gonçalves et al., 2020).

This suggests that FA-QR is not a single canonical estimator but a design space. In some papers factors are latent regressors extracted from a large panel; in others they are low-rank coefficient directions; in still others they are characteristic-driven or matrix-structured objects. The shared element is the reduction of a high-dimensional quantile problem to a small latent factor system.

5. Empirical domains and findings

Macroeconomic forecasting is a principal application. In UK data, model-averaged factor-augmented quantile autoregressions are built from lagged target variables, targeted macro variables, and two recursively estimated principal-component factors. For GDP growth, equal weights and AIC/BIC weights perform similarly, but QRIC and Jackknife weights outperform on the majority of the quantiles of interest in terms of coverage and final prediction error. For CPI inflation, by contrast, the naive QAR(1) outperforms all model-averaging methodologies, indicating that factor augmentation is variable-specific rather than uniformly beneficial (Phella, 2020).

Finance provides both multivariate and panel applications. In the factorisable multitask model’s systemic-risk illustration, daily returns of $Z_{t(m)}$5 financial institutions from 2000 to 2010 are regressed on lagged absolute and negative returns of all institutions. At $Z_{t(m)}$6 and $Z_{t(m)}$7, the estimated rank is $Z_{t(m)}$8, implying one dominant tail factor in each tail. The estimated tail factors deviate sharply from zero during the 2008–2009 crisis, and institutions with large market capitalization and high leverage exhibit the largest left-tail loadings, interpreted as high exposure to downside systemic risk (Chao et al., 2015).

The scenario-analysis literature extends FA-QRs from quantile forecasting to density and stress analysis. In the FARS framework, factors are first extracted from a multi-level dynamic factor model, then inserted into quantile regressions, and finally mapped into a full conditional density via a parametric skew-$Z_{t(m)}$9 fit to estimated conditional quantiles. Stressing the factors by moving them to the boundary of factor confidence ellipsoids produces stressed quantiles and densities, yielding measures such as Growth-at-Risk and Growth-in-Stress. In the US GDP application, Growth-in-Stress is more negative than Growth-at-Risk across horizons, especially around COVID-19, because the stressed scenario allows joint factor realizations in extreme regions of the estimated factor distribution (Bellocca et al., 14 Jul 2025).

Asset-pricing applications emphasize characteristic-driven heterogeneity. The single-index quantile factor model for US corporate bond returns from 2003 to 2020 outperforms both the benchmark quantile Fama-French five-factor model and the quantile latent factor model, particularly in the tails $h$0 and $h$1. The estimated loading functions reveal state-dependent risk exposures driven by characteristics such as bond and equity volatility, coupon, and spread, and the latent factors are given economic interpretations that differ across the lower tail, median, and upper tail (Xu et al., 24 Jun 2025).

Other empirical studies indicate that tail factors often contain information not present in mean factors. Quantile factor models have been used to study macroeconomic, climate, and financial panels; additional factors at $h$2 improve US GDP density forecasting relative to AR and AR-plus-PCA benchmarks, lower-quantile temperature factors exhibit strong Granger causality from global $h$3 growth, and finance applications show that inter-quantile-range factors are highly correlated with volatility factors constructed from squared residuals (Chen et al., 2019). Probabilistic quantile factor analysis similarly reports that high uncertainty and tight financial conditions indexes have superior predictive ability for various measures of economic activity, and that quantile factors extracted from about $h$4 daily financial series can outperform mean or median factors out of sample (Korobilis et al., 2022).

6. Software, diagnostics, and unresolved design choices

FA-QR implementation has begun to consolidate in software. The FARS package in R provides a framework for extracting global and block-specific factors from multi-level dynamic factor models, computing asymptotically valid confidence regions for estimated factors, estimating FA-QRs, recovering full predictive densities from quantile forecasts, and evaluating stressed densities when factors are moved to extreme regions of their estimated joint distribution (Bellocca et al., 14 Jul 2025). For factor extraction, the companion R package cqrfactor implements both composite and single-quantile factor estimation from large panels, with the composite quantile factor model and its factor-number information criterion available as part of the workflow (Huang, 2023).

Several design questions remain model-specific. In the forecasting FA-QAR literature, each quantile is estimated separately, model-averaging weights may differ by quantile, and there is no explicit mechanism to enforce non-crossing (Phella, 2020). In scenario-based FA-QRs, skew-$h$5 smoothing is a parametric approximation to the conditional distribution, and the factor structure drops idiosyncratic predictor information after dimension reduction, so misspecification of the factor model can contaminate the regression and density stages (Bellocca et al., 14 Jul 2025). These are not contradictions; they reflect different intended uses of FA-QRs, from forecasting and density recovery to structural factor interpretation.

Factor-number selection is itself an active subfield. Fixed-$h$6 quantile factor models use rank minimization and information criteria based on check loss (Chen et al., 2019); matrix-valued models add row/column-specific rank, information-criterion, and eigenvalue-ratio procedures (Kong et al., 2022); characteristic-based models use eigenvalue thresholding on the covariance of fitted quantile projections (Chen et al., 2023); and weak-factor-robust methods provide selectors for both the total number of factors and the subset of factors of any desired strength (Chen et al., 27 Jan 2025). This suggests that the effective dimension of an FA-QR is generally quantile-dependent and cannot be treated as a purely mean-based property of the data.

Recent direct formulations also point toward a broader FAQR agenda. The dense-plus-sparse model decomposes high-dimensional covariates into common factors and idiosyncratic components, uses convolution smoothing to regularize the quantile loss, proves factor-selection accuracy and parameter consistency under mild regularity conditions, and introduces a Bootstrap-based diagnostic procedure for testing factor adequacy (Wei et al., 1 Aug 2025). A plausible implication is that future FA-QR work will continue to move beyond the simple “PCA first, quantile regression second” template toward joint, structurally constrained, and diagnostically explicit distributional models.