Reliable Real-Time Value at Risk Estimation via Quantile Regression Forest with Conformal Calibration

Rapidly evolving market conditions call for real-time risk monitoring, but its online estimation remains challenging. In this paper, we study the online estimation of one of the most widely used risk measures, Value at Risk (VaR). Its accurate and reliable estimation is essential for timely risk control and informed decision-making. We propose to use the quantile regression forest in the offline-simulation-online-estimation (OSOA) framework. Specifically, the quantile regression forest is trained offline to learn the relationship between the online VaR and risk factors, and real-time VaR estimates are then produced online by incorporating observed risk factors. To further ensure reliability, we develop a conformalized estimator that calibrates the online VaR estimates. To the best of our knowledge, we are the first to leverage conformal calibration to estimate real-time VaR reliably based on the OSOA formulation. Theoretical analysis establishes the consistency and coverage validity of the proposed estimators. Numerical experiments confirm the proposed method and demonstrate its effectiveness in practice.

💡 Research Summary

The paper tackles the challenging problem of real‑time Value‑at‑Risk (VaR) estimation in rapidly changing financial markets, where accuracy, timeliness, and reliability must be achieved simultaneously. The authors adopt the offline‑simulation‑online‑estimation (OSOA) framework: heavy Monte‑Carlo simulation and model training are performed offline, while online risk monitoring consists of a single fast evaluation of the trained model on the latest risk‑factor observations.

In the offline stage, a large synthetic dataset of risk‑factor vectors (X_i) and corresponding portfolio losses (L_i) is generated by nested simulation. The conditional (\alpha)-quantile function (v_\alpha(x)=Q_\alpha(L\mid X=x)) is then learned using Quantile Regression Forests (QRF). QRF builds an ensemble of regression trees with bootstrap resampling and random split rules; each leaf stores the empirical distribution of the training responses that fall into that leaf. The conditional quantile is obtained directly from this empirical distribution, allowing QRF to capture highly nonlinear, high‑dimensional relationships without extensive hyper‑parameter tuning.

During online operation, the observed risk‑factor vector (x(u)) is fed into the trained forest, and the estimated VaR (\hat v_\alpha(x(u))) is returned instantly, satisfying stringent latency requirements (e.g., the 15‑minute “4:15 report”).

A key contribution is the integration of conformal calibration to guarantee the nominal coverage of the VaR estimate. After the QRF model is trained, a held‑out calibration set is used to compute residuals (R_i = L_i - \hat v_\alpha(X_i)). The empirical ((1-\alpha))-quantile of these residuals, (\hat q_{1-\alpha}(R)), is added to the raw QRF prediction, yielding the conformalized VaR (\hat v^{\text{conf}}\alpha(x) = \hat v\alpha(x) + \hat q_{1-\alpha}(R)). This correction is model‑agnostic and provides finite‑sample marginal coverage: for any real‑time risk‑factor vector, the true loss falls below the conformalized estimate with probability at least (\alpha). The authors also prove that as the offline sample size grows, the coverage guarantee strengthens to conditional validity.

Theoretical analysis establishes two main results. First, pointwise and (L^2) consistency of the QRF estimator are proved, showing that (\hat v_\alpha(x)) converges to the true conditional quantile as the number of simulated paths tends to infinity. Second, the conformalized estimator enjoys finite‑sample coverage guarantees without distributional assumptions, and the guarantee improves to conditional coverage under increasing offline data.

Empirical evaluation uses realistic high‑dimensional portfolios with complex derivatives. Performance metrics include mean root integrated squared error, mean pinball loss, and empirical coverage rate. Results demonstrate that the plain QRF achieves competitive accuracy but falls short of the target coverage (e.g., 99%). After conformal calibration, the coverage aligns closely with the nominal level while maintaining comparable or even improved accuracy, indicating that the calibration does not introduce excessive conservatism.

In summary, the paper makes four contributions: (1) formulation of real‑time VaR estimation as a conditional quantile regression problem within OSOA; (2) deployment of QRF to handle nonlinear, high‑dimensional risk‑factor–loss relationships; (3) development of a conformalized QRF estimator that provides rigorous, finite‑sample coverage guarantees; and (4) comprehensive theoretical and numerical validation of both components. The work bridges modern machine‑learning quantile methods with regulatory‑grade risk measurement, offering a practical pathway for institutions to obtain reliable, real‑time VaR estimates. Future research directions include studying the sensitivity of coverage to calibration‑set size, handling distributional drift in live data streams, and extending the framework to other risk measures such as Expected Shortfall.