Taming Tail Risk in Financial Markets: Conformal Risk Control for Nonstationary Portfolio VaR

Risk forecasts drive trading constraints and capital allocation, yet losses are nonstationary and regime-dependent. This paper studies sequential one-sided VaR control via conformal calibration. I propose regime-weighted conformal risk control (RWC), which calibrates a safety buffer from past forecast errors using exponential time decay and regime-similarity weights from regime features. RWC is model-agnostic and wraps any conditional quantile forecaster to target a desired exceedance rate. Finite-sample coverage is established under weighted exchangeability, and approximation bounds are derived under smoothly drifting regimes. On the CRSP U.S.\ equity portfolio, time-weighted conformal calibration is a strong default under drift, while regime weighting can improve regime-conditional stability in some settings with modest conservativeness changes.

💡 Research Summary

This paper addresses the challenge of maintaining reliable Value‑at‑Risk (VaR) forecasts in financial markets where loss distributions are non‑stationary, heavy‑tailed, and often organized into recurring regimes (e.g., calm versus stress periods). Classical conformal prediction provides finite‑sample coverage guarantees under exchangeability, but such an assumption is violated by sequential, drifting market data. The author therefore proposes Regime‑Weighted Conformal Risk Control (RWC), a model‑agnostic wrapper that calibrates a safety buffer on top of any conditional quantile forecaster (such as rolling‑window quantile regression or GARCH‑based VaR).

The method works as follows. At each time t a base forecaster produces a 1‑α conditional quantile (\hat q_t). The one‑sided conformity score is defined as (s_t = y_t - \hat q_t), where a large positive value indicates under‑prediction of tail risk. To compute the safety buffer (\hat c_t), past scores are weighted by two factors: (i) an exponential time decay (\exp(-\lambda (t-i))) that emphasizes recent observations, and (ii) a kernel similarity (K_h(z_i, z_t) = \exp(-|z_i - z_t|^2/(2h^2))) that gives higher weight to scores generated under market regimes similar to the current one. The regime embedding (z_t = g(x_t)) can be any fixed mapping of market features (e.g., volatility, trend, volume). Normalized weights (\tilde w_i(t)) are used to compute a weighted ((1-\alpha))-quantile of the past scores, which becomes the buffer (\hat c_t). The final VaR bound is (U_t = \hat q_t + \hat c_t).

Two theoretical results are established. First, under a “weighted exchangeability” assumption—meaning that the set of weighted scores ({(s_i, w_i(t))}) is exchangeable after conditioning on the current covariates—the paper proves an exact finite‑sample coverage guarantee: (P(y_t \le U_t \mid x_t, z_t) \ge 1-\alpha). This mirrors the classic (m+1) correction used in standard conformal inference, with a small inflation factor (\rho_t) that becomes negligible when the effective calibration weight is large. Second, recognizing that weighted exchangeability is unrealistic for finance, the author analyzes an approximate coverage bound under smooth regime drift. Assuming the conditional CDF of scores varies Lipschitz‑continuously in both regime space and time, the coverage gap (\epsilon_t) decomposes into three terms: (a) a bias proportional to the kernel bandwidth (h) (capturing regime similarity), (b) a drift term proportional to the product of the time‑drift constant (L_t) and the effective lag (\tau_t) (controlled by (\lambda)), and (c) a stochastic term of order (1/\sqrt{n_{\text{eff}}(t)}) reflecting sampling variability. This decomposition guides practitioners in selecting ((\lambda, h, m)) to balance adaptation speed, regime localization, and statistical stability.

Empirically, the method is evaluated on the CRSP value‑weighted U.S. equity index spanning 1990‑2024 (≈8,755 trading days). The dataset is split chronologically into training (1990‑2011), validation (2011‑2018), and testing (2018‑2024). Base forecasters include rolling‑window quantile regression, GARCH‑VaR, and a deep quantile network. RWC is compared against three baselines: Sliding‑Window Conformal (SWC) with an unweighted recent window, Time‑Weighted Conformal (TWC) using only the exponential decay, and Adaptive Conformal Inference (ACI) which dynamically adjusts the target exceedance level.

Results show that TWC alone already provides a strong default: it keeps the overall exceedance rate within ±0.5 % of the nominal 1 % level across most periods, with relatively small safety buffers (low conservatism). Adding regime similarity (RWC) yields modest additional conservatism but improves regime‑conditional stability, especially during high‑volatility regimes such as the 2008 financial crisis and the 2020 COVID‑19 shock. In those periods, the exceedance‑rate variance drops by 30‑40 % compared with TWC, while the average buffer size grows only slightly. SWC suffers from drift, with exceedance spikes exceeding 2 % during crises. ACI maintains the target exceedance by inflating the buffer dramatically, leading to overly conservative risk limits. The effective sample size (n_{\text{eff}}(t)) remains comfortably above 150 for typical hyper‑parameter choices ((\lambda=0.01, h=0.5) after standardizing regime features), ensuring the stochastic term in the coverage bound is negligible; when (n_{\text{eff}}(t)) falls below a preset threshold, the algorithm falls back to pure time weighting, preserving robustness.

The paper concludes that weighted conformal risk control offers a principled, distribution‑free approach to sequential VaR calibration under non‑stationarity. Time weighting alone (TWC) serves as a computationally simple and effective baseline, while regime weighting (RWC) can further stabilize performance in recurring market states at the cost of a modest increase in conservatism. Future work is suggested on (i) learning richer, possibly nonlinear regime embeddings via deep clustering or variational auto‑encoders, (ii) extending the framework to multivariate VaR/CVaR and portfolio‑level risk measures, and (iii) testing the methodology on high‑frequency data where dependence structures are stronger. Overall, the contribution bridges conformal prediction theory with practical risk management, providing both rigorous guarantees and actionable insights for practitioners dealing with drifting financial markets.