A conformal band around a VaR forecast can look perfectly calibrated on live data and still sit in the wrong place. Delay the centre forecast by 30 days and coverage of the true VaR falls to 0.54, while the realized breach rate the algorithm controls barely moves. Ivancevic, Nguyen and Luo show this on their own method. It matters because capital charges, position limits and margin calls all scale with the reported number.

A band around someone else's VaR

Point-VaR backtesting tells you whether breaches arrive at about the expected rate. Such procedures, the authors write, say little about the uncertainty of the estimated quantile. QDtACI, their quantile-adapted version of dynamically-tuned adaptive conformal inference, sits on top of an existing VaR series and produces an interval around it, using only realized returns and the forecast. It wraps whatever VaR model you already run.

The machinery is Gibbs and Candès (2024) retargeted at a quantile. The paper describes QDtACI as departing from that canonical method in three aspects: composite expert scoring, asymmetric interval construction, and an optional rank transformation of the conformity scores. The nonconformity score becomes the pinball loss, minimised at the a-quantile rather than the conditional mean. The interval's lower and upper edges are read at separate quantile positions, scaled by the recent realized violation rate. Forty experts run in parallel over learning rates log-spaced on [1e-4, 1e-1], each driving its own confidence level from an exponentially smoothed violation error rather than a binary pinball gradient. Weights come from a composite loss on coverage, width and stability at 0.90/0.05/0.05.

The synthetic evidence is 3,100 GARCH(1,1) observations with alpha 0.10, beta 0.85 and Gaussian innovations, so the true VaR is known. At the 95% level QDtACI covers the true VaR 0.961 of the time on a GARCH centre (n=2,600) and 0.936 on a CAViaR centre (n=2,850). The Delta method manages 0.895 and 0.853; a parametric bootstrap manages 0.580 and 0.858. At 99% QDtACI posts the lowest Winkler score on both centres, 0.038 and 0.047 against Delta's 0.058 and 0.049. The real application wraps CAViaR, DCC-GARCH (dynamic conditional correlation) and copula forecasts on 24 senior fixed-income bonds from Bloomberg, October 2014 to October 2024, with 2,290 evaluation days. Mean 95% widths run 2.4 to 2.9 times calm levels during the COVID crash and 2.0 to 2.3 times during the 2022-23 tightening. Width correlates with realized 30-day volatility at r between 0.49 and 0.60.

Most of the recipe is printed

Credit first. Most of the recipe is printed: forty experts, a 250-day calibration window, and the directional violation-rate window given exactly as the most recent min(50, W/5) observations. There is pseudocode, and an anonymized reference implementation, with the full repository promised on acceptance.

The optional rank transformation is handled well too. The authors motivate it and test it at curvature 0, 1, 2, 3 on the two well-specified centres at the 90, 95 and 99% levels. It is harmful in every cell of that two-centre table. On the GARCH centre at 95%, coverage drops from 0.961 to 0.898 as curvature rises to 3, width stays near 0.028, and the Winkler score worsens from 0.0301 to 0.0341. They report the negative result and set the transform to the identity. Every figure above refers to that identity version, which is the deployed one.

Where does a* come from?

The setpoint the experts drive towards is a, a target breach rate distinct from the 5% VaR level. The method section says where a comes from: it is calibrated on synthetic data, where the true VaR is observable, as the breach rate that yields nominal coverage. It declines roughly linearly from about 0.04 at 90% confidence to about 0.025 at 99%. We did not find the 95% value in the text. You interpolate between 0.04 and 0.025 and hope.

a is the judgment call, and not a peripheral one. It fixes the width and therefore the coverage. It is also fitted against the exact quantity the method is later evaluated on. The limitations concede the wider point: coverage of the true VaR is verifiable only on synthetic data, and the calibration is carried over from that setting. The authors add that the variation in coverage across the three real-data forecasts gives some evidence on how sensitive the transfer is, though it cannot confirm the calibration holds for the real return distribution. Fair as far as it goes. The synthetic generator is Gaussian GARCH(1,1) and the GARCH centre is the true data-generating process, so a is tuned in close to a best case, then shipped to a decade of bond returns.

Two smaller gaps we hit. We did not find an initial value for each expert's smoothed error. And while the directional violation-rate window is specified exactly, we did not find a window length for the rolling violation rate that enters the composite loss.

The gate the authors built, and where it stops binding

The authors concede the core objection themselves. Because true coverage cannot be measured on real data, they supply a return-only diagnostic that indicates when interval calibration can be trusted as a proxy for coverage of the true VaR. Run the standard point-VaR backtests (Kupiec, Christoffersen, DQ) on the centre first; failing any of them flags a forecast.

The lag experiment calibrates that gate. Replace the centre with the true VaR delayed by k days: coverage falls from roughly 1.0 at k=0 to 0.59 at k=10 and 0.54 at k=30, while the breach rate barely moves. The band is the right size and in the wrong place. On the synthetic centres the mapping is clean. GARCH passes (Kupiec p=0.830) and gives 0.961; CAViaR passes (p=0.764) and gives 0.936; historical simulation fails (Kupiec p=0.007, Christoffersen p=0.006, 6.1% violation rate) and gives 0.665.

Three centres, one failure. The diagnostic's whole calibration rests on that.

On the real portfolio all three centres pass (violation rates 4.98%, 5.02%, 4.85%; Kupiec p of 0.965, 0.965, 0.810), so the gate never binds in the application. The copula's Christoffersen p of 0.055 is the closest anything comes to rejection, and the authors flag it themselves, along with finite-sample power and in-sample overfitting as limits on the diagnostic. So the real-data bands are presented on the strength of a gate that has nothing to reject.

The authors note the bootstrap is restricted to parameter uncertainty. We would read its 0.580 on the well-specified GARCH centre as a weak comparator rather than a like-for-like alternative. The DtACI baseline is the more instructive one. At 95% it covers 1.000 and 0.998, purely by running about three times wider (0.087 against 0.028 on GARCH), with Winkler stuck between 0.084 and 0.091 in every cell. Coverage without a width penalty is close to free.

We could not test any of this ourselves. The paper's portfolio is 24 specific bonds behind a Bloomberg subscription, and we have no price history for those instruments. A substitute book of liquid US fixed-income ETFs would run the wrapper correctly, since it needs only a return series and a VaR series, but it would describe a different portfolio.

What would move me is a synthetic experiment under a heavy-tailed or regime-switching generator, with a* held at the value fitted under Gaussian GARCH. If coverage holds there, the transfer to real data is credible. Until then these are breach-calibrated bands, and we have seen before that conformal coverage holding is not the same as the downstream use paying off (our note on conformal Kelly). We did not find a decision rule attached to the widths in this paper; the practical use is argued rather than demonstrated.