Risk managers who distrust any single volatility model describe their uncertainty as a set of plausible models and evaluate a position under the worst one. They then re-evaluate it every day as new prices arrive. For the daily re-evaluations to stay consistent with the original verdict, the model set must be pasting-stable. Pasting stability means that if you follow one plausible model up to some random date and switch to another plausible model afterwards, the spliced model must still be in the set. Without this property, a position judged safe today can become unacceptable tomorrow even though nothing surprising happened. The natural way to build such sets in function space is a ball around a reference covariance under an information divergence. That approach runs into a basic obstacle: two Gaussian processes with different volatilities are typically mutually singular (they put probability one on disjoint sets of paths), so their Kullback-Leibler (KL) divergence is infinite. Every ball is then either the reference model alone or everything.
Operator unitization replaces a covariance operator $C$ by $C+\gamma I$ so that inverses and log-determinants become finite. This yields finite divergences even between singular Gaussian measures. This note proposes to use that regularization as the building block of pasting-stable ambiguity sets. The key observation is that unitization has a probabilistic reading: it is the covariance of the process observed with independent noise of variance $\gamma$. Under that reading, the regularized divergence should obey a chain rule over time. The chain rule makes a step-by-step ("rectangular") divergence budget pasting-stable by construction. It also makes a whole-path ball provably not pasting-stable.
We would build the resulting dynamic risk measure as a cheap backward recursion. We would choose $\gamma$ from data as the estimated microstructure noise level. We would then test the measure on the Oxford-Man realized library and CTBench, where none of the pasting-stability theory has ever been evaluated.
Core Claim
The proposal makes three claims, in decreasing order of certainty.
Chain rule. Let $D_\gamma(P,Q)$ be the unitized KL divergence between Gaussian path measures $P$ and $Q$ of Regularized divergences between covariance operators and Gaussian measures on Hilbert spaces. That paper defines regularized KL and Rényi divergences via $C+\gamma I$ and proves convergence as $\gamma\to 0$ for equivalent measures. It runs no experiments and leaves the singular case uninterpreted. We claim that on any time discretization $D_\gamma$ equals the ordinary KL divergence between the laws of the noisily observed processes. We further claim it splits at a stopping time $\tau$ as
$$D_\gamma(P_1\otimes_\tau P_2,\;Q)=D_\gamma\big(P_1|_{\mathcal F^\gamma_\tau},\,Q|_{\mathcal F^\gamma_\tau}\big)+\mathbb E_{P_1}\Big[D_\gamma\big(P_2(\cdot\mid\mathcal F^\gamma_\tau),\,Q(\cdot\mid\mathcal F^\gamma_\tau)\big)\Big].$$
Here $P_1\otimes_\tau P_2$ is the pasted measure, which follows $P_1$ until $\tau$ and $P_2$'s conditional law afterwards. $\mathcal F^\gamma_\tau$ is the information generated by the noisy observations up to $\tau$. The conditional covariances are Schur complements of the unitized operator. Concretely, they are the posterior covariances of a Kalman filter with observation noise $\gamma$.
Pasting stability. Two consequences follow. A global ball $\{P: D_\gamma(P,Q)\le r\}$ is not pasting-stable, because a pasted measure can reach divergence near $2r$. A rectangular set, which caps the conditional per-step divergence at $r_t$ almost surely, is pasting-stable. In the dual-penalty language of Cheridito, Delbaen and Kupper, it corresponds to an additive concatenation-stable penalty. Their work, Dynamic Monetary Risk Measures for Bounded Discrete-Time Processes and Coherent and convex monetary risk measures for unbounded càdlàg processes, shows that time-consistency is equivalent to stability of the dual penalty under concatenation. That equivalence supplies the target property we must verify.
Practical value. The rectangular unitized set gives tail-risk estimates that are robust to volatility regime shifts at a modest cost in conservatism. It does this while staying finite as sampling frequency increases, which is exactly where unregularized KL sets break.
Research Program
Theory. The theoretical part proves the chain rule for discretized Gaussian processes and then passes to the continuum. The main technical risk is that pasting must be defined relative to the noisy filtration $\mathcal F^\gamma$, not the clean path filtration used by Constructing sublinear expectations on path space. That paper builds time-consistent sublinear expectations on path space, without a dominating measure, assuming measurability, invariance, and pasting stability. We do not use its construction directly. Instead we show that our rectangular set, pulled back to clean paths, satisfies its three hypotheses, so its tower property applies. If the pullback fails, the measure is time-consistent only with respect to what a noisy observer can know. We would report that as a substantive finding rather than hide it.
Algorithm. For a linear position, the worst case inside a per-step budget is a one-dimensional problem: maximize conditional variance subject to a log-determinant constraint. A Lagrangian solves it in closed form up to a scalar root-find. The dynamic risk measure is then a backward recursion with a Kalman-type forward pass. Its cost per step is cubic in the local window rather than the $O(N^2(K+N))$ fitting cost reported for the global kernel Mahalanobis approach in Infinite-dimensional Mahalanobis Distance with Applications to Kernelized Novelty Detection.
Choosing γ from data. That same paper identifies sensitivity to the regularization level as an open problem with no principled selection rule. The noise interpretation gives one: set $\gamma$ to the estimated microstructure noise variance, which the high-frequency literature estimates routinely from the gap between realized variance at different frequencies.
Singular limit. As a secondary analysis, we study how $D_\gamma$ grows as $\gamma\to0$ for singular pairs. We expect logarithmic growth in $1/\gamma$, with a coefficient that counts the effectively distinct spectral directions. That would give the open singular-case question a concrete answer.
Baselines
Naive reference. The do-nothing reference is rolling historical-simulation Value-at-Risk (VaR) and Expected Shortfall (ES), with no model and no ambiguity. It is run with the same window lengths as every other method.
Strongest existing method. The strongest competitor is Deep Learning Enhanced Realized GARCH from Deep Learning Enhanced Realized GARCH. It augments realized GARCH, a model that drives daily volatility with 5-minute realized variance, using a neural network. On the 31-index Oxford-Man library it reports superior Bayes factors on 28 of 31 indices and top tail-risk (VaR/ES) performance in 22–23 of 31 categories. The file gives no per-index decimal values. We re-run it and linear RealGARCH on our splits with the authors' tuning budget, and we give our method the same budget.
On CTBench, CTBench: Cryptocurrency Time Series Generation Benchmark reports that ARMA-GARCH has the worst tail risk and FIDE has tighter risk control. The file gives no numeric Sharpe ratio or ES values for either.
Beating DeepRGARCH means one of two things on held-out years that include 2008 and 2020:
- fewer ES backtest rejections at equal VaR coverage, or
- a statistically indistinguishable backtest with zero time-consistency violations, versus a measurable rate for the global-ball variant.
Ablations. These isolate the key ingredient. The first is the global unitized ball, identical in every respect except non-rectangularity, which isolates pasting stability. The second is a rectangular KL ball with $\gamma\approx0$, which isolates unitization. The third is a rectangular Hilbert-Schmidt ball and the fourth a rectangular Bures-Wasserstein ball, which isolate the choice of divergence. The fifth is a fixed cross-validated $\gamma$ versus the noise-estimated $\gamma$, which isolates the selection rule.
Why not just…?
Why not just use a single good model like DeepRGARCH, which is trivially time-consistent? A single model satisfies the tower property automatically, so this alternative is a genuine threat. Its weakness is misspecification at regime shifts, not inconsistency. The Oxford-Man and CTBench backtests compare tail losses in crash windows directly. If DeepRGARCH matches our method there, we will report that the ambiguity layer buys nothing empirically on these benchmarks. We would also report a combined variant, a rectangular ball around DeepRGARCH's conditional covariance, which is the more likely practical winner.
Why not just use a rectangular KL ball on the discretized series? Finite-dimensional KL is finite. At daily frequency this may work about as well, and we say so. Volatility uncertainty, however, makes path measures singular in the continuum limit, so discretized KL between models differing only in volatility should grow roughly linearly in the number of observations per step. A fixed radius therefore becomes meaningless as one moves to 5-minute data. The decisive plot is required radius versus sampling frequency for KL and for $D_\gamma$. Unitization is justified only if $D_\gamma$ stays flat while KL grows.
Why not just use Bures-Wasserstein or Hilbert-Schmidt balls, which need no inverse at all? Their rectangular versions are also pasting-stable, so the question is purely about discriminating power. Estimation of Riemannian distances between covariance operators and Gaussian processes estimates unitized Riemannian distances from Gram matrices at an $O(1/\sqrt N)$ rate on synthetic centered Gauss-Markov processes. In that study the unitized metrics reach near 0% classification error, Hilbert-Schmidt reaches up to 33% error, and Bures-Wasserstein degrades to 29.2% error in 5D. That study is static and synthetic, so we rerun the comparison inside the dynamic risk task as ablations.
Why not just accept a global ball and live with some inconsistency? We measure how much inconsistency there actually is. If preference reversals are rare in practice, the theoretical argument loses force.
Proposed Experiments
Synthetic validation. We first verify the chain rule numerically on the synthetic Gauss-Markov processes of Estimation of Riemannian distances between covariance operators and Gaussian processes. We extend them with non-zero means and a volatility switch at a random time, which that paper excluded. The decisive check is that the left- and right-hand sides of the chain rule agree to numerical precision across grid refinements. Alongside it, we confirm empirically that pasted measures escape the global ball.
Time-consistency audit. On the Oxford-Man library (31 indices, 2004–2022, RV5 and daily returns) we run a rolling audit. For each method we record positions accepted at time $t$ whose re-evaluation at $t+k$ rejects them along paths with ordinary realized returns. We report this preference-reversal rate together with VaR violation rates, ES backtests, and wall-clock time.
Regime-shift robustness. On CTBench (452 Binance tokens, hourly, 2020–2024) we repeat the tail backtest. We also run a risk-constrained statistical-arbitrage overlay under the benchmark's 0.03% fee, to test whether robustness survives trading costs.
Decisive outcomes. Three results would settle the question. The rectangular set should show zero reversals while the global ball shows a positive rate. Unitized radii should stay stable across sampling frequencies where KL radii blow up. Tail-loss backtests in the 2008 and 2020 windows should be no worse than DeepRGARCH. If the third check fails while the first two hold, the contribution narrows to a consistency guarantee bought at measurable conservatism, and we would quantify that price.
References
- Regularized Divergences Between Covariance Operators and Gaussian Measures on Hilbert Spaces
- Dynamic Monetary Risk Measures for Bounded Discrete-Time Processes
- Coherent and convex monetary risk measures for unbounded càdlàg processes
- Constructing sublinear expectations on path space
- Infinite-dimensional Mahalanobis Distance with Applications to Kernelized Novelty Detection
- Deep Learning Enhanced Realized GARCH
- CTBench: Cryptocurrency Time Series Generation Benchmark
- Estimation of Riemannian distances between covariance operators and Gaussian processes