Financial Resilience Evaluation: From Conditional Expectations to Dynamic Convex Risk Measures

arXiv:2606.30070 · q-fin.MF, math.PR · Submitted 2026-06-29 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Quantitative Finance. The day in quantitative finance: models of markets, risk and prices.

Ruth: I'm Ruth.

Warren: And I'm Warren, here with Iris.

Ruth: We're looking at the paper: "Financial Resilience Evaluation".

Warren: Here's the gist of the paper. Financial resilience concerns how a position recovers or deteriorates in response to adverse conditions, evaluated by a normalized, cash-additive dynamic risk measure rho applied to an Itô process pi.

Ruth: The core evaluation is defined as: D rho s pi t:= epsilon to 0+ 1 over epsilon rho s(pi t+ epsilon - pi t), 0 at most s at most t < T, where the limit exists.

Warren: When rho is a convex dynamic risk measure induced by a Backward Stochastic Differential Equation (BSDE) with a Lipschitz or quadratic driver, this limit is well-posed and has an explicit dual representation: D rho s pi t = ess mu in A 0 E mu

bt + g(t, sigma t) F s: , 0 at most s at most t < T.

Ruth: Theorem 15 establishes that under appropriate assumptions on the driver g and the inner process pi, this evaluation is attained: D rho s pi t = ess mu in A 0 E mu

bt + g(t, sigma t) F s: , for l1-a.e.

Paper summary: Ruth: So to wrap up on "Financial Resilience Evaluation: From Conditional Expectations to Dynamic Convex Risk Measures," the authors are showing how we can move past just looking at the expected recovery rate.

Warren: They do this by introducing these dynamic risk measures, which allow us to capture a much richer picture of the risk profile beyond just the conditional mean.

Iris: The main result is Theorem fifteen which confirms that under appropriate conditions for their driver g and inner process pi, the evaluation of this dynamic resilience rate is attained in a specific way involving an essential supremum LFRZ25 <ref:2606.30070#pg2>.

Ruth: What this means for the world is that we can use these tools to create better risk-adjusted forecasts for how resilient financial positions are over time, which helps us understand what happens when things go wrong.

Warren: It’s about using the structure of dynamic risk measures derived from backward stochastic differential equations to give us a more nuanced understanding of recovery in adverse conditions LFRZ25.

Iris: The paper provides a solid framework, but they're clear that this whole approach depends heavily on the assumptions we put on the driver g and how sensitive the inner process pi is to those drivers LFRZ25.

Ruth: It’s important to remember that if you’re using a risk measure that doesn't fit this setup, like one that isn't normalized or cash-additive, this evaluation might just fail to exist LFRZ25.

Conclusion: Ruth: So, basically, this paper takes that concept of measuring risk—that dynamic resilience—and shows you how you can actually calculate it rigorously using these backward stochastic differential equations.

Warren: Right. It’s moving past just looking at a single point in time and showing how that risk level changes as time goes on, based on the path the financial position takes.

Iris: What this means for us is that we have a way to get a more detailed snapshot of future recovery rates, not just an average number.

Ruth: Exactly. The authors are proving that if you set up your model correctly—if the driver and the inner process fit their criteria—you can get this precise mathematical expression for that risk rate at any moment in time.

Warren: But it’s not magic; there are some tight conditions they have to meet, like how smooth the driver function needs to be for that calculation to actually work out.

Iris: Yeah, and they lay out a lot of technical steps in their proof—like using Girsanov’s theorem and showing convergence in L one —to make sure this isn't just some theoretical idea but something we can actually use with real data.

Ruth: It gives us a formal language to talk about how fragile or robust a portfolio is during volatile periods, which should help people manage their exposure better.

Warren: If you can track that dynamic rate, you can spot when things are about to break down before the actual disaster hits.

Iris: But we need to keep in mind what they didn't cover; they focused on those specific conditions for the driver and it’s a lot of work to get there.

Ruth: True. So, next up, we’re going to look at some concrete examples—like how this works for conditional variance or entropic risk measures—to see how different risk tools actually behave in practice.

University of Amsterdam · University of Milano-Bicocca

q-fin.MF, math.PR

Submitted: 2026-06-29

Updated: 2026-11-14

Comments: 47 pages

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 91/100

The gist: Financial resilience concerns how a position recovers or deteriorates in response to adverse conditions, evaluated by a normalized, cash-additive dynamic risk measure rho applied to an Itô process pi.

Key concepts

Dynamic Risk Measure $\rho$
This is a way to quantify financial resilience by measuring how a position recovers or deteriorates under adverse conditions. It is evaluated using a normalized, cash-additive dynamic risk measure $\rho$ applied to an Itô process $\pi$. The core evaluation involves the limit of the risk measure applied to small future increments.
Dynamic Risk Measure Evaluation $D\rho_s\pi_t$
This is the central quantity derived from the paper, defined as a limit involving $\rho_s(\pi_{t+\varepsilon} - \pi_t)$ as $\varepsilon$ approaches zero. It represents the risk rate of a future increment and serves as a time-$s$ risk-adjusted forecast for the instantaneous resilience of $\pi$ at time $t$. It is related to the slope of the exposure-risk curve.
Dual Representation
When $\rho$ is induced by a Backward Stochastic Differential Equation (BSDE) with specific driver properties, $D\rho_s\pi_t$ has an explicit dual representation. This formula expresses the evaluation as an essential supremum over a set $A_0$, involving conditional expectations of terms like $bt + g(t, \sigma_t)$ given the filtration at time $s$.
Coherent Risk Measure
A coherent risk measure is a specific type of outer dynamic risk measure induced by a positively homogeneous driver. When $\rho$ is coherent, the evaluation simplifies to an explicit form: $D\rho_s\pi_t = \text{ess} \sup_{\mu \in A_0} t E^\mu [bt + \mu_t \cdot \sigma_t| F_s]$, which provides a more explicit representation of the resilience rate.

Terminology

Summary

Financial resilience concerns how a position recovers or deteriorates in response to adverse conditions, evaluated by a normalized, cash-additive dynamic risk measure rho applied to an Itô process pi. p. 0 The core evaluation is defined as:

D rho s pi t:= epsilon to 0+ 1 over epsilon rho s(pi t+ epsilon - pi t), 0 at most s at most t < T,

where the limit exists.

When rho is a convex dynamic risk measure induced by a Backward Stochastic Differential Equation (BSDE) with a Lipschitz or quadratic driver, this limit is well-posed and has an explicit dual representation:

D rho s pi t = ess mu in A 0 E mu [bt + g(t, sigma t) F s], 0 at most s at most t < T.

Theorem 15 establishes that under appropriate assumptions on the driver g and the inner process pi, this evaluation is attained:

D rho s pi t = ess mu in A 0 E mu [bt + g(t, sigma t) F s], for l1-a.e. t in [0, T) and all s in [0, t].

The interpretation of D rho s pi t is multifaceted: it represents the risk rate of a future increment, the slope of the exposure-risk curve, an extension of the representation formula for BSDE drivers, a first-order expansion at s=t, and serves as a time- s risk-adjusted forecast for the instantaneous resilience of pi at time t, based on zero-penalty scenarios encoded by A 0. p. 2

The proof of Theorem 15 is intricate, involving seven steps. p. 3 It utilizes Girsanov's theorem, asymptotic localization arguments based on condition (H s,t), and establishing upper and lower bounds for the essential supremum. The convergence to ess mu in A 0 E mu[Y tF s] in L 1 as epsilon to 0+ is shown. p. 3

Corollary 19 provides a more explicit representation when the outer dynamic risk measure rho is induced by a positively homogeneous driver g (a coherent risk measure):

D rho s pi t = ess mu in A 0 t E mu [bt + mu t times sigma t F s],

where A 0,t is defined based on the time-dependent zero-penalty set.

The paper provides examples, including conditional expectation (reducing to the mean resilience rate), conditional variance (isolating local volatility), and entropic risk measures. p. 2 Counterexamples are provided for risk measures not induced by BSDEs or those that are not normalized or cash-additive, demonstrating that the evaluation may fail to exist. p. 4

Relevant formulas include:

(1.1) D rho s pi t:= L 1 - epsilon to 0+ 1/epsilon rho s(pi t+ epsilon - pi t), 0 at most s at most t < T, whenever the limit exists. p. 0

(3.50) For a coherent outer risk measure: D rho s pi t = ess mu in A 0 E mu bt + g(t, sigma t F s). p. 4

The entropic case formula is given as: D e gamma s pi t = E[btF s] + gamma squared E sigma t squared F s. p. 0

Lemma 36 confirms that if the driver g is deterministic and constant in time, satisfying conditions (CL) or (CQ), then the condition (H s,t) is satisfied for every 0 at most s at most t < T. p. 5 The proof establishes this by showing that a related function psi = -f is concave and non-negative, leading to the inequality dist 2(q, d g(0)) at most phi(g*(q)) at most psi(g*(q)), where phi is defined based on the distance to the subgradient of g. The proof demonstrates that x to 0+ psi(x) = 0, confirming well-posedness in this deterministic setting. p. 1

In the main theorem's Lipschitz case, Step 4 involves replacing Lemma 20(ii) with Lemma 20(i) for K=L and a conjugate exponent v > 2. This yields an upper bound (D. p. 22): ess mu in B s,t epsilon V s,t epsilon (mu) at most ess mu in A s E mu[Y t epsilonF s] + C 1 q E Y t epsilon squared F s (epsilon), where (epsilon) to 0 as epsilon to 0+. p. 3 The lower bound (D. p. 33) is established similarly. p. 9 Step 6 shows that the common factor in both bounds converges in L 1 to H s,t, and Step 7 demonstrates that both the positive and negative parts of the difference converge to zero in L 1, leading to ess mu in B s,t epsilon V s,t epsilon (mu) - H s,t!+-to 0 in L 1. This convergence is then used to conclude the main result (3. p. 547) for t in the full l1-measure Borel set T = T 1 T epsilon 2. p. 4

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Improve resilience evaluation by replacing conditional expectation with dynamic risk measures to distinguish between positions having different conditional risk profiles beyond just their expected recovery rate. This allows for a richer characterization of financial resilience as noted in the abstract.

  2. Enable explicit calculation of the resilience evaluation for BSDE-induced convex risk measures through the derived dual representation formula D rho s pi t = ess max mu in A0 E mu [bt + g(t, sigma t) Fs]. This provides a closed-form analytical tool for assessing resilience under specific risk measure structures.

  3. Develop systems capable of identifying the optimal scenario for financial resilience by leveraging the attainment of the essential supremum in Theorem 15. This allows models to pinpoint the specific risk adjustment strategy that yields the maximum resilience value D rho s pi t.

  4. Implement algorithms that can compute conditional variance as a benchmark for resilience evaluation, isolating local volatility through V(dYtFt) = Zt squared dt. This capability helps in understanding the second-order contribution to financial risk dynamics.

  5. Create models that can analyze the behavior of Value at Risk and Expected Shortfall by demonstrating their failure to generate well-defined resilience evaluations outside the BSDE-induced framework, highlighting necessary structural assumptions like normalization and cash-additivity.

Sources

Related papers