Daily Summary for 2026-11-01

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

Ruth: It's the first of November, twenty twenty-six, and this is the day's research.

Warren: 8 new papers came out today.

Ruth: I'm Ruth.

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

Ruth: We'll take the day in one pass, then pull out the papers we're staying with.

Warren: Start us off with Multiscaling in the Rough Bergomi Model: A Tale of Tails.

The day's research: Ruth: This work suggests we can discipline a policy with prior belief and update it based on realized utility instead of just predicting returns.

Warren: That incorporates uncertainty, which connects to how we model market dynamics.

Iris: Does that connection help us understand the multiscaling seen in rough volatility models?

Ruth: The paper Multiscaling in the Rough Bergomi Model: A Tale of Tails investigates if multiscaling is about temporal scaling or just fat-tailed distributions.

Warren: Understanding those complex dynamics informs pricing models, like those using Fourier-Laplace transforms in polynomial Ornstein-Uhlenbeck volatility models.

Iris: Those transforms develop numerical schemes for pricing options using Fourier and Laplace inversions on various volatility models.

Ruth: For market making, there is work on hedging impermanent loss of liquidity provision using static and dynamic approaches under Black-Scholes-Merton frameworks.

Warren: That links into automated market making, where forecasting implied volatility surface with generative diffusion models use diffusion models to generate arbitrage-free surfaces.

Iris: The paper Forecasting implied volatility surface with generative diffusion models introduces a diffusion model for one-day-ahead surfaces.

Ruth: A mathematical framework for modelling CLMM dynamics in continuous time rigorously analyzes Concentrated Liquidity Market Makers showing fees constrain admissible price processes.

Warren: That framework develops a rigorous model for concentrated liquidity market makers in decentralized finance using measure-valued processes.

Iris: Bayesian Parametric Portfolio Policies maps predictors into portfolio weights without modeling returns, making them a tool for factor timing.

Ruth: Indeed, that suggests incorporating uncertainty is key to using timing as a portfolio rule rather than relying solely on return forecasts.

Warren: The Risk-Neutral Equivalent Pricing of Model-Uncertainty moves asset pricing closer to practice by using a risk neutral equivalent approach to disentangle bias and risk.

Iris: That approach seems to disentangle the bias from the actual market risk more effectively.

Ruth: Unified Approach for Hedging Impermanent Loss of Liquidity Provision develops methods to hedge the impermanent loss experienced by liquidity providers on decentralized exchanges.

Warren: Those methods use both static and dynamic approaches, including model-based valuations under Black-Scholes-Merton frameworks.

Lucky paper: 2601.11305: Ruth: Let's look at Multiscaling in the Rough Bergomi Model: A Tale of Tails today. It’s about figuring out where this multiscaling behavior actually comes from in financial modeling.

Warren: So, the authors introduce a two-stage statistical testing procedure to isolate the source of this multiscaling behavior in the rBergomi model.

Iris: What exactly is Stage one doing when it compares the model against an uniscaling fractional Brownian motion process?

Ruth: In stage one, they use weighted least squares regression that accounts for heteroscedastic estimation errors across moments to establish if multiscaling is present at all.

Warren: And then in stage two, they apply shuffled surrogates that destroy temporal correlations to test the source of the scaling.

Iris: The null hypothesis for stage two is testing if multiscaling is purely distributional, meaning no temporal contribution from memory effects.

Ruth: The key finding here is that for roughness parameters H less than or equal to zero point zero one, the multiscaling comes primarily from fat-tailed return distributions, with kurtosis values exceeding fifty on average.

Warren: That suggests we should focus our risk management applications on those distributional properties rather than trying to exploit supposed fractal scaling laws in that model.

Iris: The paper also points out a regime transition around H equals zero point one where the nature of multiscaling shifts from being distributional to increasingly temporal as H increases.

Ruth: Even when the roughness parameter increases, say to H greater than zero point one, the multiscaling weakens in absolute terms, but volatility clustering emerges and intensifies instead of pure scaling.

Warren: We also see that for synthetic processes like the Multifractal Random Walk with a multifractal intensity lambda equal to zero point two five, genuine temporal multiscaling is confirmed by the testing framework.

Iris: That validation shows that the methodology successfully distinguishes between different sources of multiscaling in those synthetic cases, which is important because it confirms the two-stage approach works.

Ruth: The authors conclude that for risk management applications, we should focus on the model’s distributional properties instead of interpreting multiscaling as evidence of market inefficiency.

Warren: It's a step toward a more nuanced understanding because they make it clear that the rough Bergomi model still captures important features of implied volatility surfaces even with this finding.

Iris: And I think that makes sense because if we can separate the temporal scaling from the distributional effects, we get a clearer picture of what drives extreme risk estimation in those models.

Ruth: Exactly, so it's not about finding complex microstructure effects but understanding that heavy-tailed returns are the main driver here.

Lucky paper: 2511.07571: Ruth: Alright, we're looking at Forecasting implied volatility surface with generative diffusion models from Chen Jin and Ankush Agarwal today.

Warren: This paper introduces a conditional Denoising Diffusion Probabilistic Model, or DDPM, for generating one-day ahead arbitrage-free implied volatility surfaces.

Iris: So it’s not just fitting a static surface to historical data, but trying to capture path-dependent dynamics while suppressing arbitrage violations during training.

Ruth: Exactly. They frame the task as a conditional generation problem where the model learns the distribution of one-day ahead vol-surface conditioned on variables like exponentially weighted moving averages of historical vol-surfaces and asset returns.

Warren: And they enforce structural constraints using two penalty terms: a calendar arbitrage penalty, Φτ(σ), based on the non-decreasing total implied variance w(mi, τj) with respect to tau, and a butterfly arbitrage constraint, Φb(σ), ensuring convexity across moneyness m.

Iris: That means they are actively penalizing futures that violate basic financial validity during the generation process itself.

Ruth: The training signal uses a composite loss function that augments the standard score-matching objective with an explicit financial arbitrage penalty, weighted dynamically by the signal-to-noise ratio, or SNR.

Warren: Specifically, Equation (fourteen) shows the total loss is Ltotal(θ) = Lsimple(θ) + λ · Et,x(k+one)0,ϵh min(SNR(t), γ) · Φ(σˆ(k+one)).

Iris: It’s interesting how they cap the SNR at a predefined maximum threshold gamma to keep the training stable while still incorporating that penalty.

Ruth: The authors show superior performance over GAN-based approaches like VolGAN, achieving a Mean Absolute Error of zero point seven zero six two percent, compared to the benchmark's error of zero point nine one two zero percent.

Warren: And they also provide reliable uncertainty estimates; the ninety percent confidence intervals exhibit breach rates close to the theoretical ten percent target for ATM and ITM points, which is a good calibration result.

Iris: So even when adding that arbitrage penalty term, they don't compromise predictive accuracy when measured by MAPE.

Ruth: They do, actually, they show that including the arbitrage penalty term reduces all four statistics of the arbitrage penalties without compromising predictive accuracy as measured by MAPE.

Warren: When tested on real market data, the diffusion model achieves an overall MAPE of three point zero zero two six percent, which is lower than the benchmark's three point eight four one eight percent.

Iris: That lower error suggests that it’s statistically robust, and the Diebold-Mariano test confirms that the diffusion model reduces MAPE by approximately zero point eight four percentage points on average relative to VolGAN.

Ruth: The architecture itself uses a conditional U-Net with an encoder-decoder design and skip connections to reintegrate information, processing an augmented input tensor X(k) in R 4×Nm×Nτ.

Warren: They integrate time t and the conditioning vector c(k) by creating a joint context embedding e(k)ctx using a Multi-Layer Perceptron applied to sinusoidal embeddings of t and the conditioning vector c(k).

Iris: That embedding is then integrated into the convolutional blocks via Feature-wise Linear Modulation, or FiLM, which dynamically shapes the final output.

Ruth: The simulation study with synthetic data generated from the Heston model serves as a sanity check, validating that it can learn known stochastic dynamics in an arbitrage-free environment.

Warren: The conclusion is that by incorporating an explicit arbitrage penalty into the loss function that is weighted by SNR, they suppress violations without destabilizing training.

Iris: It sounds like a solid method for moving beyond simple fitting to something more constrained and financially valid in volatility forecasting.

Lucky paper: 2502.13744: Ruth: We're moving on to our third paper of this series, and today we're looking at The Risk-Neutral Equivalent Pricing of Model-Uncertainty from the University of York.

Warren: This paper is focusing on how to make asset pricing under model risk closer to practice by using a bottom-up perspective with a Risk Neutral Equivalent approach.

Iris: That sounds like it tackles the difficulty of separating bias and risk-pricing effects without having to deal with that usual joint hypothesis problem.

Ruth: Exactly, they introduce this RiskNeutral Equivalent formulation which aims to disentangle the ex-ante bias and risk-pricing effects using the fundamental theorem of asset pricing.

Warren: The authors show that viable model-risk pricing takes a canonical form featuring a price-of-risk and dynamic, which makes it more intuitive.

Iris: They set up the analysis by considering a binary model risk B with an unknown probability law pB0, and they focus on 'one-off' B-risks linked to salient events.

Ruth: If pB0 is unknown, they run into issues where the bias piB0 minus pB0 is greater than zero, and this one-off feature amplifies that effect while decoupling the pricing of B-risk from classical risk.

Warren: That decoupling leads to a simpler structure compared to existing approaches because it isolates the two effects more cleanly.

Iris: The viable asset-price dynamic is derived as a martingale measure equivalent to that of the expected benefit process under the reference law, which has three 'engines'.

Ruth: Those engines are specifically the B-sure risk-premium drift, the model-drift, and the RNE B-inference process.

Warren: Economically requiring a differentiable function, they find that viable pricing must have B-risk only pricing kPi t equal to (K one/two minus K minus one/two)sigma Pi t.

Iris: That expression introduces a conserved quantity K which scales up the plain reference gain-loss ratio, which is written as L pi plus zero over L pi plus zero.

Ruth: They establish that risk-pricing is positive if K is greater than one, and unpriced B-risk has K equal to one.

Warren: Under normal market competition, asset-pricing under B-risk satisfies RPMax equal to zero meaning K stays in the range of

one two). [Iris: The paper suggests that model-risks whose laws are unknown might reveal themselves through exploitable excess returns expressed in variables that are more readily observed and measured.

Ruth: Momentum trading exploits the momentum effect by ranking assets by excess-to-date and profiting from their differing average excess-to-come.

Warren: Volatility trading exhibits a Low-Risk effect, which inverts the classical dictum that high risk means high reward, and they give the peak volatility for volatility trading as v sigma max equal to one over rho plus one.

Iris: The analysis demonstrates how model-risks whose laws are unknown might show up through exploitable excess returns expressed in variables that are more readily observed and measured.

Ruth: So, the paper moves away from the joint hypothesis problem by showing how you can separate bias effect from ex-ante risk-pricing using this RNE formulation.

Lucky paper: 2407.05146: Ruth: So we're looking at this paper, Unified Approach for Hedging Impermanent Loss of Liquidity Provision, which covers how to hedge the impermanent loss of liquidity provision on Uniswap V2 and V3 protocols.

Warren: It defines impermanent loss in three ways: nominal for funded LP, nominal for borrowed LP, and relative IL for borrowed LP.

Iris: The paper then develops a static model-independent replication method using traded European vanilla call and put options to hedge the IL protection claim if you have a liquid options market.

Ruth: That's interesting because it shows that if you have that liquid options market, you can perfectly hedge the IL protection claim with option portfolios.

Iris: But when there isn't a liquid options market, they switch to a model-based approach decomposing the IL function into vanilla options, digital options, and an exotic square root payoff.

Warren: They use Fourier transform techniques applied to the moment generating function to compute the analytic valuation formulas for IL protection claims under both Black-Scholes-Merton and log-normal stochastic volatility models.

Ruth: The application to Uniswap V3 shows that the IL of funded LP can be decomposed into a linear part, an exotic square root price, and a vanilla option part.

Iris: So the payoff of the IL protection claim for funded LP under Uniswap V3 is given by Proposition three point six in that section.

Warren: The authors conclude that for a trading desk's total P andL, you look at the difference between premiums received from selling IL protection claims and the variance realized through delta-hedging.

Ruth: They suggest liquidity providers can focus on selecting DEX pools and liquidity ranges where expected fees could exceed the hedging costs.

Iris: The paper shows that the protection seller must hedge dynamically using option delta under a BSM model for that claim.

Ruth: That means a dynamic hedging strategy is needed, not just a static one, when you're dealing with those protocols.

Warren: The work builds on previous literature regarding automated market makers and loss-versus-rebalancing in decentralized exchanges.

Lucky paper: 2412.00658: Ruth: Alright, let's look at this paper, Fast Approximate Modular Inference for Option Price Prediction. The authors are Worapree Maneesoonthorn, David T. Frazier, and Gael M. Martin from Monash University.

Warren: This method uses multiple information sources: daily spot returns, high-frequency spot data, and option prices to predict future option prices without needing an explicit statistical model for those errors.

Iris: So they are using the Heston pricing model as a simulator while avoiding specifying models for the errors in the option prices themselves.

Ruth: Exactly, they use a modular Bayesian approach that cuts feedback between components to guard against misspecification of certain model parts.

Warren: That modularization means they don't need to specify full probabilistic models for high-frequency measures or option pricing errors.

Iris: How exactly does the training signal get built for this approximate posterior?

Ruth: They construct an approximate posterior using a Gaussian mixture model to approximate the cut posterior, pˆcut(ΦS(y1:T)), which they generate by simulating artificial data from the theoretical model.

Warren: And what about the computational efficiency gains they claim?

Iris: They say this method achieves predictive accuracy comparable to likelihood-based modular inference while delivering a four thousand-fold reduction in computational time compared to modular MCMC.

Ruth: That speed is significant, especially when you're looking at real-time updates of these predictive distributions for short-maturity contracts.

Warren: The paper shows this approach is most accurate in predicting options that mature within a trading day, with empirical coverage statistics hovering just under the nominal level.

Iris: Despite using the Heston model, which they note is almost certainly misspecified as a theoretical pricing model, the method still produces reasonable coverage for short-maturity contracts because of those continuous real-time updates.

Ruth: So it works even when the underlying pricing model isn't perfect, as long as you have fast enough posterior updates.

Warren: The modular ABI method relies only on summary statistics, and that modularization is not dictated by the rigid structure of the joint likelihood function.

Iris: That flexibility lets them use this framework with other theoretical models that incorporate additional stochastic features, provided they build informative summary statistics.

Ruth: It sounds like the speed of updating those cut posteriors is what makes this approach viable for practical derivative trading rules.