Daily Summary for 2026-10-20

daily

Transcript

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

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

Warren: 2 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 Arbitrage-Aware Probabilistic Multi-Step Forecasting of Implied Volatility Surfaces.

The day's research: Ruth: We focused today on improving forecasts for implied volatility surfaces.

Warren: Those surfaces are critical for numerous financial applications; what's the core issue?

Iris: Accurately modeling their changes demands handling two-dimensional geometry and time dependence.

Ruth: Uncertainty is another factor we have to incorporate into the model structure.

Warren: Ensuring the resulting models remain economically sound is a major challenge.

Iris: Soundness depends on how we weigh the geometric and temporal inputs, doesn't it?

Ruth: DYSANOS presents the first generative market model for smooth SANOS option surfaces.

Warren: It aims to be free of static arbitrage across those surfaces.

Iris: The design generates entire paths of daily spot and option prices covering years into the future.

Ruth: This uses a baseline AR1 model to generate hidden states simply.

Warren: How does this compare against pure implied volatility PCA models?

Iris: The research compares DYSANOS performance on IvyDB S&P Index data from two thousand twenty to two thousand twenty-five.

Ruth: That comparison covers the period of two thousand twenty to two thousand twenty-five, correct?

Warren: Yes, the study uses that specific timeframe for its benchmark analysis.

Iris: The comparison is against pure implied volatility PCA models for performance metrics.

Iris: The main idea is combining surface learning with diffusion for evolution, creating more realistic probabilistic scenarios.

Ruth: That generative modeling works alongside DYSANOS's focus on smooth, arbitrage-free paths for option surfaces.

Warren: What does that combination specifically allow us to generate that isn't currently available?

Iris: It creates more realistic probabilistic scenarios than what exists now.

Ruth: So the latent diffusion processes help evolve those surface representations into those paths?

Warren: Can you detail how this differs from standard simulation methods for option surfaces?

Iris: It builds future paths for implied volatility surfaces and stock returns while avoiding arbitrage.

Ruth: That avoidance of arbitrage is key when generating these trajectories.

Warren: And the DYSANOS work focuses on creating smooth, arbitrage-free option surfaces across many years?

Iris: Yes, it introduces a generative model for smooth surfaces over all strikes and expirations.

Ruth: So we have surface evolution via diffusion and smooth generation via DYSANOS.

Warren: What is the concrete benefit of using latent diffusion in this context?

Iris: It creates realistic future paths by modeling how those surfaces evolve dynamically.

Ruth: That dynamic modeling provides a richer set of scenarios for testing theories.

Warren: How does the forecasting aspect improve upon static surface fitting methods?

Iris: It allows for multi-step forecasting while maintaining arbitrage avoidance in the process.

Ruth: So we are moving toward joint generation of surface-return trajectories using latent diffusion.

Warren: And DYSANOS generates smooth, arbitrage-free non-parametric option surfaces for all strikes and expirations.

Iris: That paper is worth hearing because it introduces a model creating smooth, arbitrage-free surfaces across many years.

Ruth: We also have Arbitrage-Aware Probabilistic Multi-Step Forecasting of Implied Volatility Surfaces: Joint Generation of Surface-Return Trajectories Using Latent Diffusion, which uses latent diffusion to create realistic future paths for implied volatility surfaces and stock returns while avoiding arbitrage.

Warren: That paper is valuable because it uses a latent diffusion model to create realistic future paths for implied volatility surfaces and stock returns while avoiding arbitrage.

Iris: Those are the papers worth hearing today.

Ruth: Arbitrage-Aware Probabilistic Multi-Step Forecasting of Implied Volatility Surfaces: Joint Generation of Surface-Return Trajectories Using Latent Diffusion is worth hearing because it uses a latent diffusion model to create realistic future paths for implied volatility surfaces and stock returns while avoiding arbitrage.

Warren: DYSANOS Generative Dynamic Smooth Arbitrage-free Non-parametric Option Surfaces is worth hearing because it introduces a generative model that creates smooth, arbitrage-free option surfaces for all possible strikes and expiration dates over many years.

Iris: Agreed.

Ruth: That covers what we did today with those findings.

Warren: Indeed.

Iris: I conclude the review now.

Lucky paper: 2608.22478: Ruth: So, let's get into some specifics on this paper now. We're talking about the Arbitrage-Aware Probabilistic Multi-Step Forecasting of Implied Volatility Surfaces: Joint Generation of Surface-Return Trajectories Using Latent Diffusion.

Warren: It uses a twostage latent diffusion model to generate joint thirty-day trajectories for implied volatility surfaces and underlying returns while keeping it economically admissible.

Iris: The autoencoder compresses the daily surfaces into low-dimensional codes, and then the conditional diffusion model captures that evolution based on recent history.

Ruth: That separation assigns representation learning and economic regularization to the autoencoder, with the conditional law over future paths handled by the diffusion model.

Warren: Before we get to how they built their data, I want to ask about their data construction protocol.

Iris: The protocol involves constructing daily implied volatility surfaces from OptionMetrics data on WRDS covering January two thousand to August two thousand twenty-five.

Ruth: They use a pretrained graph neural operator maps the irregular BSM IV quote set to a smooth, essentially static-arbitrage-free surface on prescribed coordinates.

Warren: So they are transforming those daily t-IVS into a normalized log surface before compressing it into an eight-dimensional latent state using a deterministic RAE-inspired autoencoder.

Iris: The training objective for the autoencoder includes reconstruction terms and latent regularization terms to guide the model towards admissible surfaces.

Ruth: When they generate the next thirty days of surfaces, they use anchored displacements where persistence of the latest observed state corresponds to future anchored displacements remaining at their value at that forecast origin.

Warren: That means a mean-scaling gate is fitted to correct the predictive mean by adding a term related to persistence, specifically zˆ scaled h equals zˆ persistence plus lambda b h (zˆ mean h minus zˆ persistence).

Iris: The evaluation on SPX surfaces showed that the framework outperforms the persistence benchmark in point forecasting, even when generating realistic probabilistic multi-step scenarios.

Ruth: They said the generated trajectories preserve the dominant smile and term-structure geometry, recovering the realized level, skew, and curvature factors of surface movements.

Warren: What about their limitations? I see they mention underdispersion as a key issue.

Iris: The principal limitation is that predictive intervals widen with the horizon but remain too narrow because the model captures the directions and dependence of future movements more accurately than their amplitude.

Ruth: They also noted that while underdispersion is a problem, the predictive mean contains information not available from persistence alone.

Warren: So they are showing gains over persistence starting from approximately weekly horizons, and those gains are strongest at short maturities and in the wings.

Iris: The CRPS score shows that diffusion outperforms bootstrap persistence in eight of nine surface descriptors by horizon ten indicating better distributional improvements at medium and longer horizons.

Ruth: That means the predictive mean's gains are driven by horizondependent shape adjustments rather than uniform level shifts.

Warren: And what about the arbitrage aspect? They evaluated restrictions using finite differences on the grid.

Iris: The results showed virtually no calendar violations and only sparse butterfly violations in terms of arbitrage restrictions.

Ruth: But they also found that the return channel matches the zero-return baseline, producing informative distributions without implying any return predictability.

Warren: So, to summarize what this paper does for someone just listening, it moves beyond just copying the last surface forward and generates probabilistic scenarios that respect market structure over longer time steps.

Iris: Exactly. The main limitation is that the predictive intervals are too narrow, even though the mean prediction itself is more informative than simple persistence.

Ruth: The paper provides a reproducible protocol for constructing these datasets from OptionMetrics data on WRDS, which is a big piece of infrastructure for anyone trying to benchmark IVS forecasting.

Warren: And they also provided code on GitHub so other researchers can try to replicate the setup and see if it holds up.

Iris: The authors suggest that future work should focus on identifying which metrics are most appropriate for evaluating these functional trajectories in practice.

More episodes

← Home