Arbitrage-Aware Multi-Step Forecasting of Implied Volatility Surfaces: Modelling Surface Trajectories Using Latent Diffusion
This paper proposes an arbitrage-aware conditional latent diffusion framework that generates realistic, economically admissible multi-step probabilistic forecasts of implied volatility surfaces and underlying returns, outperforming persistence benchmarks in point forecasting while effectively modeling complex surface geometry and temporal dependence.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Financial markets are vast, complex ecosystems where the price of an asset is only half the story; the other half is the uncertainty surrounding that price. When investors buy options, which are contracts giving them the right to buy or sell something later, they are essentially paying for a prediction of how much that price might swing. To make sense of these predictions, traders look at a "volatility surface." Imagine a three-dimensional map where every point represents how much uncertainty the market expects for a specific price target and a specific future date. This map is not flat; it has hills and valleys that shift constantly, reflecting the collective mood of the market. If this map is drawn incorrectly, it can lead to massive financial losses or allow for risk-free profits that should not exist, known as arbitrage. Because these maps are so central to pricing and managing risk, the ability to predict how they will change over time is a holy grail for finance. However, predicting the future shape of this map is incredibly difficult because it must follow strict economic rules while also accounting for the chaotic, unpredictable nature of human behavior and market news.
Researchers at the University of St.Gallen have developed a new way to forecast these shifting maps, treating the problem not as a simple guess but as a journey through time. Instead of trying to predict a single future point, their system generates a full thirty-day movie of how the volatility surface might evolve, day by day, alongside the movement of the stock market itself. They built a two-part machine learning system to achieve this. The first part acts like a translator, compressing the complex, three-dimensional map into a simpler, low-dimensional code that captures its essential shape without the noise. This translator is trained with a strict rule: it must never produce a map that allows for risk-free profits. By baking this economic rule directly into the translator, the system ensures that every future scenario it creates is financially valid. The second part of the system is a generative engine that takes this simplified code and imagines the next thirty days of movement. It learns from years of historical data to understand how the map usually changes, how it reacts to stock price drops, and how uncertainty tends to spread over time.
The results of this approach reveal a system that is remarkably good at capturing the rhythm of the market, though it has a specific blind spot. When the researchers tested their model on data from the S&P 500 index, they found that the generated maps looked very real. They preserved the familiar "smile" shape that options markets typically show, where uncertainty is higher for extreme price moves, and they correctly reproduced the way uncertainty changes as time passes. The model also successfully captured the deep connection between stock prices and volatility: when stocks fall, uncertainty usually rises, and the model learned to generate this negative relationship accurately over periods of a week or a month. In fact, for forecasts extending beyond a few days, the model's predictions were more accurate than simply assuming the market would stay exactly as it is today. This is a significant finding because financial markets are often so efficient that the best guess for tomorrow is usually just today's state.
However, the model is not a crystal ball for the immediate future. For the very next day, the system struggled to beat the simple assumption that nothing will change, often predicting too much movement where the market actually stayed still. Furthermore, while the model was excellent at predicting the direction and shape of future changes, it consistently underestimated how wild those changes would be. The range of possible outcomes it generated was too narrow; it captured the path the market might take but missed the full intensity of the storm. This suggests the model is better at understanding the structure of market movements than the sheer magnitude of their chaos. Despite this, the system produces a set of plausible, economically sound scenarios that can help investors visualize a range of futures rather than a single, rigid prediction.
The researchers also made a point to ensure their work could be checked and repeated by others, releasing the code and data protocols they used. This transparency is crucial in a field where models are often black boxes. By providing a clear path for others to test their ideas against the same data, they are helping to build a standard for how these complex problems should be solved. The study concludes that while this tool is not perfect for predicting the very next day's price, it is a powerful engine for generating realistic, multi-week stories about how market uncertainty might unfold. It offers a way to see the future not as a single line, but as a cloud of possibilities that respects the fundamental laws of economics, giving traders a better sense of the terrain they might have to navigate in the weeks ahead.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.