A Small Neural Network Calibrates VIX Options Fast

When markets lurch around, pricing VIX options can become a computational traffic jam. This paper tackles that bottleneck without discarding the mathematics behind the price. The authors keep the original pricing transform used for stochastic volatility models and split the formula into two parts: data-independent integrals and a market-dependent remainder. They precompute the integrals with GPU acceleration, then let a small neural network, a compact software model that learns patterns from data, handle only the remaining market-sensitive piece. The method is tested on a rough volatility model with tempered-stable jumps, designed for power-type volatility derivatives, and fitted to VIX options with a global-to-local search. The result is a pure-jump rough volatility model that captures VIX dynamics well, in line with prior empirical findings. The paper reports that this calibration framework achieves both high accuracy and speed.

VIX options are hard to price quickly. Each quote can send a calibration run into a long wait. That matters when markets jump and models must keep up. This paper keeps the original pricing transform, so the finance math stays intact. Then it splits the job into two parts. One part depends only on the model, not the live data. The other part depends on the market. A tiny neural network learns just that moving part. The surprise is simple. You do not need to replace the old pricing formula to speed it up.

What the model found in VIX options

The workflow lands on a rough volatility model with tempered-stable jumps. Rough volatility means the market’s shake follows a very jagged path. Tempered-stable jumps means the model allows sharp moves, but reins them in. That setup fits power-type volatility derivatives, including VIX options. The calibration uses a global-to-local search. That means it first hunts widely for good settings. Then it fine-tunes nearby. The result is plain. A pure-jump rough volatility model can capture VIX dynamics well. That matches earlier empirical findings. The method also gives high accuracy and speed.

How the speedup works

The method preserves the old pricing transform. That matters because the transform already encodes the model’s shape. First, it splits the pricing formula into data-independent integrals and a market-dependent remainder. Then it computes the data-independent integrals with GPU acceleration. GPUs are graphics chips that crunch many numbers at once. After that, a small neural network approximates only the remaining map from market inputs to prices. The network does less work than a full pricing engine. That design keeps the physics-like structure while shaving off wasted effort.

Why this matters for live pricing

Speed is not the only prize here. Robust calibration means the model can be fit without falling apart when the market shifts. That is the point of using a structure-preserving design. The framework keeps the finance formula recognizable, but it moves the slow part out of the way. This can help whenever a pricing desk needs many updates, not one perfect answer. It also shows that a pure-jump rough volatility view can be enough for VIX dynamics. That is a useful result for anyone who cares about volatility products, not just model builders.

The next test is broader market data

The sharp next check is whether the same split-and-learn trick holds beyond VIX options. The paper names stochastic volatility models with characteristic functions, so that is the natural home. It would also be worth testing the same global-to-local search on other volatility derivatives. The present result already gives one clear lesson. A smart shortcut can speed up calibration without throwing away the original pricing law. That makes the method feel less like a hack and more like a careful refit of the machine.