Why Deep Learning Rarely Helps Estimate Implied Volatility

Summary

The discussion distinguishes estimating implied volatility from observed option prices from using machine learning to approximate prices for complex derivatives. Inverting a standard option pricing model to find implied volatility is described as a simple, well-understood numerical task, so deep learning offers little obvious advantage for that step. The greater opportunity is speeding up valuation when a product or risk calculation is computationally demanding.

A learned pricing model can shift work offline: expensive model-generated examples train it, and the trained model may provide faster online estimates. The trade-off is possible loss of accuracy. The approach also depends on having enough training data; complex derivatives often lack plentiful observable market prices, so training data may need to come from costly valuation models. Taylor approximations and precomputed price grids are cited as existing alternatives. The discussion is conceptual and does not provide empirical comparisons or specify when machine learning will outperform those methods.

Key ideas

• Implied volatility can generally be recovered from observed prices with established numerical methods.
• Machine learning may be more useful for approximating prices of computationally expensive derivatives.
• Offline training can trade greater preparation costs for faster online valuation.
• Synthetic training data may require running the costly pricing model the surrogate is meant to accelerate.
• Taylor approximations and precomputed price tables are alternative ways to speed up valuation.

Tags

# How can deep learning methods measure implied volatility?


# How can deep learning methods measure implied volatility?












Why and how should we utilize deep learning methods to calculate implied vol of options? I've also heard that finding the fair price of the option is not nearly as important as finding a numerical method to accurately measure IV, is this true? Is this a big point of research in Quant Fin?

## Answer by Kermittfrog (score 1)

https://quant.stackexchange.com/a/58795

As @Jesper Tidblom already stated in his comment, the quant finance problem is not in inverting observed prices to estimate the implied volatility; this is a well understood and, admittedly, simple problem these days.

Finding (model) prices for very complex derivatives products is a potential field for applied ML. Especially in counterparty and market risk applications, improvements of computation times are sought after. Using ML, you would be 'swapping' training time (offline) for online calculation speed, at the cost of potentially reduced accuracy.

The intricacy here is that in order to apply an ML model, you need data... And in terms of (complex) derivatives valuation, there is usually not sufficient observable data in the markets, e.g. from Bloomberg and such; thus you'd have to run your valuation model (costly) to train your ML model (costly) to have better intraday online performance. To add to that, there already exist known methods to arrive at approximate prices, e.g. Taylor expansions, (batch/overnight) pre-computation of price for various parameter combinations ...

Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.