project

Deep Hedging under Rough Volatility

resultResearch prototype with repo; no public performance number claimed.

A rough volatility hedging prototype comparing smooth GBM intuition with jagged volatility paths and learned path dependent hedges.

status: research prototypeartifact: GitHub repoconfidence: medium

Problem

Standard Brownian volatility paths are often too smooth for the hedging behavior I wanted to study. The question was whether rougher, path dependent volatility dynamics change what a learned hedge has to pay attention to.

Approach

Generated fractional Gaussian noise with the Davies Harte algorithm to study Hurst values around 0.1, then used neural SDE style dynamics and reinforcement learning to test hedging policies under non Markovian volatility.

The useful engineering work was the simulation stack: path generation, policy training, risk measurement, and comparison against simpler smooth volatility baselines.

What I tried that failed

The early write up leaned too hard on a single headline risk metric. I am keeping the project, but not treating that metric as a final public claim until the evaluation is rerun under the new notebook standard.

Naive smooth volatility paths made the agent look cleaner than the market structure warranted, which is exactly the kind of result I now try to audit before advertising.

Finding

Rough paths change the hedging problem in a way smooth volatility does not expose. The current value is methodological: path generation, learning loop, and a concrete reminder that one headline risk metric is not enough.

How to reproduce

The public repo contains the implementation. The next cleanup step is a locked seed suite and fee aware benchmark before I put a performance number back on the site.