Why Rough Volatility Models Must Be Asset-Specific for Accurate Pricing

Rough Volatility Across Assets

The Cross-Asset Result

The authors ran a common data pipeline through every market rather than stitching together separate estimations, and the conclusion repeats: log-volatility paths are not smooth. They do not approximate a continuous, differentiable process on any of these instruments.

This is not news to anyone running a vol-of-vol book. Roughness has been the working assumption in serious quantitative circles for years. What I find useful here is the cross-asset confirmation under a single framework. The roughness is not an artifact of one dataset, one window, or one estimation technique. It survives a unified pipeline. That matters because most retail options strategies still assume the opposite — and most retail vol-targeting backtests still price it.

What the Implied Surface Is Telling You

The more interesting half of the paper is the evaluation of implied roughness — the roughness you back out from option prices rather than from realized paths. This is where the practical edge sits for anyone running a vol-of-vol model, a surface arbitrage, or a delta-hedged book. The study finds that implied roughness estimates are reliable in some places and unreliable in others, and the differences are systematic, not random.

The practical implication is direct. If your edge is in fitting the vol surface, your roughness prior cannot be a global constant applied across every product. It needs to be asset-class-specific. A single roughness parameter plugged into every options book is a hidden assumption you pay for in drawdown. Cross-asset confirmation of roughness does not mean cross-asset uniformity of the roughness coefficient.

This also reframes vol targeting. If realized volatility is rough, the correct sampler for bootstrap, stress test, and Monte Carlo work is not geometric Brownian motion. Most retail "vol-targeting" backtests assume smooth paths and inflate Sharpe as a consequence. That is the kind of edge that evaporates the moment you cross a regime boundary.

Where the Signal Decays

The authors are appropriately tight-lipped about specific implied-roughness values, which is the right level of caution for this sample size. A unified pipeline standardizes estimation concerns; it does not eliminate them. The breadth makes the cross-asset claim robust, but per-asset-class implied-roughness numbers still carry noise.

What I will watch in replication: does the implied roughness estimate converge as you extend the surface into the wings, or does it diverge? Where it diverges, the surface is not telling you about roughness. It is telling you about liquidity, skew stickiness, or dealer hedging pressure. Those are different signals with different follow-throughs.

The retail takeaway, since someone always wants one: stop fitting vol surfaces under the assumption that volatility is smooth. It is not. The data across every liquid asset class now confirms it in a single framework, and there is no remaining excuse.