Volatility Is Rough: the Discovery That Broke Fifty Years of Models

Sometimes a paper does not propose a better model. It goes and measures something, and the measurement turns out to be so far from what everybody assumed that a whole literature has to be rebuilt.

Jim Gatheral, Thibault Jaisson and Mathieu Rosenbaum asked a question that sounds almost too simple to be interesting: how jagged is volatility, really? They went to high-frequency data, measured it, and got an answer that nobody expected and that, once you follow the consequences, explains a stubborn puzzle that had defeated the field for twenty years.

The problem: every model assumes volatility is smoother than it is

Every stochastic volatility model, from Hull-White to Heston to SABR, assumes volatility is driven by a standard Brownian motion. That is a specific assumption about how rough the path is.

Roughness can be measured. The standard tool is the Hurst exponent, a number between zero and one that describes how jagged a random path is. It works like this:

• H = 0.5 is ordinary Brownian motion. Increments are independent. This is what every classical model assumes.
• H > 0.5 means the path is smoother than Brownian, with trending, persistent increments.
• H < 0.5 means the path is rougher than Brownian: increments tend to reverse, the path is more violently jagged, more crinkly at every scale.

Here is the interesting thing about the literature before 2014. The long-memory tradition in volatility modelling, going back to work on fractionally integrated volatility, had concluded that volatility is smoother than Brownian, with a Hurst exponent above 0.5. That was the consensus: volatility has long memory, it trends, it is persistent.

Gatheral, Jaisson and Rosenbaum went and measured it directly from high-frequency realised volatility, and found the opposite.

The key idea via analogy: a coastline, not a hillside

Their central empirical result is stark. Log-volatility behaves essentially like a fractional Brownian motion with a Hurst exponent of about 0.1.

Not 0.5. Not 0.6. About 0.1. That is dramatically rougher than anything anyone was modelling.

To get an intuition for what H = 0.1 looks like: standard Brownian motion is already famously jagged, nowhere smooth, wiggling at every scale. A path with H = 0.1 is far worse. Zoom in on any piece of it and you find more violent crinkle, not less. It is less like a hillside and more like a fractal coastline: rough at every level of magnification, and dramatically so.

And they found this across essentially every asset they looked at, and at every reasonable timescale. It is not a quirk of one market. It appears to be a universal feature of financial volatility, which is itself a remarkable claim.

Why the roughness explains the smile

Here comes the payoff, and it resolves the oldest embarrassment in stochastic volatility modelling.

Every classical stochastic volatility model has the same fatal flaw: it cannot produce a steep enough smile for short-dated options. Look at options expiring in a week, and the real market shows a violently steep skew. Heston, calibrated to fit it, requires a volatility-of-volatility so absurdly high that it destroys the fit everywhere else and implies dynamics nobody believes.

The structural reason is simple. In a classical model, volatility is driven by a Brownian motion, which over a very short interval barely moves. So over one week, volatility is nearly constant, returns are nearly normal, and the smile is nearly flat. The model cannot generate a steep short-dated skew, no matter how you tune it. To patch this, the field bolted on jumps (Bates, Kou, Variance Gamma). That works, but it always felt like a workaround.

Rough volatility solves it without jumps. A path with H = 0.1 moves an enormous amount over a very short interval. Precisely because it is so jagged, volatility can shift substantially in a day, and the fat tails appear immediately rather than accumulating over months. So a rough volatility model produces a steep short-dated skew naturally, as a direct consequence of the measured roughness, with no jumps and no implausible parameters.

Even better, the model predicts the shape of how the skew decays with maturity, and the prediction matches the observed market with striking accuracy, using very few parameters. The rough Bergomi and rough Heston models that followed fit the entire volatility surface with roughly three parameters, where classical models struggle with five or more.

That is a rare and powerful kind of result. An empirical fact measured from time-series data (how rough realised volatility is) turned out to explain a completely separate puzzle in cross-sectional pricing data (why the short-dated smile is so steep). Two unrelated bodies of evidence pointing to the same number is the sort of thing that makes a claim believable.

Why it mattered

• It is a genuine anomaly that turned into a new field. Within a few years of the first working paper, rough volatility went from an obscure observation to one of the most active research areas in quantitative finance, with rough Heston, rough Bergomi, and a large body of work on its numerics.
• It gave the steep short-dated skew a cause. For twenty years the answer was "jumps." Rough volatility offers a different, arguably more parsimonious answer: volatility is simply far more jagged than we thought.
• It fits the surface with fewer parameters. Rough models achieve better fits with three parameters than classical models manage with five or more, and the parameters are far more stable across time, which is a serious practical advantage: unstable calibration means unstable hedges.
• It has a plausible microstructural story. Rosenbaum and co-authors showed that rough volatility emerges naturally from models of order flow in which trades beget trades: the self-exciting, clustered nature of high-frequency order arrival produces exactly this kind of roughness at longer timescales. The roughness is not a mathematical curiosity, it may be the macroscopic fingerprint of how markets actually trade.
• It challenged a consensus with data. The prevailing view was that volatility is smooth and persistent. The paper says the opposite, and backs it with measurement.

The honest limitations

• The measurement itself is contested. This is the most important caveat. Estimating the Hurst exponent from realised volatility is delicate, because realised volatility is itself an estimate of an unobservable quantity, computed from noisy high-frequency prices contaminated by microstructure effects such as the bid-ask bounce. Several subsequent papers have argued that this estimation noise biases the Hurst exponent downwards, and that the true value may be considerably higher than 0.1, potentially even consistent with a smoother process. This debate is genuinely unresolved. The pricing implications of roughness are widely accepted; the precise value of H, and whether volatility is "really" rough at all, is still argued about.
• The mathematics is punishing. Fractional Brownian motion with H below 0.5 is not a semimartingale and not Markovian. Almost every tool of standard mathematical finance, Ito's lemma, the Feynman-Kac equation, standard PDE methods, tree methods, assumes those properties. Rough volatility throws away the entire toolkit. Even the arbitrage theory has to be handled carefully to ensure the model does not accidentally permit arbitrage.
• It is slow. Because the process has no Markov property, simulation requires you to carry the whole history at every step, which is expensive. Considerable ingenuity has gone into making rough models fast enough for production use, and it remains a real obstacle. The rough Heston model, which does admit a semi-analytic characteristic function, is popular partly because it is one of the few tractable options.
• It describes, but does not fully explain. The microstructural story is suggestive and elegant, but the claim that order-flow self-excitation is the cause of rough volatility is still an active area of argument rather than a settled fact.
• Better fit is not the same as better hedging. A model that fits the surface superbly may still produce poor hedge ratios. The evidence on whether rough models actually hedge better in practice is thinner than the evidence that they fit better, and hedging is what a desk lives on.

The one-line takeaway

Gatheral, Jaisson and Rosenbaum measured how jagged volatility actually is, found a Hurst exponent around 0.1, far rougher than the Brownian motion assumed by every model since 1987, and discovered that this single empirical fact naturally explains the brutally steep short-dated smile that had defeated classical stochastic volatility for two decades, spawning an entire new modelling literature at the cost of abandoning most of the mathematics that made the old one tractable.