Stochastic Local Volatility Model

The stochastic local volatility (SLV) model marries two seemingly incompatible ideas: stochastic volatility (volatility is a random process) and local volatility (volatility depends on spot and time). The result is a framework that can fit the observed volatility surface exactly while producing forward dynamics that are more realistic than either parent model alone.

Why pure models fail

Pure local volatility (Dupire’s framework) lets volatility be any smooth function of spot price and time. It fits the surface perfectly: you calibrate it to option prices across all strikes and expirations, and by definition, you will price every liquid vanilla option correctly. The catch: the model implies that volatility moves are entirely determined by the spot price. If the stock rallies, the smile rotates and flattens as if by physics, not as a result of regime change. When you simulate the forward distribution, the paths are too concentrated; realized volatility in the tails is too small, and the smile evolution is too passive. Exotic option prices depend critically on path distribution, so you overprice exotics and underprice certain dynamic strategies.

Pure stochastic volatility (Heston, SABR, etc.) lets volatility follow a random process independent of spot. It is conceptually cleaner: volatility has its own shocks, its own mean reversion, its own term structure. But here is the problem: once you choose a parametric form (say, square-root process in Heston), you have only a handful of parameters. You can fit ATM volatility and the term structure well, but you cannot fit the smile shape across all strikes and expirations. You will always overprice some out-of-the-money options and underprice others, leaving money on the table and inviting smart traders to exploit your mispricing.

The reconciliation: leverage functions

The SLV model solves this by introducing a leverage function, usually denoted $L(S, V)$, that scales local volatility by the current level of the stochastic volatility process. The spot dynamics become:

$$dS_t = \mu S_t , dt + L(S_t, V_t) \sqrt{V_t} , S_t , dW^S_t$$

where $V_t$ is stochastic volatility (say, variance in Heston), and $W^S$ is the Brownian motion driving the spot.

The leverage function is typically normalized so that its expected value is one. This means: if volatility is at its mean level, the local-vol function is unchanged. But if volatility spikes, the leverage function amplifies or dampens the local-vol component, depending on spot. This coupling ensures the model can fit the entire surface while maintaining independent, stochastic volatility dynamics.

Calibration in practice

SLV calibration is a two-stage procedure. First, you fit a base stochastic volatility model (Heston or a similar structure) to ATM volatility, skew, and term structure. You keep this calibration fixed as your “volatility engine.” Second, you use the leverage function to absorb the residual smile: the difference between the prices the stochastic vol model generates and the market prices you observe.

The leverage function is fit by a least-squares minimization: you compute leverage values at a grid of spot levels and volatility levels such that option prices generated by the SLV model match the observed volatility surface as closely as possible. Modern implementations use finite difference or Monte Carlo grids, combined with optimization algorithms, to recover the leverage surface.

Once calibrated, the model is arbitrage-free: option prices are consistent with no static replication opportunities, and you can hedge spot and volatility risk simultaneously.

Path realism and forward variance

The payoff of SLV is in path dynamics. Simulation of the underlying and volatility produces a forward distribution that respects both the realized volatility clustering captured by the stochastic vol component and the strike-dependent tail behavior implied by the leverage function. Paths in the tails are fatter when volatility is high (good) and thinner when volatility is low (realistic). The smile itself evolves stochastically, reflecting both spot moves and volatility shocks.

This matters for variance swaps and volatility derivatives. A variance swap pays the difference between realized variance and a strike. Pure local vol dramatically underestimates the variance swap rate because it misses the independent volatility shocks. Pure Heston cannot fit the surface, so you start with a mispriced set of calibration tools. SLV does both: fits the surface and correctly prices the variance swap.

For barrier options, the path distribution is critical. A knockout option loses value if the spot touches a barrier; the probability of touching it depends on the full path, not just the endpoint. SLV gives you paths that respect smile dynamics and volatility clustering, so your barrier option prices are more reliable.

Practical variations

Mixing ratios vary. Some practitioners use a simpler form, where local volatility is scaled by a single scaling factor (not a full leverage function) that depends on volatility alone, not spot. This reduces calibration burden at the cost of less flexibility.

Markovian vs. non-Markovian. Pure SLV is Markovian: the future depends only on the current state (spot, volatility), not on the history. In practice, traders sometimes append a rough volatility component to capture memory effects and path dependence, producing a quasi-Markovian model that still allows fast Monte Carlo simulation.

Multi-asset extensions. For a basket of stocks (or a copula of correlated assets), you can define leverage functions for each asset, coupling local volatility to cross-sectional stochastic volatility models. This is common in structured products and correlation hedging desks.

Computational trade-offs

SLV is more expensive than pure local vol but cheaper than a full-featured Heston implementation with smile fitting. You need to:

1. Calibrate a base stochastic vol model (Heston in ~seconds).
2. Compute the leverage function on a grid (1–2 minutes on modern hardware).
3. Simulate paths or solve PDEs to price exotics (depends on contract complexity).

For a desk running hundreds of book positions, the leverage function is recomputed at market close or intraday as needed. If the ATM implied volatility moves 0.5 vols, you might recalibrate; if only the smile twists, a leverage-function refresh is enough.

Limitations and when to use alternatives

SLV is powerful but not universal. If your book is dominated by plain-vanilla options, pure local vol is cheaper and sufficient. If you need extreme scenario analysis—say, a 50% stock drop—SLV may struggle because the leverage function was fit in normal market conditions and extrapolates poorly into stress regimes.

For American options and early-exercise features, you need a pricing algorithm that can handle the leverage function in a PDE solver, not just in Monte Carlo. Some SLV implementations use tree methods, which are slower but can handle early exercise naturally.

For correlation and basket products, SLV works, but you must carefully specify the correlation between spot and volatility (a crucial parameter), and between assets. Misspecifying this correlation can undo all the smile-fitting benefit.

SLV’s relationship to rough volatility

Rough volatility models add non-Markovian memory to the volatility process, allowing path-dependent leverage. A rough-SLV hybrid (sometimes called rough-SLV) retains the leverage function but lets volatility have fractional Brownian motion dynamics with Hurst parameter less than 0.5. This produces even more realistic smile evolution and is an active research frontier, though calibration complexity rises sharply.

See also