Interface Volatility in SV Models

• Interface volatility is defined as the mutual information between observed returns and latent volatility, framing the inferential limits in SV models.
• The methodology applies Euler discretization and Shannon mutual information to quantify key effects like persistence and leverage in volatility inference.
• Findings reveal that daily returns provide modest information, suggesting that high-frequency data or additional observables are needed for precise volatility recovery.

Searching arXiv for the specified paper and closely related material to ground the article. “Interface volatility” (Editor's term) can be used to denote the information-theoretic interface between observed returns and latent volatility in stochastic volatility (SV) models. In the formulation studied in “Volatility Inference and Return Dependencies in Stochastic Volatility Models” (Pfante et al., 2016), stock returns $r_t$ are driven by an unobserved process $v_t$ capturing the random dynamics of volatility, and the central question is how much information about $v_t$ and future returns can be inferred from past returns in terms of Shannon mutual information. The resulting framework treats volatility inference as a hidden-state problem, identifies volatility persistence and leverage as the main sources of inferability, and shows that predictability of returns is bottlenecked by the information carried by the latent volatility. Across the model classes and parameterizations examined, the mutual information available from daily returns is modest, and the information required for precise volatility recovery is substantially larger (Pfante et al., 2016).

1. Stochastic-volatility model classes

The continuous-time SV diffusion model is specified by

$$\begin{aligned} dS_t &= \mu(v_t)\,S_t\,dt + f(v_t)\,S_t\,dW^0_t,\ dv_t &= b(v_t)\,dt + \gamma(v_t)\,dW_t, \end{aligned}$$

where $W^0_t$ is a 1D Brownian motion for returns noise, and $W_t=(W^1_t,\ldots,W^n_t)$ is $n$-dimensional Brownian motion for volatility; instantaneous leverage is captured via $$\mathrm{d}W^0_t\,\mathrm{d}W^i_t=\rho_i\,\mathrm{d}t$$. Smooth coefficients and uniform ellipticity are assumed for well-posedness, and the volatility process admits a unique stationary distribution $\rho$ (Pfante et al., 2016).

For returns sampled at interval $\tau$, the Euler discretization is

$v_t$0

with $v_t$1, $v_t$2 jointly Gaussian with correlation vector $v_t$3.

A mean-reverting one-factor family is given by

$v_t$4

with cases $v_t$5. The parameters are $v_t$6 for the risk-free rate, $v_t$7 for the mean-reversion level, $v_t$8 for volatility-of-volatility, $v_t$9 for the scaling exponent on diffusion, $v_t$0 for the exponent in drift, and $v_t$1 for the leverage coefficient.

The exponential Ornstein–Uhlenbeck (Exp-OU) one-factor model is

$v_t$2

with stationary $v_t$3. A two-factor Exp-OU variant uses two OU factors $v_t$4 and $v_t$5, and empirical work typically takes $v_t$6 independent. Later analysis uses the Gaussian volatility proxy $v_t$7.

These model classes establish the common setting in which returns are directly observed while volatility is latent. A plausible implication is that any inference procedure, whether parametric or nonparametric, inherits the same structural dependence on persistence, leverage, and observation noise encoded by $v_t$8 and $v_t$9.

2. Mutual-information formulation of inferability

The mutual information about current volatility from past returns is defined as

$$\begin{aligned} dS_t &= \mu(v_t)\,S_t\,dt + f(v_t)\,S_t\,dW^0_t,\ dv_t &= b(v_t)\,dt + \gamma(v_t)\,dW_t, \end{aligned}$$0

and the mutual information about the next return from the past is

$$\begin{aligned} dS_t &= \mu(v_t)\,S_t\,dt + f(v_t)\,S_t\,dW^0_t,\ dv_t &= b(v_t)\,dt + \gamma(v_t)\,dW_t, \end{aligned}$$1

Under the Euler scheme, the return-volatility system is a hidden Markov model: $$\begin{aligned} dS_t &= \mu(v_t)\,S_t\,dt + f(v_t)\,S_t\,dW^0_t,\ dv_t &= b(v_t)\,dt + \gamma(v_t)\,dW_t, \end{aligned}$$2 This allows the use of the data processing inequality and the chain rule to obtain tight upper bounds in terms of simpler mutual informations involving the latent volatility (Pfante et al., 2016).

The key proposition is

$$\begin{aligned} dS_t &= \mu(v_t)\,S_t\,dt + f(v_t)\,S_t\,dW^0_t,\ dv_t &= b(v_t)\,dt + \gamma(v_t)\,dW_t, \end{aligned}$$3

and

$$\begin{aligned} dS_t &= \mu(v_t)\,S_t\,dt + f(v_t)\,S_t\,dW^0_t,\ dv_t &= b(v_t)\,dt + \gamma(v_t)\,dW_t, \end{aligned}$$4

These bounds isolate two sources of inferability. The term $$\begin{aligned} dS_t &= \mu(v_t)\,S_t\,dt + f(v_t)\,S_t\,dW^0_t,\ dv_t &= b(v_t)\,dt + \gamma(v_t)\,dW_t, \end{aligned}$$5 measures persistence of volatility, while $$\begin{aligned} dS_t &= \mu(v_t)\,S_t\,dt + f(v_t)\,S_t\,dW^0_t,\ dv_t &= b(v_t)\,dt + \gamma(v_t)\,dW_t, \end{aligned}$$6 measures the leverage-mediated information in the contemporaneous return. Return predictability is correspondingly bottlenecked by $$\begin{aligned} dS_t &= \mu(v_t)\,S_t\,dt + f(v_t)\,S_t\,dW^0_t,\ dv_t &= b(v_t)\,dt + \gamma(v_t)\,dW_t, \end{aligned}$$7. In this formulation, dependence among returns is not treated as primitive; it is induced by the latent volatility process and then compressed by the observation channel from $$\begin{aligned} dS_t &= \mu(v_t)\,S_t\,dt + f(v_t)\,S_t\,dW^0_t,\ dv_t &= b(v_t)\,dt + \gamma(v_t)\,dW_t, \end{aligned}$$8 to $$\begin{aligned} dS_t &= \mu(v_t)\,S_t\,dt + f(v_t)\,S_t\,dW^0_t,\ dv_t &= b(v_t)\,dt + \gamma(v_t)\,dW_t, \end{aligned}$$9.

3. Analytical bounds and discrete-time analogues

For jointly Gaussian variables $W^0_t$0, the mutual information is

$W^0_t$1

This identity yields closed-form expressions in Gaussian-linear special cases. For log-OU volatility,

$W^0_t$2

because the OU transition is Gaussian with variance $W^0_t$3 (Pfante et al., 2016).

For a general diffusion, Theorem 3.1 gives the Euler-proxy expansion

$W^0_t$4

where $W^0_t$5 is the entropy of the stationary volatility. The same theorem gives the leverage-driven information gain

$W^0_t$6

In one-factor models,

$W^0_t$7

Theorem 3.2 provides a Fisher-information-based upper bound via a logarithmic Sobolev inequality (LSI). If the stationary $W^0_t$8 satisfies LSI with constant $W^0_t$9,

$W_t=(W^1_t,\ldots,W^n_t)$0

where

$W_t=(W^1_t,\ldots,W^n_t)$1

If $W_t=(W^1_t,\ldots,W^n_t)$2, then

$W_t=(W^1_t,\ldots,W^n_t)$3

so the bound becomes independent of sampling interval $W_t=(W^1_t,\ldots,W^n_t)$4 whenever $W_t=(W^1_t,\ldots,W^n_t)$5 is constant. If LSI holds with $W_t=(W^1_t,\ldots,W^n_t)$6, the Kullback–Leibler divergence to stationarity decays as $W_t=(W^1_t,\ldots,W^n_t)$7, implying that volatility persistence information decays exponentially in time.

A discrete-time SV analogue is the Gaussian log-vol AR(1) model,

$W_t=(W^1_t,\ldots,W^n_t)$8

possibly with leverage via $W_t=(W^1_t,\ldots,W^n_t)$9. Its persistence is

$n$0

and its instantaneous leverage contribution is

$n$1

Taken together, these formulas show that the inferability of latent volatility is governed by a small set of structural quantities: temporal persistence, the leverage coefficient, and the geometry of the stationary law through LSI and Fisher information.

4. Return dependence, squared-return autocorrelation, and leverage

In SV models, $n$2 are nearly uncorrelated, but $n$3 show positive autocorrelation driven by volatility persistence. For lognormal volatility with

$n$4

the squared-return autocorrelation is approximately

$n$5

For OU volatility,

$n$6

which yields exponential decay in $n$7. In two-factor Exp-OU models, the covariance is a sum of exponentials, producing short- and long-memory components (Pfante et al., 2016).

The leverage effect appears as instantaneous negative correlation between $n$8 and $n$9, or between $$\mathrm{d}W^0_t\,\mathrm{d}W^i_t=\rho_i\,\mathrm{d}t$$0 and $$\mathrm{d}W^0_t\,\mathrm{d}W^i_t=\rho_i\,\mathrm{d}t$$1, when $$\mathrm{d}W^0_t\,\mathrm{d}W^i_t=\rho_i\,\mathrm{d}t$$2. Information-theoretically, leverage increases $$\mathrm{d}W^0_t\,\mathrm{d}W^i_t=\rho_i\,\mathrm{d}t$$3 through

$$\mathrm{d}W^0_t\,\mathrm{d}W^i_t=\rho_i\,\mathrm{d}t$$4

Stronger leverage therefore means more extractable information about current volatility from the contemporaneous return, but only modest gains for realistic $$\mathrm{d}W^0_t\,\mathrm{d}W^i_t=\rho_i\,\mathrm{d}t$$5.

A common interpretive error is to equate strong squared-return autocorrelation with high-precision volatility recovery. The reported results indicate a different conclusion: autocorrelation of $$\mathrm{d}W^0_t\,\mathrm{d}W^i_t=\rho_i\,\mathrm{d}t$$6 is a manifestation of volatility persistence and does increase inferability, yet even sizable autocorrelation translates into modest mutual information under realistic parameters.

5. Empirical parameterizations and reported magnitudes

The paper reports numerical and analytical results for one-factor mean-reverting models, one-factor Exp-OU, and two-factor Exp-OU under empirical parameterizations (Pfante et al., 2016). In the one-factor mean-reverting family with annual parameterization and daily sampling $$\mathrm{d}W^0_t\,\mathrm{d}W^i_t=\rho_i\,\mathrm{d}t$$7, the upper bound

$$\mathrm{d}W^0_t\,\mathrm{d}W^i_t=\rho_i\,\mathrm{d}t$$8

is approximately $$\mathrm{d}W^0_t\,\mathrm{d}W^i_t=\rho_i\,\mathrm{d}t$$9 to $\rho$0 nats, depending on $\rho$1 and fitted parameters. The leverage term contributes $\rho$2; for $\rho$3, this is approximately $\rho$4–$\rho$5 nats. The return-predictability bound

$\rho$6

is approximately $\rho$7–$\rho$8 nats. The information required to quote daily returns to $\rho$9 precision is

$\tau$0

and for an annualized volatility index,

$\tau$1

The required sampling frequency to bridge the volatility information gap with the bound $\tau$2 is on the order of $\tau$3–$\tau$4 million returns per year, that is, “secondly” resolution or finer. For the Heston case $\tau$5, the stationary solution may fail for empirical parameter values, preventing meaningful MI bounds in those cases; the paper proves non-existence in one subcase.

For one-factor Exp-OU with daily parameters from Perelló et al. (2008), volatility persistence is exactly

$\tau$6

the LSI/Fisher bound on return predictability is

$\tau$7

and the reported numerical values are

$\tau$8

The required information is

$\tau$9

and at least $v_t$00 returns per day are needed to reach volatility precision via returns alone.

For two-factor Exp-OU with parameters from Alizadeh et al. (2002), the volatility proxy $v_t$01 is Gaussian but not Markov, and the data processing inequality gives

$v_t$02

The information increases with history length $v_t$03 but saturates quickly, with an approximate maximum of $v_t$04 nats after $v_t$05 days. The reported numerical values are

$v_t$06

Despite two time scales, inferability is lower than in the one-factor case because the transient component reduces temporal dependence.

The same results can be expressed in bits: $v_t$07 nat $v_t$08 bits, so $v_t$09 nats is approximately $v_t$10 bits, while $v_t$11 nats is approximately $v_t$12 bits. The gap between the information required for precise volatility recovery and the information delivered by daily returns is therefore large.

6. Estimation perspective, limitations, and interpretive consequences

State-space filtering and smoothing methods, including particle filters and extended or unscented Kalman procedures for transformed or approximated models, as well as MLE and Bayesian estimation for parameters and latent volatility, exploit the hidden Markov structure of SV models. The paper’s mutual-information bounds do not replace these algorithms; instead, they quantify a fundamental limit on inference accuracy from return data alone (Pfante et al., 2016).

Several consequences follow directly from the reported bounds. High persistence, such as $v_t$13 in the AR(1) model or small $v_t$14 in OU dynamics, increases $v_t$15, but the mutual information remains modest once realistic volatility-of-volatility $v_t$16 is taken into account. Larger $v_t$17 increases measurement noise in volatility estimation via returns. Predicting volatility benefits from higher sampling frequency, but the increase is practically useful only at very high frequency. Predicting returns is more constrained: under broad conditions, including constant drift or drift as a function of $v_t$18, the bound $v_t$19 is independent of the sampling interval $v_t$20, so shorter $v_t$21 does not fundamentally increase predictability. Past data length also has limited impact, because most usable information accrues quickly and then stabilizes at low levels.

The analysis is subject to explicit assumptions. It uses normal innovations and parametric SV structures; stationarity is crucial for many results, and some Heston parameter sets may violate it. The Euler discretization approximates continuous-time dynamics, so accuracy depends on $v_t$22. LSI-based bounds require convexity, specifically a uniform lower bound on the Hessian of $v_t$23, and may be loose for heavy-tailed or weakly convex stationary laws. Model misspecification can alter dependencies; jumps and rough volatility are identified as cases outside the diffusive SV focus.

These qualifications also delimit several common misunderstandings. The reported results do not imply that latent-volatility filtering is ineffective; rather, they imply that returns alone impose a strict information ceiling. They do not imply that more factors automatically improve inference; multi-factor models may reduce inferability from returns by introducing transient components that weaken temporal dependence. They also do not imply that very long historical windows are intrinsically valuable; exponential decay of $v_t$24 quantifies diminishing returns from distant observations.

In aggregate, the information-theoretic picture is consistent across the reported model classes: past returns contain only limited information about current volatility and future returns in standard diffusive SV models. Persistence and leverage help, but realistic daily-frequency mutual information remains well below the amount required for precise volatility inference. A plausible implication is that effective volatility estimation and forecasting require richer observables, such as option-implied information, realized measures, or very high-frequency returns, when the inferential target exceeds the limits quantified by these mutual-information bounds.