Ito Calculus: Why Curvature Changes the Economics of a Hedge

The practical discipline is to maintain one coherent chain: define the stochastic process, identify the integration convention, specify the probability measure, state financing and payoff terms, and then derive the value and hedge. A mathematically correct formula becomes a useful financial calculation only when every link describes the same contract and decision.

An option payoff is nonlinear in its underlying asset. When that asset moves randomly, curvature contributes to the evolution of the option’s value. Ito calculus is the framework that keeps this contribution rather than discarding it as ordinary first-order intuition might.

This reference connects stochastic integrals, the Ito correction, alternative integration conventions and derivative valuation. Its central practical point is that a formula’s timing convention, measure and financing assumptions must agree.

Why the integral needs a timing convention

For a simple trading strategy, holdings $H_i$ are selected at time $t_i$ and applied to the following increment of a Brownian driver. The associated sum is

$$\sum_i H_i(W_{t_{i+1}}-W_{t_i}).$$

Each holding uses information available before the next shock. Under suitable measurability and integrability conditions, limiting these sums defines the Ito integral $\int H_t\,dW_t$.

The construction is not simply an ordinary Riemann integral with an unusual integrator. Brownian paths have unbounded variation, and their squared increments accumulate to elapsed time. Classical bounded-variation arguments do not apply in general. Some special pathwise integrals may still exist under additional regularity; that does not remove the need to specify which integral is intended.

For a predictable integrand satisfying $\mathbb E\int_0^T H_t^2dt<\infty$, the Ito integral has mean zero and obeys the isometry

$$\mathbb E\left[\left(\int_0^T H_t\,dW_t\right)^2\right] =\mathbb E\left[\int_0^T H_t^2dt\right].$$

In words: The average squared random trading gain equals the average accumulated squared exposure over time.

The right side converts random trading exposure into accumulated variance. With weaker, local conditions, one obtains a local martingale; a zero-mean claim then needs additional justification.

Ito’s formula keeps curvature

Suppose a state variable follows

$$dX_t=a(X_t,t)dt+b(X_t,t)dW_t.$$

Here $a$ is drift, $b$ is diffusion and $W$ is Brownian motion. For a function $f(x,t)$ with one time derivative and two space derivatives, Ito’s formula gives

$$df(X_t,t)=\left(f_t+af_x+\tfrac12b^2f_{xx}\right)dt+bf_xdW_t.$$

In words: Value changes with time, the expected state move, curvature acting on uncertainty and the random state move.

All derivatives are evaluated at the current state and time. The extra second-derivative term is the effect of quadratic variation. It survives because stochastic increments have size proportional to the square root of elapsed time.

For $f(x)=x^2$ and $X_t=W_t$,

$$d(W_t^2)=2W_tdW_t+dt.$$

Integrating gives the useful identity

$$\int_0^T W_t\,dW_t=\frac12(W_T^2-T).$$

The ordinary-calculus answer $W_T^2/2$ would have the wrong expectation. Subtracting $T/2$ restores the zero mean expected from the square-integrable Ito integral.

Multiplicative risk and the logarithmic correction

Let an asset price satisfy geometric Brownian motion,

$$dS_t=\mu S_tdt+\sigma S_tdW_t,$$

with expected proportional drift $\mu$, volatility $\sigma$ and positive starting price. Applying Ito’s formula to $\log S_t$ gives

$$d\log S_t=(\mu-\tfrac12\sigma^2)dt+\sigma dW_t.$$

The expected arithmetic growth rate and expected log-growth rate differ. This is a distributional consequence of multiplicative uncertainty, not an extra cash expense.

The exact solution is

$$S_T=S_0\exp\left[(\mu-\tfrac12\sigma^2)T+\sigma W_T\right].$$

Its mean is $S_0e^{\mu T}$, while its median is $S_0e^{(\mu-\sigma^2/2)T}$. A forecast of the average level across scenarios therefore should not be presented as the most representative realised path without further explanation.

Several risk factors create cross terms

For two states driven by Brownian motions with instantaneous correlation $\rho$, a function of both states acquires a mixed derivative term. If the diffusion coefficients are $b_1$ and $b_2$, that contribution is

$$\rho b_1b_2f_{x_1x_2}dt.$$

It records the effect of simultaneous co-movement. Omitting it changes the valuation equation for multi-asset products and stochastic-volatility models. Correlation is not just an input to a final portfolio aggregation; it can enter the evolution of the value itself.

This explains why a price-volatility correlation assumption matters for option smiles and why cross-gamma exposure cannot always be reduced to two separate single-factor exposures.

Ito and Stratonovich represent different conventions

The Stratonovich integral uses a symmetric convention and supports the classical chain rule. For continuous semimartingales, the conversion is

$$\int_0^T H_t\circ dW_t =\int_0^T H_t\,dW_t+\frac12[H,W]_T,$$

where $[H,W]$ is quadratic covariation and the circle denotes Stratonovich integration.

If $dX_t=a_I(X_t,t)dt+b(X_t,t)dW_t$ is written in Ito form, its equivalent Stratonovich drift is

$$a_S=a_I-\frac12b\,\frac{\partial b}{\partial x}.$$

For additive noise, $b$ does not depend on the state, so the correction vanishes. For multiplicative noise it generally does not. The equation $dX_t=X_t\circ dW_t$, for example, becomes $dX_t=\tfrac12X_tdt+X_tdW_t$ in Ito form.

The two conventions can describe the same process when the drift is converted correctly. It would be inaccurate to say that Stratonovich calculus is intrinsically incompatible with finance or that it ignores quadratic variation. It incorporates that effect through a different convention. Ito’s non-anticipating sums remain the natural formulation for self-financing trading gains.

Self-financing means rebalancing is funded

Let a replicating portfolio hold $\Delta_t$ units of an underlying asset and $\beta_t$ units of a money-market account $B_t$. Its value is $V_t=\Delta_tS_t+\beta_tB_t$. Self-financing means

$$dV_t=\Delta_tdS_t+\beta_tdB_t.$$

Changes in holdings are funded by selling or buying the other component. They do not create an external cash injection. This is why a derivation must not differentiate $V-\Delta S$ while silently discarding the financing effects of a changing $\Delta$.

For constant risk-free rate $r$, $dB_t=rB_tdt$. If a smooth derivative value is $V(S,t)$, Ito’s formula identifies $\Delta=V_S$ as the underlying holding that matches the diffusion term. Matching the drift of the derivative with that of the self-financing portfolio yields

$$V_t+\frac12\sigma^2S^2V_{SS}+rSV_S-rV=0.$$

This is the basic Black–Scholes equation for a non-dividend-paying underlying under the familiar continuous-trading assumptions. Expected physical return $\mu$ disappears because the risk is replicated within that model.

Pricing expectations and discounting

Under the corresponding risk-neutral dynamics, $dS_t=rS_tdt+\sigma S_tdW_t^{\mathbb Q}$. Applying Ito’s formula to the discounted price $e^{-rt}S_t$ removes its drift. With suitable integrability, the discounted price is a martingale.

For terminal payoff $g(S_T)$, the valuation is

$$V_t=\mathbb E_{\mathbb Q}[e^{-r(T-t)}g(S_T)\mid\mathcal F_t].$$

In words: The current value is the expected discounted terminal payoff under pricing probabilities, using information available now.

The expectation and the PDE are complementary representations of the same valuation problem when their assumptions match. A mismatch in dividends, discounting, boundary conditions or payoff timing prevents their numerical results from agreeing.

The result is not a claim that risk-neutral probabilities describe actual outcomes. Physical-measure scenarios are used for risk forecasting and economic expectations; pricing-measure scenarios are calibrated and interpreted for valuation.

Product rules explain where financing terms go

For continuous semimartingales $X_t$ and $Y_t$, the product rule is

$$d(X_tY_t)=X_tdY_t+Y_tdX_t+d[X,Y]_t.$$

The final term is quadratic covariation. It vanishes if one factor is a continuous finite-variation process, but not generally when both contain correlated diffusion risk.

For the deterministic discount factor $e^{-rt}$, the cross term with an asset price is zero. The ordinary derivative of the discount factor cancels the asset’s pricing-measure drift, producing the familiar driftless discounted process. If the discounting account is itself built from stochastic rates, it still has finite-variation paths under the usual money-market definition, but its value is random and must remain inside the conditional expectation.

For a portfolio $\Delta_tS_t+\beta_tB_t$, direct differentiation produces changes in holdings and their covariations as well. The self-financing condition imposes the relationship that makes these rebalancing terms net to zero. It is an economic constraint on the holdings process, not permission to ignore terms arbitrarily.

A complete one-step hedge ledger

Consider an explicitly illustrative short call worth 10 with delta 0.6 when the underlying is 100.

Initial hedge ledger records the option, underlying and funding

After a small time step, suppose the underlying rises to 102 and the option’s delta becomes 0.65. Before rebalancing, the stock holding is worth 61.2, the short option must be marked at its new model or market value, and the borrowing has accrued financing cost. Buying another 0.05 underlying units costs 5.1, which increases borrowing unless other cash is provided.

The hedge’s profit cannot be calculated as the change in stock value minus the change in option value while omitting financing and the cost of new holdings. This ledger is also why gross trading volume is not profit: cash spent to rebalance has a corresponding position change.

No new option price is invented in this example. A numerical hedge profit would require that price, the time interval, financing convention and transaction costs. Stating what is missing is preferable to completing a persuasive-looking calculation with unspecified inputs.

Curvature connects local risk and expected change

For a small shock with conditional mean zero and conditional variance $v$, a second-order Taylor approximation gives expected value change of roughly $\tfrac12 f''(x)v$, ignoring higher-order effects. A convex function benefits in expectation from dispersion under that local approximation; a concave function loses.

Ito’s formula is the continuous-time counterpart with the appropriate quadratic variation. It explains why simply replacing a random input by its expectation can misprice a nonlinear payoff. The conclusion is conditional on the distribution, curvature and approximation region; jumps and large shocks can require a fuller calculation.

This connection provides a practical diagnostic. If a model with positive gamma shows no curvature-related effect when volatility changes, the implementation deserves inspection. If the model applies the same second-order approximation to an extreme move across a payoff discontinuity, the approximation itself may be inadequate.

Interpreting the reported sensitivities

Delta is $V_S$, gamma is $V_{SS}$, and vega is the derivative of value with respect to a specified volatility input. For a small underlying change $\Delta S$, a local value approximation includes $\Delta\,\Delta S+\tfrac12\Gamma(\Delta S)^2$, alongside time, volatility and other risk-factor terms.

A delta hedge removes the model’s first-order underlying exposure at the rebalancing instant. It does not remove gamma, changing volatility, jumps, discrete trading intervals, transaction costs or all model error. The Ito isometry can quantify a specified residual stochastic integral, but total hedging loss may contain drift, costs and unmodelled components as well.

Distinguish a mathematical hedge from a risk limit

A model hedge specifies holdings that neutralise particular local risks under stated dynamics. A risk limit specifies how much loss or exposure the institution is willing to tolerate. The hedge can reduce measured exposure without eliminating the need for the limit.

For example, a delta-neutral book can remain short gamma and lose when prices move sharply before the next rebalance. A vega-neutral book can remain exposed to changes in the shape of the volatility surface because different strikes and maturities move differently. Those are distinct risks rather than evidence that the derivatives were computed incorrectly.

The valuation model also determines which state changes the Greeks recognise. If the actual market introduces a jump or basis movement absent from the model, the local expansion may not explain the realised profit and loss. Reconciliation should distinguish implementation error, approximation error and a missing risk factor.

This makes the calculus operationally useful. It supplies a precise local explanation, while the risk process examines whether the local model is sufficient for the size, horizon and liquidity of the position.