Impact-Neutral Measure Change

• Impact-Neutral Measure Change is a technique that adjusts a targeted drift while preserving key model invariants such as variance, covariance structures, and diffusion coefficients.
• It leverages tools like Girsanov’s theorem and Radon–Nikodym derivatives to transform between physical and risk-neutral measures, ensuring no-arbitrage conditions.
• The approach underpins practical applications in derivative pricing, stochastic programming, and risk models by maintaining structural properties while altering long-horizon expectations.

Searching arXiv for the cited papers and topic context. “Impact-neutral measure change” (Editor’s term) can denote a class of measure transformations in which a designated effect is neutralized while a selected structural component is preserved. In the sources considered here, the neutralized component may be the physical drift embedded in historical-return dynamics, the pricing-oriented drift of a risk-neutral model, the friction-induced drift in an incomplete-market stock model, or the decision effect of coherent risk aversion; the preserved component may be the diffusion coefficient, one-step return variance, Gaussian/affine tractability, positivity of a square-root factor, covariance structure, or the optimal value of an optimization problem under a worst-case measure (Tiwari, 21 Mar 2026, Alaya et al., 2024, Berninger et al., 2020, Liu, 2020, Liu et al., 2019).

1. Conceptual pattern and mathematical setting

The common template is a change from one reference measure to another, typically from the physical measure $\mathbb P$ to the risk-neutral measure $\mathbb Q$, or conversely from $\mathbb Q$ to a deliberately engineered real-world measure $\mathbb P$. In diffusion settings, the operative mechanism is Girsanov’s theorem: the diffusion coefficient is preserved while the drift is altered. In optimization settings, the corresponding mechanism is a Radon–Nikodym density $dQ=\zeta\,dP$, which transfers the effect of a coherent risk measure into a modified probability law (Alaya et al., 2024, Liu et al., 2019).

This suggests a unifying interpretation. The measure change is “neutral” not in the sense of removing all model content, but in the narrower sense of removing one targeted contribution while leaving the rest of the analytical architecture intact. In some cases the preserved object is local variance; in others it is an affine bond-pricing formula, the OU structure of Gaussian factors, or the state-space dynamics after a Lamperti transform. In yet others it is the optimal-value equivalence between a risk-averse and a risk-neutral optimization problem. The terminology is therefore structural rather than universal.

A recurrent misconception is that such a measure change must be economically or probabilistically innocuous. The sources do not support that interpretation. They support a narrower claim: drift can be modified without altering certain chosen invariants. The economic consequences may still be substantial, because altered drifts change term structures, long-horizon expectations, scenario distributions, derivative values, and policy choices (Berninger et al., 2020, Liu et al., 2019).

2. Score-space neutralization in diffusion-based derivative pricing

The most explicit “neutralizing correction” appears in the DDPM framework for derivative pricing under GBM. On a filtered probability space

$$(\Omega,\mathcal F,(\mathcal F_t)_{t\ge 0},\mathbb P),$$

the stock follows

$dS_t=\mu S_t\,dt+\sigma S_t\,dW_t,$

and the log-price $X_t=\log S_t$ satisfies

$$dX_t=\Bigl(\mu-\tfrac12\sigma^2\Bigr)\,dt+\sigma\,dW_t.$$

For a fixed $\Delta t$, the one-step log-return

$\mathbb Q$0

is Gaussian under $\mathbb Q$1: $\mathbb Q$2 Derivative pricing, however, requires simulation under $\mathbb Q$3, where

$\mathbb Q$4

so that

$\mathbb Q$5

The key structural fact is that the mean changes from $\mathbb Q$6 to $\mathbb Q$7 while the variance $\mathbb Q$8 is unchanged, and without this measure change generated paths violate the no-arbitrage martingale condition required for

$\mathbb Q$9

The paper’s exact contribution is to implement this measure change inside the reverse DDPM dynamics as an additive correction in score space, hence in $\mathbb Q$0-space (Tiwari, 21 Mar 2026).

With forward DDPM noising

$\mathbb Q$1

the forward marginals under $\mathbb Q$2 and $\mathbb Q$3 remain Gaussian with common variance

$\mathbb Q$4

and means

$\mathbb Q$5

Because two Gaussians with equal variance and different means have score functions differing by a constant,

$\mathbb Q$6

Using the DDPM score–noise relation, this becomes the closed-form correction

$\mathbb Q$7

with corrected predictor

$\mathbb Q$8

The reverse update is then run with $\mathbb Q$9, so the correction enters only through the mean term while the innovation variance $\mathbb P$0 is unchanged. The paper interprets this as adjusting drift while preserving the learned volatility and any higher-order structure encoded by the trained network.

The martingale consequence is explicit. If the generated one-step returns satisfy the risk-neutral mean and variance, then

$\mathbb P$1

so $\mathbb P$2 is a $\mathbb P$3-martingale. Empirically, under a GBM benchmark, the shifted model reproduces the risk-neutral terminal law and prices both European and arithmetic Asian options accurately. For a one-month ATM call with $\mathbb P$4, the reported diagnostics include mean one-step return $\mathbb P$5, standard deviation $\mathbb P$6, discounted mean $\mathbb P$7 versus $\mathbb P$8, KS statistic $\mathbb P$9, KS $dQ=\zeta\,dP$0-value $dQ=\zeta\,dP$1, and call price $dQ=\zeta\,dP$2 versus Black–Scholes $dQ=\zeta\,dP$3. For arithmetic Asians, the paper summarizes pricing errors as typically around $dQ=\zeta\,dP$4 of option value. The paper is also explicit that this closed-form additive correction is exact only in the Gaussian affine-score setting; beyond GBM, the $dQ=\zeta\,dP$5-shift becomes only an approximation (Tiwari, 21 Mar 2026).

A further clarification in that source is terminological. In this context, “impact-neutral” is not market-impact modeling. It is the neutralization of the physical drift impact embedded in a DDPM trained under $dQ=\zeta\,dP$6, so that arbitrage-free pricing can be carried out under $dQ=\zeta\,dP$7.

3. Engineered real-world deformations from a risk-neutral starting point

A different but related construction begins from a diffusion under $dQ=\zeta\,dP$8 and engineers a real-world measure through Girsanov’s theorem. The generic state process

$dQ=\zeta\,dP$9

is transformed by the Lamperti map

$$(\Omega,\mathcal F,(\mathcal F_t)_{t\ge 0},\mathbb P),$$0

into an additive-noise process

$$(\Omega,\mathcal F,(\mathcal F_t)_{t\ge 0},\mathbb P),$$1

or more generally

$$(\Omega,\mathcal F,(\mathcal F_t)_{t\ge 0},\mathbb P),$$2

under assumptions including local Lipschitz continuity, local integrability of $$(\Omega,\mathcal F,(\mathcal F_t)_{t\ge 0},\mathbb P),$$3, $$(\Omega,\mathcal F,(\mathcal F_t)_{t\ge 0},\mathbb P),$$4 regularity of $$(\Omega,\mathcal F,(\mathcal F_t)_{t\ge 0},\mathbb P),$$5 for Lamperti, monotonicity of $$(\Omega,\mathcal F,(\mathcal F_t)_{t\ge 0},\mathbb P),$$6, and a Feller-type boundary condition ensuring a unique strong solution in the state space (Alaya et al., 2024).

The distinctive feature is that the paper does not independently postulate the real-world dynamics. Instead it constructs an auxiliary process $$(\Omega,\mathcal F,(\mathcal F_t)_{t\ge 0},\mathbb P),$$7 such that under $$(\Omega,\mathcal F,(\mathcal F_t)_{t\ge 0},\mathbb P),$$8

$$(\Omega,\mathcal F,(\mathcal F_t)_{t\ge 0},\mathbb P),$$9

and, after a Girsanov change of measure with kernel

$dS_t=\mu S_t\,dt+\sigma S_t\,dW_t,$0

the same process satisfies under $dS_t=\mu S_t\,dt+\sigma S_t\,dW_t,$1

$dS_t=\mu S_t\,dt+\sigma S_t\,dW_t,$2

The Radon–Nikodym derivative is explicit: $dS_t=\mu S_t\,dt+\sigma S_t\,dW_t,$3 under the Novikov condition

$dS_t=\mu S_t\,dt+\sigma S_t\,dW_t,$4

The structural preservation is clear: the diffusion coefficient $dS_t=\mu S_t\,dt+\sigma S_t\,dW_t,$5 is unchanged, and the real-world process under $dS_t=\mu S_t\,dt+\sigma S_t\,dW_t,$6 has the same functional drift form $dS_t=\mu S_t\,dt+\sigma S_t\,dW_t,$7, now applied to $dS_t=\mu S_t\,dt+\sigma S_t\,dW_t,$8. At the same time, the model admits a deterministic deformation in Lamperti space,

$dS_t=\mu S_t\,dt+\sigma S_t\,dW_t,$9

hence

$X_t=\log S_t$0

This is not merely a conventional exogenous market-price-of-risk insertion. It is a target-fitting program in which $X_t=\log S_t$1 is calibrated and the Girsanov kernel $X_t=\log S_t$2 is then implied.

In the CIR++ credit-spread application, the square-root factor

$X_t=\log S_t$3

is Lamperti-transformed by $X_t=\log S_t$4, giving

$X_t=\log S_t$5

The intensity is

$X_t=\log S_t$6

with deterministic shift $X_t=\log S_t$7 fitting the initial survival curve, and under the real-world measure

$X_t=\log S_t$8

where

$X_t=\log S_t$9

The cumulative hazard rate obeys

$$dX_t=\Bigl(\mu-\tfrac12\sigma^2\Bigr)\,dt+\sigma\,dW_t.$$0

which is the basis for calibrating $$dX_t=\Bigl(\mu-\tfrac12\sigma^2\Bigr)\,dt+\sigma\,dW_t.$$1 to prescribed target curves. This makes the measure change explicitly scenario- and objective-dependent rather than unique. The paper therefore supports a structured RN-to-RW measure change, not a unique canonical one (Alaya et al., 2024).

4. Structure-preserving measure changes in Gaussian term-structure models

A closely related notion of neutrality appears in the Gauss2++ model, equivalently the two-factor Hull–White Gaussian short-rate model. Here the preserved object is analytical tractability. On

$$dX_t=\Bigl(\mu-\tfrac12\sigma^2\Bigr)\,dt+\sigma\,dW_t.$$2

with $$dX_t=\Bigl(\mu-\tfrac12\sigma^2\Bigr)\,dt+\sigma\,dW_t.$$3, the short rate is

$$dX_t=\Bigl(\mu-\tfrac12\sigma^2\Bigr)\,dt+\sigma\,dW_t.$$4

and under $$dX_t=\Bigl(\mu-\tfrac12\sigma^2\Bigr)\,dt+\sigma\,dW_t.$$5 the factors are OU processes,

$$dX_t=\Bigl(\mu-\tfrac12\sigma^2\Bigr)\,dt+\sigma\,dW_t.$$6

with $$dX_t=\Bigl(\mu-\tfrac12\sigma^2\Bigr)\,dt+\sigma\,dW_t.$$7. The zero-coupon bond price is affine-exponential,

$$dX_t=\Bigl(\mu-\tfrac12\sigma^2\Bigr)\,dt+\sigma\,dW_t.$$8

where $$dX_t=\Bigl(\mu-\tfrac12\sigma^2\Bigr)\,dt+\sigma\,dW_t.$$9, and the variance term $\Delta t$0 is explicit (Berninger et al., 2020).

The paper’s theorem-level statement is that any progressive and square-integrable function can be used to specify the change of measure without losing analytic tractability of zero-coupon bond prices in both worlds. In the independent-Brownian representation, the Girsanov kernel $\Delta t$1 can be parameterized through two time-dependent functions $\Delta t$2 and $\Delta t$3, interpreted as local long-run risk-premium levels. Under the new measure $\Delta t$4, the factors become

$\Delta t$5

with unchanged volatilities, unchanged correlation structure, and unchanged variance formula for the integrated short rate. The bond-pricing formula retains exactly the same affine-exponential form; only the state variables now evolve under the real-world drift.

This is a precise sense in which the measure change is impact-neutral with respect to analytical structure. The drifts change, but the Gaussian OU state dynamics, the volatility structure, the affine form of bond prices, and the variance $\Delta t$6 are preserved. The paper is equally explicit that this mathematical gentleness is not economically neutral. The choice of $\Delta t$7 and $\Delta t$8 changes long-horizon expectations of rates and yields, scenario generation, and risk measures.

The practical issue is the specification of the real-world risk premia. The paper studies constant, step, and linear specifications for $\Delta t$9 and $\mathbb Q$00. The constant case forces a single premium across the full horizon. The step case separates short-term and long-term premiums through a switching horizon $\mathbb Q$01. The linear case smooths the transition toward long-run levels and imposes differentiability at $\mathbb Q$02. In the empirical application, the time-varying step and linear specifications produce much more stable long-term interest-rate forecasts than the constant specification, while preserving tractability; the paper reports that the long-run absolute risk premium of the short rate typically lies between about $\mathbb Q$03 and $\mathbb Q$04 under the time-varying specifications (Berninger et al., 2020).

5. Correlation-sensitive neutralization in incomplete markets

The hard-to-borrow stock model introduces a different kind of neutralization problem. Under the physical measure, the stock and borrow-tightness dynamics are fully coupled: $\mathbb Q$05

$\mathbb Q$06

with correlated Brownian drivers

$\mathbb Q$07

The pricing objective is to replace the stock drift contribution $\mathbb Q$08 by $\mathbb Q$09, but the model is incomplete because buy-in risk is nontraded and a unique risk-neutral measure does not exist (Liu, 2020).

The technical correction in the paper is that the correlated Brownian motions must first be converted to an independent two-dimensional Brownian motion: $\mathbb Q$10 Defining

$\mathbb Q$11

the correct vector Girsanov kernel is

$\mathbb Q$12

The Radon–Nikodym density is therefore

$\mathbb Q$13

with

$\mathbb Q$14

Under $\mathbb Q$15, the stock becomes

$\mathbb Q$16

The neutralized component is the expected HTB-induced drift in the tradable stock. What is not neutralized is the state-dependent mechanism itself: $\mathbb Q$17 remains stochastic, the buy-in jump mechanism remains present, the state variable $\mathbb Q$18 remains coupled to stock returns, and the pricing measure is not unique because the market price of buy-in risk is represented by an arbitrary function $\mathbb Q$19. The paper therefore supports a limited but precise notion of neutralization: expected carry-like drift is removed from the stock equation, but the endogenous friction mechanism is retained. It also shows that when the stock shock and the borrow-tightness shock are correlated, the corresponding risk premia cannot be adjusted independently; the second kernel component must contain the correction term $\mathbb Q$20 (Liu, 2020).

6. Risk-neutral reformulation in stochastic programming and general limitations

A non-diffusion analogue appears in risk-averse stochastic programming. The paper studies the static problem

$\mathbb Q$21

and the multistage nested problem

$\mathbb Q$22

where the one-step conditional risk mapping is

$\mathbb Q$23

For coherent risk measures,

$\mathbb Q$24

so the risk-averse problem can, at a saddle point, be rewritten as a risk-neutral expectation minimization under a worst-case measure. In the multistage setting, the corresponding dual representation is

$\mathbb Q$25

This is the paper’s “risk neutral reformulation”: the effect of risk aversion is transferred into a modified measure that overweights bad scenarios (Liu et al., 2019).

For the specific mixed expectation–AVaR risk measure, the finite discrete reweighting is explicit. If $\mathbb Q$26 and $\mathbb Q$27, then

$\mathbb Q$28

with

$\mathbb Q$29

Hence the tail scenarios $\mathbb Q$30 receive larger probabilities under the stressed measure. In the multistage finite-scenario case, the paper uses stagewise frequencies of “bad” outcomes to construct modified scenario probabilities $\mathbb Q$31, then solves the resulting risk-neutral problem with standard SDDP machinery. The transformed problem is therefore neutral only in form: the objective becomes a plain expectation, but the scenario law has been deliberately deformed to preserve the impact of risk aversion on decisions (Liu et al., 2019).

Taken together, these works impose clear scope conditions on any “impact-neutral” interpretation. First, exact closed-form corrections are model-dependent. In the DDPM pricing paper, the additive score shift is exact because the $\mathbb Q$32 and $\mathbb Q$33 forward marginals are Gaussian with equal variance and different means; outside that setting, the same construction is only approximate (Tiwari, 21 Mar 2026). Second, validating the measure change can be technically difficult: the CIR++ RN-to-RW construction relies on Novikov’s condition and model-specific exponential-integrability arguments for $\mathbb Q$34 (Alaya et al., 2024). Third, preservation of tractability does not imply uniqueness of the target measure. The RW measure in the CIR++ framework depends on the calibrated deformation $\mathbb Q$35, and the HTB pricing measure is nonunique because the market price of buy-in risk is not pinned down by no-arbitrage alone (Alaya et al., 2024, Liu, 2020). Fourth, a structure-preserving measure change may still produce economically large effects: in Gauss2++, time-varying market-price-of-risk specifications materially alter long-horizon rate forecasts even though the affine-Gaussian machinery is unchanged (Berninger et al., 2020).

In that sense, the most defensible encyclopedia-level definition is narrow. An impact-neutral measure change is not a universal doctrine but a family of constructions in which the measure is changed so as to neutralize one specified effect while preserving a chosen invariant. The invariant may be martingale pricing, diffusion variance, Gaussian/affine solvability, positivity, covariance structure, or optimizer equivalence; the neutralized component may be a physical drift, a pricing drift, a friction-induced carry term, or the weighting of adverse scenarios. The technical content lies in specifying exactly what is neutralized, exactly what is preserved, and exactly under which assumptions the equivalence is valid.