Pricing a put takes two to three times the CPU time here, but estimator variance falls by three to four orders of magnitude. Both the Vasicek and CIR simulations deliver that gain under correlated equity and short-rate dynamics. The useful product is a Monte Carlo estimator. Whether listed options misprice stochastic-rate corrections is a separate question, one the paper does not ask.
The pricing problem is a European put whose underlying stock is correlated with a random short rate. A change of numéraire is the standard route under stochastic interest rates. As the paper observes, that route often ends in Monte Carlo anyway, especially for long-dated options and fast recalibration. The proposed gain is equal precision for less CPU. There is no identified mispricing to harvest: the evidence consists entirely of simulations under two rate models, Vasicek and CIR.
What does the Itô representation buy?
Antonelli and Ramponi use a log-price with constant volatility σ and drift equal to the short rate. Its Brownian motion has correlation ρ with a one-factor short rate. They avoid the usual change of numéraire to the forward measure. Their starting point is Itô's formula applied to the discount martingale multiplied by the Black-Scholes put. They evaluate that put at the current log-price and at R_s, the conditional expectation of the average rate between s and maturity.
This splits the discounted payoff into a finite-variation part and stochastic integrals with zero mean. Taking expectations eliminates those integrals. Three price components remain. First comes today's Black-Scholes term, N(t,t)P_t, with N denoting the discount martingale. Next is the integral ∫E[N(t,s) r_s P_s] ds, which contains r_s times the discounted put. Neither rate volatility nor correlation multiplies this term.
The remaining correction integrals carry weights in rate volatility and ρσ. Asset-rate covariation itself equals ρσ times the rate diffusion b(r). For affine rates (Hull-White, CIR++), the Clark-Ocone formula gives the rate martingale explicitly, with ρσ entering linearly outside the integrands. The authors emphasize that the joint asset-rate system need not be affine. CIR shows why this matters: asset-rate covariance contains √r, where transform methods cease to work.
The representation yields two tools.
One removes the zero-mean martingale on each path: Y_rep = Y_pay − M_{0,T}. In continuous time, the expectation stays the same while noise falls sharply. This is a martingale control variate in the family of Fouque and Han, whom the authors cite. The contribution is a control derived from a Black-Scholes decomposition that already incorporates stochastic rates and correlation. The other tool produces a semi-analytical price by freezing the Black-Scholes derivatives at time zero and retaining closed-form rate expectations. A further Itô expansion handles the leading term.
The data are entirely simulated. The ATM put has X0 = K = log(100), κ = 0.8, β = 0.04, r0 = 0.035, σ = 0.2 and ρ = 0.1. Tests use rate volatility ε of 0.01 or 0.05 and maturities from 91 to 365 days. Every run has 10^6 paths on a daily grid, without a deterministic shift (θ ≡ 0).
Paying 2x to 3x more CPU
A third to a half of the variance gain survives the extra computing cost. That still gives efficiency ratios of 795 to 21,800.
Vasicek variance-reduction factors range from 1.71×10^3 (ε = 0.05, one year) to 4.61×10^4 (ε = 0.01, three months). The representation estimator takes 2.075 to 2.148 times as much CPU as the naive estimator. Adjusting for that cost gives efficiency ratios from 7.95×10^2 to 2.18×10^4. CIR delivers similar results: variance reduction ranges from 2.69×10^3 to 4.62×10^4, with a CPU ratio of 2.624 to 3.031 and efficiency from 9.32×10^2 to 1.73×10^4.
Longer maturities erode the gain. At ε = 0.01, Vasicek efficiency drops from 2.18×10^4 at three months to 1.28×10^3 at one year. Even the weakest result makes equal precision about 800 times cheaper.
The benchmarks behind these tables warrant different levels of confidence. Vasicek supplies an exact closed-form price; the representation estimate is within about 5.1×10^-4 of it. The ε = 0.01, one-year row makes the benefit concrete. The naive estimate is 6.191148, compared with 6.175060 from the representation and an exact price of 6.174733. We did not find standard errors alongside the estimates.
Under correlation, CIR has no closed form. The authors instead benchmark against a Monte Carlo estimate of the same representation, arguing that this isolates approximation error. That argument works for approximation error alone. Bias shared by the representation and the discretization would escape this comparison. The rate uses full-truncation Euler, while the variance-reduction runs use trapezoidal time integrals. We did not find a quantification of that bias.
One put, one correlation, four maturities
The semi-analytical errors increase quickly with maturity. For Vasicek at ε = 0.01, the error grows from 1.7539e-05 at three months to 1.0930e-03 at one year, still below 0.02% relative. At ε = 0.05, it grows from 8.9538e-05 to 8.6329e-03, or about 0.14%. Between three months and one year, the error increases about 96-fold.
The relevant comparison is the price change when ε rises from 0.01 to 0.05. At one year, the exact put price increases from 6.174733 to 6.287819, a move of 0.113. The approximation error of 0.0086 at ε = 0.05 and one year amounts to 7.6% of that move. At three months, the price move is 0.0116 and the ε = 0.05 error is 8.95e-05.
Almost all the error comes from the first integral, which has neither ε nor ρ as a multiplier. At ε = 0.05 and one year, its error is 9.8284e-03. The two correction terms contribute −1.6528e-04 and −5.1645e-04. The CIR row for ε = 0.05 and one year tells the same story: the first term contributes −3.1025e-03, versus at most 5.8e-06 from the other four. Total error is −3.1180e-03 (about 0.05%).
The authors acknowledge the maturity limit. They describe the linear-in-ρ approximation as suitable for moderate maturities and call for a refined expansion at longer maturities. Yet the abstract says the experiments confirm accuracy. How far that claim extends beyond one year remains open.
We have doubts about the analytical bound. The local L2 error of order √(s−t) depends on uniformly bounded Black-Scholes x-derivatives. The paper's own formula for (∂xx−∂x)P, however, includes a 1/(σ√(T−s)) factor that blows up near expiry. We would use the tables as evidence for the approximation and regard the bound as unproven.
Before pricing SPY options
We are evaluating a strategy on SPY built from this method. It is our extension of the pricing method: the paper supplies no trading rule to reproduce. The representation prices European options on a non-dividend asset, whereas SPY options are American-style and SPY pays dividends. Screening reduces the mismatch from early exercise and dividends. It does not eliminate it.
Rate parameters pose a second gap. Historical short-rate dynamics do not identify risk-neutral parameters. We have no market-calibrated rate-derivatives surface, so calibration must use yields and option quotes under stated assumptions. Transaction costs and an independent pricing check also remain to be addressed.
The paper tests ρ = 0.1 and ε no higher than 0.05. Within those tests, increasing ε from 0.01 to 0.05 changes the one-year put price by 0.113. Any trading signal has to outperform flat-rate Black-Scholes using the term-structure rate after spreads.
We would use the estimator for pricing now. The relative-value trade remains unproven. An independent, converged one-year price matching the CIR representation price would increase our confidence in the semi-analytical formula and earn it a place as a fast benchmark.