Revisiting Stochastic Collocation with Exponential Splines for an Arbitrage-Free Interpolation of Option Prices
Fabien Le Floc’h · 2025
Read Revisiting Stochastic Collocation with Exponential Splines for an Arbitrage-Free Interpolation of Option Prices by Fabien Le Floc’h (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-12419.txt.
Introduction
Fabien Le Floc'h takes a fresh look at a math tool used to price options, in a solo paper from two thousand twenty-five. The tool is stochastic collocation, and here it uses the exponential of a quadratic spline. His aim is a way to fill in option prices that does not allow free-money gaps.
He checks two ways to build the spline in close detail. One way fixes the height points and tunes the place points. The other fixes the place points and tunes the terms of a B-spline form. By weighing which choice works best, the paper helps users pick a sound setup for this price task.
Catalog record
| CREATOR | Fabien Le Floc’h |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 27 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2508-12419
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Revisiting Stochastic Collocation with Exponential Splines for an Arbitrage-Free Interpolation of Option Prices Fabien Le Floc’h Abstract We revisit the stochastic collocation method using the exponential of a quadratic spline. In particular, we look in details whether it is more appropriate to fix the ordinates and optimize the abscissae of an interpolating spline or to fix the abscissae and optimize the parameters of a B-spline representation. keywords stochastic collocation; implied volatility; quantitative finance; arbitrage-free; risk neutral density; B-spline † † firstpage: 1 † † volume: xx † † issue: 1 † † articlenumber: 5 † † year: 2025 † † copyright-year: 2025 † † authornames: Fabien Le Floc’h † † address: Independent researcher † † corresponding: Correspondence: [email protected] \history Received: Aug 15 2025 1 Introduction Le Floc’h and Oosterlee 2019 introduce the stochastic collocation method utilizing either a spline or the exponential of a spline to achieve an arbitrage-free interpolation of vanilla option prices, or, equivalently, their implied volatilities. The formula for the undiscounted price of a call option C with strike K with the exponential spline collocation reads \displaystyle C(K) \displaystyle=\int_{-\infty}^{\infty}\max(e^{g(x)}-K,0)\phi(x)dx\,, (1) where g is a monotonic quadratic spline function interpolating (x_{i},\ln y_{i})_{i=0,...,N} , y_{i} are option strikes, \phi is the normal probability density function. Let (\bar{x}_{j})_{j=0,...,M} be the spline knots, and g(x)=a_{j}+b_{j}(x-\bar{x}_{j})+c_{j}(x-\bar{x}_{j})^{2} on [\bar{x}_{j},\bar{x}_{j+1}] , with k the index such that \bar{y}_{k}\leq K<\bar{y}_{k+1} , we have then \displaystyle C(K) \displaystyle=I(\tilde{x}_{-1},\bar{x}_{0})+I(\tilde{x}_{M},\infty)+K\Phi(-\tilde{x}_{k}) \displaystyle+\sum_{j=k}^{M-1}e^{a_{j}-b_{j}\bar{x}_{j}+c_{j}\bar{x}_{j}^{2}+\frac{1}{2}m_{j}^{2}}\frac{\Phi\left(\bar{x}_{j+1}\sqrt{1-2c_{j}}-m_{j}\right)-\Phi\left(\tilde{x}_{j}\sqrt{1-2c_{j}}-m_{j}\right)}{\sqrt{1-2c_{j}}}\,, (2) with m_{j}=\frac{b_{j}-2c_{j}\bar{x}_{j}}{\sqrt{1-2c_{j}}} and defining \tilde{x}_{k}=g^{-1}(\ln K) , \tilde{x}_{i}=\bar{x}_{i} for i>k . We consider the cumulative normal density function in the complex plane: when 1-2c_{j}<0 , we may use the imaginary error function erfi to compute its value. For a linear extrapolation with slope s and passing by the point (x,y) , corresponding to g(u)=s(u-x)+y , we find \displaystyle I(a,b) \displaystyle=\int_{a}^{b}e^{s(u-x)+y}\phi(u)du \displaystyle=e^{y-sx+\frac{1}{2}s^{2}}\left(\Phi(b-s)-\Phi(a-s)\right)\,, (3) with some abuse of notation \Phi(-\infty)=0 , \Phi(\infty)=1 . The left wing extrapolation corresponds to x=\bar{x}_{0} and y={a}_{0} , with s=s_{L} a free parameter and the right wing extrapolation corresponds to x=\bar{x}_{M},y={a}_{M} with s=s_{R} . In order to keep the continuity of the derivative at the boundaries, we choose s_{R}=g^{\prime}(x_{M}) and s_{L}=g^{\prime}(x_{0}) . The first moment is given by \displaystyle M_{1} \displaystyle=\int_{-\infty}^{\infty}e^{g(x)}\phi(x)dx \displaystyle=I(-\infty,\bar{x}_{0})+I(\bar{x}_{M},\infty) \displaystyle+\sum_{j=0}^{M-1}e^{a_{j}-b_{j}\bar{x}_{j}+c_{j}\bar{x}_{j}^{2}+\frac{1}{2}m_{j}^{2}}\frac{\Phi\left(\bar{x}_{j+1}\sqrt{1-2c_{j}}-m_{j}\right)-\Phi\left(\bar{x}_{j}\sqrt{1-2c_{j}}-m_{j}\right)}{\sqrt{1-2c_{j}}}\,. (4) In particular, Le Floc’h and Oosterlee 2019 explore the use of a B-spline parameterization with fixed knots \displaystyle g(x)=\sum_{j=0}^{M}\alpha_{j}B_{j,3}(x)\,, (5) where the B-spline parameters \alpha_{j} are calibrated such as to minimize the \ell_{2} -norm between weighted model prices and market prices. One challenge is to find a proper choice of knots in the x space. Another approach, which is suggested but not fully explored is to consider a regular spline interpolation, fixing the ordinates y instead of the knots x . The knots x are calibrated and y may be taken to be simply the market strikes. A standard B-spline parameterization is then not practical, since it would require solving a quadratic programming problem at each iteration of the non-linear least-squares minimization in order to enforce positivity of the B-spline coefficients, and thus monotonicity (see Wolberg and Alfy 2002 ). Instead, it is easier to rely on a monotonic spline interpolation representation. For quadratic splines, Schumaker 1983 proposes an algorithm to insert knots such that the interpolant preserves monotonicity and convexity and the spline is of class C^{1} . This latter approach has been rediscovered by Roos 2025 . Motivated by its encouraging results, we study in this note the choice of fixing y versus fixing x in more details. 2 When the Schumaker spline does not preserve monotonicity The quadratic spline from Schumaker 1983 preserves the shape, but not necessarily the monotonicity as ( Schumaker 1983 , Property 3.2) is not guaranteed to hold. A counter-example presented in Table 1 and corresponds to a calibration on the market data of Table 3 (case II). Table 1: Counter-example where the Schumaker spline g on (x,y) is not monotonic but the data is. We have g^{\prime}(3.03)<-0.0472 . x y -3.8732183006023453 -3.34887695753723 -3.0522883128452993 -3.0139892617835105 -1.7943713634054417 -2.6791015660297828 -1.6512340998496167 -2.344213870276071 -1.6259385165727456 -2.009326174522346 -1.6006430286358648 -1.6744384787686222 -1.578610520539529 -1.3395507830148954 -1.5563726083805842 -1.0046630872611706 -1.4706329281117003 -0.6697753915074481 -1.3853673387319743 -0.33488769575372446 -1.2808142988169438 0.0 -0.974725079691088 0.3348876957537252 -0.799531535567272 0.6697753915074496 -0.2964451400432737 1.00466308726117 0.40549865892657655 1.3395507830148956 0.8534594889549447 1.6744384787686188 2.5550732737844375 2.009326174522343 2.822138011027902 2.3442138702760635 2.887184595942125 2.6791015660297917 3.1735718560441963 3.0139892617835136 3.294759066619504 3.348876957537241 A solution is to use a harmonic mean estimate for the derivatives at the initial knots ( Lam 1990 ) . This is the variation on the original Schumaker algorithm that we make use for our numerical experiments. 3 The issue with fixing the ordinates Le Floc’h and Oosterlee 2019 pay attention to preserve the first moment exactly in order to guarantee the arbitrage-free property of the interpolation over the full range of strikes. Instead, the earlier stochastic collocation on a polynomial of Grzelak and Oosterlee 2017 was computing the call option price and deducing the put option price from the put-call parity relationship. In this latter approach, there is no guarantee then that the put option price will be positive for all strikes as the undiscounted call option price obtained by such a polynomial collocation may be below the intrinsic value F-K . For a positive asset, if we assume absorption at zero, the put-call parity must be adjusted using the call option price at strike zero: \tilde{F}=C(0) . As the forward price comes from the market, the value of \tilde{F} is given, and we are back to enforcing the preservation of the first moment. The simplest way to enforce the first moment condition is to shift the ordinates of the interpolation by \ln F-\ln C(0) . This means that the ordinates are no longer truly fixed: at each iteration, the first moment condition will shift the ordinates. It is not obvious a priori if the shift is large or small. In this analysis, the B-spline is always calibrated using fixed knots (abscissae) by optimizing its parameters, and the Schumaker spline is always calibrated from a fixed initial set of ordinates (but which moves to enforce the first moment condition) by optimizing its abscissae. We consider the example market data of Jäckel 2014 presented in Table 3 (case II), and use two different initial guesses, with initial ordinates at the market strikes: (i) The equivalent Black-Scholes implied volatility is constant equal to the at-the-money volatility. \displaystyle a_{j}=\ln K_{j}\,,\quad b_{j}=\sigma(F(T),T)\sqrt{T}\,,\quad x_{j}=\frac{1}{b_{j}}\left(a_{j}+\frac{b_{j}^{2}}{2}-\ln F(T)\right)\,, (6) where \sigma(K,T) is the market 1 1 1 The at-the-money volatility may computed by interpolating with a quadratic on the nearest neighbours. Black-Scholes implied volatility at strike K for maturity T . (ii) The abscissae and slopes use the full implied volatility smile. \displaystyle a_{j}=\ln K_{j}\,,\quad b_{j}=\sigma({K_{j},T})\sqrt{T}\,,\quad x_{j}=\frac{1}{b_{j}}\left(a_{j}+\frac{b_{j}^{2}}{2}-\ln F(T)\right)\,. (7) In (i), the first moment is actually preserved directly and the initial ordinates correspond exactly to the market strikes. In (ii), a small shift of around 0.18 is incurred, which is well below the ordinate of next strike for this data (the spacing in log-strikes is approximately 0.33). During the Levenberg-Marquardt minimization starting from the initial guess (i), the ordinates move significantly off the original strikes. This is not as much pronounced with initial guess (ii), but there is no guarantee that this will hold for other market data. (a) Schumaker (b) B-spline Figure 1: First ordinate of the exponential spline collocation as a function of the iteration of the Levenberg-Marquardt minimizer. The minimizer is applied