Technical report : Risk-neutral density recovery via spectral analysis

TL;DR

This paper introduces a spectral analysis-based method for directly estimating the risk-neutral density from option prices, offering a fast, robust, and nonparametric approach that improves upon existing techniques.

Contribution

The paper proposes the spectral recovery method (SRM), a novel, efficient, and nonparametric technique for risk-neutral density estimation using spectral analysis and quadratic programming.

Findings

SRM is fast and simple to implement.

It does not require calibration or data smoothing.

It performs well on real and simulated data.

Abstract

In this paper, we propose a new method for estimating the conditional risk-neutral density (RND) directly from a cross-section of put option bid-ask quotes. More precisely, we propose to view the RND recovery problem as an inverse problem. We first show that it is possible to define restricted put and call operators that admit a singular value decomposition (SVD), which we compute explicitly. We subsequently show that this new framework allows us to devise a simple and fast quadratic programming method to recover the smoothest RND whose corresponding put prices lie inside the bid-ask quotes. This method is termed the spectral recovery method (SRM). Interestingly, the SVD of the restricted put and call operators sheds some new light on the RND recovery problem. The SRM improves on other RND recovery methods in the sense that: - it is fast and simple to implement since it requires solution of a single quadratic program, while being fully nonparametric; - it takes the bid ask quotes as sole input and does not require any sort of calibration, smoothing or preprocessing of the data; - it is robust to the paucity of price quotes; - it returns the smoothest density giving rise to prices that lie inside the bid ask quotes. The estimated RND is therefore as well-behaved as can be; - it returns a closed form estimate of the RND on the interval [0,B] of the positive real line, where B is a positive constant that can be chosen arbitrarily. We thus obtain both the middle part of the RND together with its full left tail and part of its right tail. We confront this method to both real and simulated data and observe that it fares well in practice. The SRM is thus found to be a promising alternative to other RND recovery methods.