A nonparametric Bayesian approach to term structure fitting. Issue 3 (12th July 2019)

• Record Type: Journal Article
• Authors: Lapshin, Victor
• Is Part Of: Studies in economics and finance. Volume 36:Issue 3(2019)
• Page Start: 600
• Page End: 615
• Publication Date: 2019-07-12
• Subjects: Regularization -- Bayesian approach -- Bond pricing -- Infinite-dimensional Bayesian inverse problems -- Term structure fitting -- Markov Chain Monte-Carlo; Economics -- Periodicals; Finance -- Periodicals; 330
• DOI: 10.1108/SEF-01-2018-0025
• Languages: English
• ISSNs: 1086-7376

Abstract

Abstract : Purpose: This paper aims to illustrate how a Bayesian approach to yield fitting can be implemented in a non-parametric framework with automatic smoothing inferred from the data. It also briefly illustrates the advantages of such an approach using real data. Design/methodology/approach: The paper uses an infinite dimensional (functional space) approach to inverse problems. Numerical computations are carried out using a Markov Chain Monte-Carlo algorithm with several tweaks to ensure good performance. The model explicitly uses bid-ask spreads to allow for observation errors and provides automatic smoothing based on them. Findings: A non-parametric framework allows to capture complex shapes of zero-coupon yield curves typical for emerging markets. Bayesian approach allows to assess the precision of estimates, which is crucial for some applications. Examples of estimation results are reported for three different bond markets: liquid (German), medium liquidity (Chinese) and illiquid (Russian). Practical implications: The result shows that infinite-dimensional Bayesian approach to term structure estimation is feasible. Market practitioners could use this approach to gain more insight into interest rates term structure. For example, they could now be able to complement their non-parametric term structure estimates with Bayesian confidence intervals, which would allow them to assess statistical significance of their results. Originality/value: The model does not require parameter tuning during estimation. It has its own parameters, but they are to be selected during model setup.