Neural Network Calibration for Interest Rate Modelling, with Application to Regulatory Compliance
Milliman consultants from Warsaw and Paris look at the regulatory context of the weights calculation problem and some examples and results of calibration of interest rate models using neural network techniques.
Until recently most companies adopted rather arbitrary set of calibration instruments (swaptions) in calibrating of risk neutral economic scenarios. For example, a traditional preference was to use “10 year at the money swaptions”, without much justification. In some countries (e.g. Italy and France) regulators started to challenge the companies whether “the calibration process is relevant given the characteristics of [their] obligations” (EIOPA Guideline 57 on the valuation of technical provisions). Intuitively it is clear that for example insurance products with high guarantees should be calibrated to out of the money payer swaptions (thus payer swaptions with high strikes), while for low guarantees in the money payer swaptions should be preferred. This gave rise to the problem of proper selection of swaptions (e.g. in terms of maturities, tenors or strikes) for the calibration of risk neutral interest rate models so that they would “fit” the nature of insurance liabilities. This problem is referred by us as a weight selection problem, where weights refer to relative importance of each swaption for a given liability profile.