Yield Curve Quantization and Simulation with Neural Networks

Abstract

We present a method for simulating yield curve dynamics by learning the curve distribution from historical data using Artificial Neural Networks (ANN) in a two step procedure. The first step involves an autoencoder which performs a quantization of curve moves, generating a set of representative curve shapes. The second step learns a probability distribution over the quantized shapes, conditional on the current curve and the shift of a single pivot tenor point. This allows to simulate the curve by first drawing the the pivot tenor shift and then the shape of the curve move from its dynamic distribution. A suitable choice of regularizers allows to keep the simulation statistics close to the original data.