A1871
Title: Historical calibration of SVJD models with deep learning
Authors: Milan Ficura - University of Economics in Prague (Czech Republic) [presenting]
Jiri Witzany - University of Economics in Prague (Czech Republic)
Abstract: The aim is to propose how deep neural networks can be used to calibrate the parameters of stochastic-volatility jump-diffusion (SVJD) models to historical asset returns. 1-dimensional convolutional neural networks (1D-CNN) are used for that purpose. The accuracy of the deep learning approach is compared with methods based on shallow neural networks and generalized moments, as well as with standard statistical approaches including MCMC and QMLE. The deep learning approach is found to be highly accurate and robust in simulation tests, surpassing even the best-performing statistical approaches with significantly lower miss-convergence rates. The main advantage of the deep learning approach is that it is fully generic and can be applied to any kind of SVJD model as long as simulations from the model can be drawn. An additional advantage of the approach is its speed in situations when the parameter estimation needs to be done repeatedly as the re-estimation of the SVJD model on new data is nearly instantaneous.