This master thesis attempts to introduce a Bayesian estimation which used the NonParametric Simulated Maximum Likelihood Estimator (NPSMLE) in order to compute an approximation of the agent-based model likelihood. The main achievement of this master thesis is the creation of an adaptive Gibbs sampler which takes into account the shape of the likelihood function in order to explore the parameter space in a proper way.
To test the accuracy of our Gibbs sampler, laboratory experimentation have been conducted in order to assess the extent to which our simulated Bayesian method is better than the NPSMLE. Finally, our Bayesian method has been tested on the S\&P500 index in normal and crisis economic conditions. We used the most famous Heterogeneous Agent Models, the Adaptive belief system (Brock and Hommes, 1998). This new Bayesian method is more accurate than its frequentist equivalent in laboratory conditions because it has a smaller variance and bias. The sample size of our bias-variance analysis is small, therefore it constitutes a strong limitation to our work. Cloud computing has to be used in order to increase this sample and have a better estimation of the true bias and variance of our estimator. To conclude, it is clear that a paradigm shift is necessary in order to improve calibration accuracy in agent-based models.
|la date de réponse||18 juin 2020|
|Superviseur||Sophie Bereau (Promoteur)|