In this thesis, we investigate model selection in a general setting and perform several exercises in financial econometrics. We present the multi-hypothesis testing (MHT) framework, with which we design different type of model comparisons. We distinguish between test of model performance significance, of relative and absolute model performance and apply our framework to market risk forecasting model, to latent factor jump-diffusion models employed for the estimation of the statistical measure of an equity index, as well as to equity option pricing models. We develop original tests and, with regard to the proper exercise of model selection from an initial battery of models without any reference to a benchmark model, we combine the MHT approach with the model confidence set (MCS) to deliver a novel test of model comparison that is performed along with the established version of the MCS, as well as with an alternative simplified new MCS test that are detailed in the course of this work. We collect empirical evidence concerning model comparison in several subjects. With respect to market risk forecasting models, we have found that models capturing volatility clustering or targeting directly an auto-correlated conditional distribution percentile, perform better than the target model set and in particular, better than the historical simulation, widely employed by practitioners, and better than the so called RiskMetrics model. With respect to the equity index data dynamics, we have found that the popular affine jump-diffusion model requires a CEV augmentation to perform appropriately and that those models are slightly overperformed by an alternative stochastic volatility model, characterised by stochastic hazard with high frequency small jumps. The test performed over a large model set employed in the option pricing exercise points to a wide similarity of the results obtained by the many model specifications of the superior exponential volatility model, therefore suggesting a more careful adjustment of the model complexity. The model selection framework has proven very flexible in dealing with the varied collection of statistical problems. In particular, our main contribution represented by the generalised MHT based MCS test provides a method for model selection that is robust to finite sample distribution and that has the advantage of an adjustable tolerance for false rejections, allowing conservative to aggressive testing profiles.