In this thesis, we analyse the relationship between the hedging activities and return on equity, and the relationship between profit on hedging and other factors. Fully conditional specification is used to impute the missing values. Instrumental variable estimation and finite mixture of regression models are then used to predict the return on equity and hedging gain. We find the instrumental variable estimation is better than the OLS estimation to deal with the hedging data since it eliminates the endogeneity. By finite mixture of regression models, we show that different firms have different hedging strategies, which cause different profits in hedging. We also find the companies with large total assets prefer to hedge.

Submitted by Divisão de Documentação/BC Biblioteca Central (ddbc@ufam.edu.br) on 2018-05-29T14:38:33Z No. of bitstreams: 2 license_rdf: 0 bytes, checksum: d41d8cd98f00b204e9800998ecf8427e (MD5) VersaoFinal.pdf: 2882901 bytes, checksum: a35c6d27fe0f9aa61dfe3a96244b3140 (MD5) / Approved for entry into archive by Divisão de Documentação/BC Biblioteca Central (ddbc@ufam.edu.br) on 2018-05-29T14:38:46Z (GMT) No. of bitstreams: 2 license_rdf: 0 bytes, checksum: d41d8cd98f00b204e9800998ecf8427e (MD5) VersaoFinal.pdf: 2882901 bytes, checksum: a35c6d27fe0f9aa61dfe3a96244b3140 (MD5) / Made available in DSpace on 2018-05-29T14:38:46Z (GMT). No. of bitstreams: 2 license_rdf: 0 bytes, checksum: d41d8cd98f00b204e9800998ecf8427e (MD5) VersaoFinal.pdf: 2882901 bytes, checksum: a35c6d27fe0f9aa61dfe3a96244b3140 (MD5) Previous issue date: 2018-03-12 / CAPES - Coordenação de Aperfeiçoamento de Pessoal de Nível Superior / The traditional estimation of mixture regression models is based on the assumption of normality of component errors and thus is sensitive to outliers, heavy-tailed and/or asymmetric errors. Another drawback is that, in general, the analysis is restricted to directly observed predictors. We present a proposal to deal with these issues simultaneously in the context of mixture regression by extending the classic normal model by assuming that, for each mixture component, the random errors and the covariates jointly follow a scale mixture of skew-normal distributions. It is also assumed that the covariates are observed with error. An MCMC-type algorithm to perform Bayesian inference is developed and, in order to show the efficacy of the proposed methods, simulated and real data sets are analyzed. / A estimação tradicional em mistura de modelos de regressão é baseada na suposição de normalidade para os erros aleatórios, sendo assim, sensível a outliers, caudas pesadas e erros assimétricos. Outra desvantagem é que, em geral, a análise é restrita a preditores que são observados diretamente. Apresentamos uma proposta para lidar com estas questões simultaneamente no contexto de mistura de regressões estendendo o modelo normal clássico. Assumimos que, conjuntamente e em cada componente da mistura, os erros aleatórios e as covariáveis seguem uma mistura de escala da distribuição normal assimétrica. Além disso, é feita a suposição de que as covariáveis são observadas com erro aditivo. Um algorítmo do tipo MCMC foi desenvolvido para realizar inferência Bayesiana. A eficácia do modelo proposto é verificada via análises de dados simulados e reais.