Middleware technologies have been widely adopted by the software industry to reduce the cost of developing computer systems. Nonetheless, predicting the performance of middleware-based applications is difficult due to specific implementation details of a middleware platform and a multitude of settings and services provided by middleware for different deployment scenarios. Thus, the performance management of middleware-based applications can be a non-trivial task. Autonomic computing is a new paradigm for building self-managed systems, i.e., systems that seek to operate with minimal human intervention. This work investigates the use of statistical approaches to building autonomic management solutions to control the performance of middleware-based applications. Particularly, we investigate this issue from three perspectives. The rest is related to the prediction of performance problems. We propose the use of classiffcation techniques to derive performance models to assist the autonomic management of distributed applications. In this sense, different classes of models in statistical learning are assessed in both online and online learning scenarios. The second perspective refers to the reduction of false alarms, seeking the development of reliable mechanisms that are resilient to transient failures of the classifiers. This work proposes an algorithm to augment the predictive power of statistical learning techniques by combining them with statistical tests for trend detection. Finally, the third perspective is related to diagnosing the root causes of a performance problem. For this context, we also propose the use of statistical tests. The results presented in this thesis show that statistical approaches can contribute to the development of tools that are both effective, as well as effcient in characterizing the performance of middleware-based applications. Therefore, these approaches can contribute decisively to different perspectives of the problem.