I think this is a bit dangerous. Attempting to ascertain sentimental correlations and apply them to huge financial decisions may work in certain contexts and you could definitely train a model with 99% training accuracy on this task, but future situations that are dependent on complex human action can never be adequately represented by a numerical parameterization and a finite state machine. If the model is not large enough, we will not learn all the possible combinations of interactions. If it is too large, then we only learn the context of our training dataset. That being said, you could do both and get good results during training. Personally, I do not have extensive NLP experience or Bayesian experience in production, but their fundamentals suggest that they would learn this type of model well independently or in conjunction. Naïve Bayes is good for state estimation-based decision making, and NLP can be used to model language and extract sentiment. However, these models depend completely on the input dataset that one utilizes, and the chosen labels (if using a supervised method) that are often subjective. Using data from the internet is also dangerous because it is next to impossible to have humans annotate every piece of training data without spending a large amount of money, and learning from problematic input data can lead to problematic situations.