Recent machine learning techniques such as deep learning and reinforcement learning were built on specific assumptions of the underlying data generation process.  Financial time series frequently do not satisfy these assumptions.  In this talk, we discuss the possible problems if these techniques are applied blindly.  The solutions to these problems are in general problem specific.  However, some of the pain can be alleviated by combining recent machine learning techniques with more classical statistical and econometrics insights.  We will discuss these probable solutions with examples.