Duda & Hart's "Pattern Classification" is one of the best introductions to machine learning IMO. It assumes very little in the way prerequisites, which is nice for first time exposure.
Hastie & Tibshirani's "Elements of Statistical Learning" can be a little intimidating without having been exposed to the ideas of the previous two texts. Afterwards, however, it is a gem.
It's been offered during the spring term for the past two years, so maybe Feb 2016 will see the next run.
https://www.edx.org/course/introduction-probability-science-...
In math, the model and axioms are sound (by definition) within the mathematician's world.
However, in order to make judgments about reality by using statistics, one has to come up with reasonable models and assumptions, otherwise the resulting deductions can be worthless. The leads to a lot of subjectivity and grey areas for debate that a mathematician may not be accustomed to.
In contrast, all of the science and engineering disciplines can make use of very interesting math. Not deep compared to research math, but used in a much more interesting way than in finance. E.g when you study the statistics of markets, you are just playing a game, and don't care that much about external reality per se. On the other hand if you study the statistics of DNA or gene expression, you are doing real science.
I think the best advice to a young person studying math is what was given to me at the age when I was doing the IMO (and interestingly, after I graduated by someone else): Don't neglect statistics.