Analytic methods based on modeling to transform data sets or data driven methods where models are extracted from data sets? Mathematics or machine learning?
It is the AND effect that will matter - computational mathematics AND machine learning - in quantitative finance (like in many other fields).
But it will take a little more time, to make this happen. We all know the reasons that start with different scientific cultures and end in overhyping ostensible new methodologies by journalists, ... - the exabyte age, big data, data science ... that is countered by the physicists, mathematicians, ...: only soft methods.
In Finance: Extrapolate the Presence or Model the Future?
Predictive modeling is the task to create models that forecast. Simplifying, it seems the sell side is seeking extrapolation, and the buy side and risk management want to model the future?
Simplifying even more one could say extrapolating the presence (derivative pricing) applies to most sophisticated symbolic and numerical math methods to solve the stochastic models adequately and robust. Data come into play as input for the calibration engines. The objective is to get "deterministic" answers.
In risk management data play a diversified roll as input and output. It is especially expected that output data tell a story.
Models need to be Understandable and Computational
Independent on how you create a model, by thinking or automatically from data, it shall tell you how the transformation works and it shall be computational and calculate results. With dada driven methods this is not always the case. Crisp rule systems are understandable, but hardly computational, neural nets are computational but not so easy to interpret. Fuzzy rules are both.
All type of equations are both. In general mathematics represents a language with an operational semantic. The model speaks and its solvers are responsible for the quality of the i/o relation it describes.
Combine Analytic and Data driven Methods Intelligently
If you know nothing you can apply machine learning to get a model quickly. It might say you much about dependencies, but it can approximate reality only by the sample data you input. Generalization is difficult.
If you have a model that is not very accurate, you can apply machine learning to close the gap between model results and reality.
If your models has parameters you can apply machine learning to optimize them adaptively to real data.
It is quite obvious that the last approach is the most promising.
Future Job Descriptions of Quants?
Dan Tudball, in Wilmott Magazine Jan-13, says: quant finance practitioners in ten years will need to act at the convergence between both trends scientific problem solving and opportunities that Big Data present .... and to keep pace with the technology development.
They need to create good knowledge of data, good knowledge of financial theory, a pragmatic knowledge of market behaviors, but still multi-methods enabling the creation of computational knowledge systems.
But the Challenge is Already Here
CVA/DVA bilateral introduces much more complexity to the valuation space, relevant to trading, risk management and accounting that hybrid skills and approaches will be indispensable. Hard AND Soft.