Applied Finance and The Third Culture

Abstract

In his 2001 paper, Leo Breiman describes two diametrically opposed cultures in data science: the model driven statistician and the algorithm driven data scientist. His argument, that an algorithmic approach is often preferable to an iterative model-driven approach is unlikely to resonate with many financial professionals for whom theory is king, and simplicity is a virtue. This third culture of applied finance is as far removed from either culture described in the original paper.

We propose a horse race between these three cultures: applied finance and an iterative search of simple heuristic models, traditional econometricians armed with classical regression techniques and data science with a mature approach nonlinear algorithmic exploration. We compare several approaches to modeling market volatility and downside risk and explore the role that modeling discipline has on the stability of the outcome. Our tests are performed both on historical equity market and currency market returns, and on a return-mimicking data generating process, from which we can more thoroughly test the stability of each approach.

We find that each culture has its own idiosyncrasies and limitations, which need to be carefully understood as practitioners start to experiment with tools from other disciplines. The necessity of algorithmic approaches to harness the burgeoning opportunities of alternative data may be tempered by the limited work that has been done to bridge the knowledge and techniques of the three cultures.

Keywords: Machine Learning, Econometrics, Applied Finance, Investment Process, Volatility Modeling, GARCH, Elastic Net Blending, Research Culture

JEL Classification: C50, G11