- SURD Method (the algorithm discussed in the article): https://github.com/Computational-Turbulence-Group/SURD
I don’t understand: isn’t this stuff already dealt with, theoretically and practically, in statistics? Regression, propensity score, statistical analysis and similar stuffs? So what the MIT did, is just a computer program to score and estimate the cause-effect links or is there something fundamentally new?
Moving beyond statistical association to establish causal relations i non-trivial.
Their very first reference is Judea's Pearl's work who pioneered this domain.
Pearl, J. Causality: Models, Reasoning, and Inference (Cambridge University Press, 2000)
He also has a slightly more accessible book (The book of Why).
He argues that most of classical statistical analysis is but the first of three levels (association), with intervention (what if) and counterfactual analysis being the path to understanding.
To add to openrisk's comment, classical statistics doesn't formally define or attempt to tackle causality. Regressions and the like only describe associations between variables. For a statement regarding causality to hold, you need stronger evidence than just associativity.
Pearl has published a number of primers/papers/books with instructive examples that outline the limitations of classical statistics and applications of casual diagrams/graphs.
Can someone explain how this measures causality, rather than correlation?
Best bet is to read the paper (linked in a sibling comment). The introduction is very approachable and has a great introduction to causality/correlation/association and how they interact.
+1 to this - I think the authors did a wonderful job at explaining the fundamentals of causality in a self-contained way.
This is also useful to hedge funds.
sssssh. * wink *