Data mining, machine learning and rival discount rates

A fool’s gold (or data) mine

Quants are building statistical toolkits to avoid the pitfalls of data mining

OK, computer: hurdles remain for machine learning in credit risk

Concerns over cost, applicability and oversight give pause to banks’ use of ML techniques in credit risk

Clearers diverge on SOFR swaps discounting

CME switches to new rate for clearing; rival LCH stays with Fed funds

COMMENTARY: Torturing the data

Eating too much cheese causes Americans to strangle themselves with their own bedsheets. Buying margarine makes the people of Maine leave their spouses. Every new Nicolas Cage film brings in its wake an epidemic of drowning in swimming pools. It must be true – the datasets correlate almost perfectly.

These and other spurious correlations (collected by the author and consultant Tyler Vigen) point to one of the biggest problems of big data – with a large enough universe of data to pick from, you can produce a convincing statistical story to explain pretty much any series of points. But, of course, that isn’t the same as a reliable story. Banning margarine would not lead to a sudden outbreak of marital harmony among the pine trees and lobsters. It’s unlikely that even The Wicker Man caused any Americans to fling themselves despairing into the nearest body of water.

Big-data partisans in the financial sector are coming to a very similar conclusion. Too many of those promising big data and machine learning projects in quantitative finance have repaid their backers only with fool’s gold; they have dredged through enormous data sets in search of a profitable discovery and found only a spurious correlation with no predictive power. Many more have only rediscovered an already-known factor.

Some of the most heated arguments centre around the role of human judgement. Many quants argue data mining without forming a hypothesis first is unscientific – that way lie swimming pools, Nicolas Cage and other irrational conclusions. And those cautious about the use of these technologies in credit risk modelling point out that regulators will be sceptical of models developed without human intervention. Their scepticism is justified – blind trust in a black box model is a very serious risk.

But there are arguments on the other side too: humans are biased and blinkered reasoners, and basing models on human intuition risks enshrining human bias in code, or missing unorthodox profit opportunities, argue machine learning supporters.

A more productive area for quants would be encouraging a more sophisticated understanding of probability and statistical robustness. Many academics are already arguing in this direction – including those pushing for more widespread use of Bayesian statistics. It will mean abandoning many expensively developed and dearly held beliefs; but that’s what science is all about.

STAT OF THE WEEK

Figures published by the Fed as part of this year’s Comprehensive Capital Analysis and Review show the 35 participating banks would suffer total losses of $578 billion in a severe recession – $98 billion more than last year’s result.

QUOTE OF THE WEEK

Boards have gone from turning up once a quarter for a prawn sandwich to being down in the weeds of what you do – Stephen Creese, Citi