Mifid, machine learning and swaps compression

The week on Risk.net, March 10–16, 2017

MIFID alarms the rest of the world

MACHINE LEARNING attracts quant and sales attention



COMMENTARY: The problems with machine learning

The banking industry has not always been the best at adopting new technology. For many years, it was the last redoubt for such items of industrial archaeology as fax machines and telex terminals; even today, the internal networks of many major banks are tangled nightmares of obsolete software and incompatible standards. The reasons for this vary – from regulation and staff turnover to a widespread belief in secrecy and proprietary software over open standards – but in one area, progress is speeding up.

Machine learning, after many false starts under various names, is now flavour of the year in financial technology, promising better regulation, smarter risk management and sharper sales tactics, by providing financial institutions with a tool that can process the immense amounts of information being collected under 'big data' schemes. Automated systems that were intended to detect suspicious transactions and prevent fraud or money laundering can also be used to profile retail customers for marketing purposes. Now derivatives sales teams are using the same approaches to guide their own sales and client relations efforts – either backing up human salespeople, or replacing them altogether.

But risk managers need to be more cautious, delegates at the Quant Europe conference heard this week: training machine-learning software on historical records risks judging the future by the standards of the past, which might make sense in medicine but not in an evolving field like finance, Riccardo Rebonato warned. But, as other delegates noted, in many areas, the volumes of data involved may leave banks no choice but to rely on machine-learning approaches to an ever greater degree. “[Machine learning] is a way for us to find a solution without writing equations, without building the model – we just let our data tell us what the optimal solution is,” one commented.

The crucial consequence of this is that modellers and risk managers must be able to defend the results. If the process is obscured within a black-box machine-learning process such as a neural net, the output must still be comprehensible – to regulators, and to the senior managers who will rely on it for decisions on new products, risk appetite and business strategy generally.



Mortgage service provider Ocwen Financial paid $225 million to re-enter the California mortgage servicing business, settling charges that it had sent time-sensitive letters to homeowners days after the dates stated on the documents – so-called letter dating – and sent borrowers inaccurate information on notices of default. This and other offences led to its being banned from the mortgage market by the state’s department of business oversight in 2015, as well as a series of other fines.


I don’t know how long it takes for CVA to blow up in our faces in the event of a big bank default, but frankly the numbers we get out of this exercise are based on wishful thinking; I’m not sure what they really mean” –  Rama Cont, professor of mathematics and chair of mathematical finance at Imperial College London.

  • LinkedIn  
  • Save this article
  • Print this page  

Only users who have a paid subscription or are part of a corporate subscription are able to print or copy content.

To access these options, along with all other subscription benefits, please contact [email protected] or view our subscription options here: http://subscriptions.risk.net/subscribe

You are currently unable to copy this content. Please contact [email protected] to find out more.

You need to sign in to use this feature. If you don’t have a Risk.net account, please register for a trial.

Sign in
You are currently on corporate access.

To use this feature you will need an individual account. If you have one already please sign in.

Sign in.

Alternatively you can request an individual account here: