Print

Quants vs decimals

In early 2009, customers of AXA Rosenberg began complaining of ‘industry overexposure’ in their portfolios, according to an SEC order published on Thursday.

Fast forward two years and AXA — which was one of the pioneers of quant models for portfolio investments — has now been fined $242m by the US securities watchdog, for failing to disclose the discovery of a coding error in their investment model.

Axa Rosenberg neither admitted nor denied wrongdoing, the FT reports.

According to the SEC, AXA’s model consisted of three components: the Alpha model, Risk model, and Optimizer. The Alpha model would look at public firms based on their earnings and valuation, while the Risk model would identify ‘risk’ based on things like sector-specifics, sovereign risk or stock price risks (P/E ratios etc.).

The Optimizer, meanwhile, would bring the two other models together; balance them against each other, and recommend an optimal portfolio for the AXA client based on a benchmark chosen by the customer, such as the S&P 500 or whatever.

And it was in the Optimizer that AXA’s error took place. How bad was it exactly?

From the SEC order:

Starting in 2009, a BRRC employee began work as part of BRRC’s effort to implement a new version of the Risk Model. In June 2009, this employee noticed certain unexpected results when comparing the new Risk Model to the existing one that was rolled out in April 2007. He learned that the Optimizer was not reading the Risk Model’s assessment of common factor risks correctly because an error in the code caused a failure to perform the required scaling of information received from the Risk Model. Some Risk Model components sent information to the Optimizer in decimals while other components reported information in percentages; therefore the Optimizer had to convert the decimal information to percentages in order to effectively consider all the information on an equal footing. Because proper scaling did not occur, the Optimizer did not give the intended weight to common factor risks.

Whoops!

The error was discovered by an employee of Barr Rosenberg Research Center, which is the model’s developer and part of AXA, in June 2009. He or she duly discussed it with senior staff at AXA and Barr, the SEC says. However, an unnamed senior official told the employees to keep quiet, according to the commission.

It wasn’t until November 2009 that the error was actually fixed for all of AXA’s portfolios. And it was only in November that year, when an unnamed employee “felt compelled” to tell AXA’s CEO about the mistake that an internal investigation was launched, culminating in disclosure to the SEC in late March 2010.

In the meantime though, and according to the SEC’s estimates, some 608 of AXA’s 1,421 clients were affected by the error — generating almost $217m worth of losses.

Whoops, again!

Related links:
Number-crunchers crunched - Economist
Moody’s CPDO ratings error - FT Alphaville

Print