05 January 2016

A Matrix of Hessians

If you're interested in being a corporate quant, drop off your morals at the door. It's no secret that I regularly warn wannabe (retail) financial quants (who think they'll get rich from their quant skills) that they're playing against the House, who's dealing from a stacked deck. Motive and incentive drive the data, not the other way round, modulo some consumer advert testing, which is boring as all hell. It's not often that malfeasance in corporate quant is outed so dramatically as now.

Today's news of Takata's evil is stunning, even to one as deeply cynical as I am. If the data doesn't "prove" the current corporate profit scheme is within bounds; well just fudge the data.
"Happy Manipulating!!!" a Takata airbag engineer, Bob Schubert, wrote in one email dated July 6, 2006, in a reference to results of airbag tests. In another, he wrote of changing the colors or lines in a graphic "to divert attention" from the test results and "to try to dress it up."

That last line is of particular importance, since it displays the underhandedness of some quants serving the hand that feeds him. The pie chart is widely despised amongst the academic quant community, but along with Excel, is widely used in Business. Why? Because a narrative, either oral or text, can be developed to spin the underlying data the spinner's desired direction, since the pie chart is ambiguous from the git go.

It gets worser. The Takata quants went so far as to camoflage the meaning of distributions.
"I showed all the data together, which helped disguise the bimodal distribution," Mr. Schubert wrote. "Nothing wrong with that. All the data is there. Every piece," he added. But then he suggested using "thick and thin lines to try and dress it up, or changing colors to divert attention."

One might wonder what Hadley Wickham would have to say about that. Or William Cleveland. Or, God protect them, Edward Tufte.

A picture may be worth a thousand words, but I see many dead people in the picture. How many is stat sig?

As a GM engineer admitted:
a bimodal distribution showed the parts being tested were not consistent -- generally a requirement for meeting quality standards for automotive safety products.

For those, esp. the financial and micro quants, who think that data is judgment free, guess again.

No comments: