Thin slicing

Not sure why this is just occurring to me now, sometimes it just takes a while to make a connection.

Malcolm Gladwell is one of my favorite authors. In his book Blink he looks at why experts can come to correct conclusions almost instantly when a far more methodical examination of the data by non-experts turns up nothing (or, worse, conclusions that are false). He refers to the phenomenon as "thin slicing", the ability to cut through all the noise and recognize just the information that matters. This is often subconscious. Ask a master chess player why a certain positional move is bad and you might get a response like "it's just not the right move for that position." They're right, and they know they are right, but the reason they know it isn't easily articulated. They've seen thousands of similar positions and simply "know" what works and what doesn't. Tactical moves, of course, do require computation even by masters, but they will do it much faster because they are quicker at discarding variations that are clearly going nowhere.

At yesterday's lecture in Data Mining, it was mentioned that all this work is really just boiling huge data sets down to reveal the actionable information. In other words, it's thin slicing by machine.

Here's the rub. While a renowned expert may be able to get away with an opinion supported only by their reputation, machines can't. Or, at least, they shouldn't. The actuaries in my company are not going to accept conclusions without supporting documentation, even if the results match what their gut is telling them. They can't. Regulators require that the means by which decisions are made are documented and conform to accepted practice. One such accepted practice is that results be stated not as point estimates but confidence intervals.

So, how do you state at thin-sliced opinion as a confidence interval? Well, if I already had that worked out, I'd already have my PhD, but I should be able to tell you in a couple years.