Peter Hodge’s article in last weekend’s The Weekend Australian, “Critical thought beats blind faith in models“ brings up the possibility of what is effectively a Nuremburg defence: “It wasn’t our fault, the statistical model told us to do it”.

Although the article goes on to correctly define the boundaries of modelling and its applicability, visiting the argument from the position of a decision analyst brings further insight. Peter Hodge writes that mathematics is an artificial construct and so should be treated with scepticism. I would argue mathematics is fine – it’s what some do with it that is weak.

We obviously use statistical modelling a lot, and spend a lot of time ensuring everyone involved knows what is going on and where the rules will break down. Why do we spend such an effort? Because if we didn’t we’d never solve a problem.

You can’t replace decision making with a machine. And I would argue that trying to is foolish, and will ultimately fail. I am aware of Artificial Intelligence, various intelligent learning and genetic algorithms. I am a fan of many. But decision making is distinctly biological.

There are two facets to this:

• Firstly, there are too may variables, including the background of the person ultimately taking responsibility for the outcome. Models, particularly statistical ones, only have knowledge of what they’ve been allowed to see. They are necessarily reactive and based in history. A human absorbs from a variety of stimuli, and has an inherent junk filter that allows synthesis of all these inputs in ways a model simply can’t replicate.
• Secondly, we have the commercial influence. People working in BI (Business Intelligence) will tell you that selling a black box forecaster or any other form of analytics is an uphill battle. And it should be. To presume that a computer can manage a system cheapens the skills of those managers we employ to perform just such a task.

Every manager I’ve ever worked with as an analyst has wanted the facts, the interpretation, and the reasoning behind it all, so they could make their decision. I’ve never been allowed to make the decision for them, and neither have any of my analyses.

And that’s the way it ought to stay. Analytics and modelling are decision support tools, for aiding a decision maker in their task. Having said that, if people were relying on the black boxes to make their decisions, I’m not surprised it went belly up.

The Black Swans (after Nassim Taleb’s book) that people speak of when bemoaning the failure of analytical modelling arise not due to a short coming in the modelling, but one in the interpretation. The models are being asked what the likelihood of a “black swan” arising is. And they answer truthfully. The question most managers ought to care about is “what are the costs associated with the those likelihoods?” That matrix is far more meaningful to decision making. Risk bears a cost – it is shortsighted to measure a risk without the associated cost.