Now that I'm back from my 3-week blackout I can comment on this.

The dirty secret of bankers is that they are bad at science and maths, and do not understand that a model, however sophisticated, cannot provide output of a quality better than the input.

Even people who are bad at science and mats are bad at modelling, so nothing new there. And GIGO (garbage-in-garbage-out) is a cautionary concept that still tells you little about how to get good data

Lots of data does not mean better data

To think otherwise is to engage in what people call "datamining", meaning sifting data looking for correlations which may well turn up to be spurious (correlation is not causation - another catchphrase which is true but hard to put in practice productively)

what makes data "good" is qualitative analysis, i.e. risk assessment by bankers doing their job instead of relying on fancy models.

To put a contrarian/charitable spin on this, the hope seems to have been that the knowledge of bankers could have been translated into models giving rise to "expert systems" which would be able to do a banker's risk-assessment job. In other words, the philosophy of serious quantitative investing seems to have been to avoid acting on hunches: if you have an idea or a judgement you should be able to elucidate the reasons why and make a model about it. One of the fund managers I interviewed with nearly 3 years ago prided himself on a model of "no human intervention" precisely on this theory that if you have an intuition you should be able to make it a model.

They do use extremely smart mathematicians to play around with data, but these guys' jobs are not that of bankers.

My impression is that people are being hired on the basis of numeracy or IQ to the exclusion of funancial knowledge or experience, expected to be able to swim by themselves when thrown into the water, with little to no in-house training (all on-the-job, basically) and in many cases given very little, if any, exposure to "real" markets. Sometimes computer programmers, not portfolio managers or risk managers, or traders are in charge.

I mean, "models typically predict the future on the basis of past data"

Models have parameters. Parameters need to be estimated from past data.

- anybody that has ever bought any financial instrument gets told (or sees written in small print) right from the start that the past is no indicator of the future...

What the small print says is that past performance is not an indicator of future performance. This is intended to preven prople from marketing on the basis of a promise to obtain the same returns in the future as in the past. The disclaimer is always seen alongside displays of past performance intended as advertising, and is therefore being ignored both by managers (advertisers) and investors (consumers).

Thus, models work until they don't. LTCM's lesson has visibly not been learnt.

Hey, LTCM's predictive models are claimed to have worked in the end, what LTCM didn't have is proper risk management as they crashed because they didn't have deep enough pockets to weather one particularly bad adverse movement, and this is because they got greedy and reduced their outside investor base hoping to keep all the return to themselves, but reducing their capital base prevented them from meeting their margin calls...

By the way, a common accusation levelled against hedge fund managers is that if their models really work they keep them to themselves and only seek outside investors when returns are no longer that good, and make money from the (often outrageous) fees they charge. There is some truth to that.

Can the last politician to go out the revolving door please turn the lights off?