The Ferguson lynch mob… strike.

My day job is being a quant researcher and programmer for hedge funds writing predictive models that forecast stock price movements. In the 20yrs I’ve been doing it I’ve managed 25 people and hired about 50 working in some of the most model driven funds in the world.

Ferguson is a typical academic mathematician with a model, turned coder. I’ve hired, managed, worked with, and for, 10s of people like him. As much as they might be “crap” coders, it doesn’t make their maths wrong.

This lynch mob on lockdown sceptics miss the point. The Imperial College research department is not Google. It does not, have the money to justify, or the speed of release cycle to need, lots of people working on the same code in a commerical capacity to a deadline, with the need to make sure that people who don’t understand the “clever” maths don’t break its functionality when pushing to the deadline of a release.

Unit tests, regression tests, continuous integration — all automated — are vital for any commercial enterprise to stay in business. I’ve driven the managerial realisation and implementation of this as a mathematician and coder in all the businesses I’ve worked for. The author is right, these things matter, stable results are much easier to work with.

… but, none of this makes his answer wrong.

All I say is that regardless of the github revision control mining you might do, you can never know the manual regression testing processes that Ferguson would employ, and force his “employees” (fellow researchers) to follow to make sure that all the very real short falls in the code they highlight were mitigated to give him the consistent answers he had spent a lot of time convincing himself were right. Don’t assume what he did, or did not do, on top of the code, interview him, don’t just guess he did nothing.

Being able to see this better from all sides than this Google employee who is likely far more of a coder than a mathematician, and certainly does not talk with a managerial understanding — you start by looking at their answer, not their process. Here is his actual paper. If the prediction is genuinely off, then their review has merit, if not, it is just an indication he needed the help Microsoft seem to have been giving him since to move his code to something more robust.

Is it just a simple SIR model?

To all the critics of this just being a simple SIR model, look at page 4 and consider the description of the model behaviour. This model may take an input background r0, but it is heavily varying the magnitude of that r0 depending on actual UK demographic data. As in modelling the spread in different parts of the country, and in different settings, as faster or slower depending on the demographic descriptions of the actual population.

Did he get his predictions really wrong?

Everyone seems to share this image of his predictions vs reality in Sweden as some sort of proof the man was incompetent. If you actually look at the prediction plots in his paper the yellow area would correspond to household case isolation, maybe with closing schools, bear in mind Sweden has shut universities and 6th form colleges. This is not a suppression strategy of stay at home orders. As I consider in this article you can’t say Sweden is running a “moderate” mitigation strategy when 65% of people are working from home. They are in a voluntary lockdown, and any attempt to describe it as a utopia of freedom and liberty is blinkered or making manipulative attempts to influence policy, certainly ignoring reality.

What is worse is if you actually look at the paper his prediction for a suppression strategy were 75% work from home is the green line in this top plot (in the blue shaded area), that is where we are right now according to his predictions. If you’re having trouble seeing it, it is the line that says about 1-5 deaths not 250, as in, trundling right along the bottom. Which looks very much like the red area in the first plot. Essentially the assassination plot is the data equivalent of a “misquote”.

Far more conclusively, if you look on page 13 of his Mar16 paper, he provides numeric estimates, including the famous 550k, but there are 15 other estimates too. Under “Total Deaths > 2.6 > On Trigger 400 > PC_CI_HQ_SD” (code meanings page 6) he was predicting 48k deaths by riding the health service up to ICU capacity in little wavelets (given he thought the ICU ceiling as lower) using a stay-at-home suppression strategy. The numbers of ICU cases treated in total looking very similar to those we have seen treated during this first wave.

It is cancel-culture-mania to use the “do nothing” 550k estimate to destroy a man when he clearly in the same paper predicts 48k deaths with the very strategy the government actually followed.

Now you can ride his technical skills all you want, but it doesn’t make his answer wrong. What we do with hires who are bright, good at numbers, but crap at coding was pair them with great coders who can understand (and validate) the maths, and tell them to learn. What comes out of the pair is code that has unit tests, can fit into our automated regression testing and continuous integration platform for releases… but, most importantly, we don’t end up just dismissing the value of the maths because the person wasn’t good at everything, we give what they are good as a chance to sing.

The lynch mob above… will lead to the exact opposite.

Now I totally agree with all the other complaints, you can’t advise a policy, have the privilege of that affording you a test, and then use that to allow you to flout you’re own policy advice. No matter how hot the girl is. He’s a pillock, who can’t code to the production standards of a modern tech firm… but, it still doesn’t make his maths wrong, or give us the right to decimate what’s left of his reputation through misquote.