08 July 2016

Event Driven Data

It's become a standing meme in these endeavors: events drive data, not the other way round. In some realms, notably the physical sciences, the rules of the game are God's and mostly known, so there are fewer events of the asteriod mass extinction type. Had some very future generation of humans been around at the time, one which had orders of magnitude of more computing power now available, detecting and diverting should such a bad boy come cruising by, is likely. In such realms, time-series and quant modeling generally can prove useful predictively. The National Weather Service, with its massive supercomputers, is one example.

When it comes to financial engineering, and other human driven venues, not so much. The catastrophes over the last four or five decades have gone unpredicted just because the models assumed stable context, while the bad actors were behind the curtain pulling levers and pushing buttons that they shouldn't have. They even went so far as to re-wire the controls. Didn't tell anyone, pocketed the moolah, and, on the whole, got away without a scratch.

So, where, if any, do the two venues overlap? One I'd offer up is biotech/pharma. Most of pharma these days is re-patenting compounds and patenting reformulations of mixtures old drugs. Not much to that.

But there is time and money being spent on pushing back the frontier. Oncology gets the prize for greatest effort. This week one of those pushs, Juno, blew up, big time. Adam Feuerstein's Twitter feed has some interesting postings. Since it's a continuing feed, depending on how soon you go explore, you may have to go looking. The thing about really new drug development is that the scientists are never really sure about how a drug works: the MoA or mechanism of action. Thus clinical trials are used to test, empirically, what really happens. If the understanding of the science is spot-on, clear sailing. On the other hand, what looks like a winner in early, small trials more often than not proves no better than saline in larger ones.

The big dipper is killing patients, and that's what happened to Juno as discussed in AF's Twitter feed. Doesn't happen often, but when it does the specter of guilt rises. Was the fatal effect predicted? Was it predictable? Did it result from previously unknown interactions?

Events drive the data, so Juno's (and others pursuing the same type of compound) share price has taken a crash that wasn't predicted by the financial engineers.

My two faves from the feed, so far:
1-
I love the "grade 5 toxicity" euphemism. I'm hearing more and more about CNS AEs for CAR-Ts. A bit scary sounding.
-- Ozgur Ogut/3 May 2015

2-
Essentially, the heavy Flu/Cy may well be driving the bulk of the apparent CAR-T efficacy
-- Vikram/29 March 2016

In sum, then: even in science, which obey's God's rules, predicting the future based on past data is fraught with danger.

No comments: