A thread on the limits of using statistics for human decision making in real time.
The NYT election now-cast needle malfunctioned.
It reminds me of the 1998 LTCM crash that roiled financial markets.
The NYT election now-cast needle malfunctioned.
It reminds me of the 1998 LTCM crash that roiled financial markets.
Some very smart people (with Nobels) at LTCM failed to anticipate the inherent instability of their statistical model as they arbitraged on-the-run versus off-the-run treasuries etc.
Why? because the real world is not stationary in technical terms.
Why? because the real world is not stationary in technical terms.
In simpler terms, there is always the risk (even with "big data") that the statistical model on which we base our predictions is not exactly right - that it is "misspecified", and can change.
The "misspecification errors" can lead to surprises that are outside of the "confidence intervals" implied by the statistical models themselves.
The world is moving fast towards statistical models with AI and all (often for good reasons). But as we do so, it is important to maintain some skepticism for how far we can go. And most importantly, maintain appropriate safe guards that tell us what happens when models fail.
