30 July 2014

Bayes Rum and Coke

Markus Gesmann provides the most lucid, yet succinct, presentation of Bayes I can recall seeing. Doesn't convince me that Bayes has any justification in data analysis or data science or stats. But that's just me I guess. The data are, and that's the end of it. By the bye: his example is far more concrete than is typically used.

While I prattle on that policy trumps data all the time, Bayes provides a mechanism of injecting policy sub-rosa. Not my cup of tea. Look to incentives/policy if you want to predict the future. It was policy, not data, which generated The Great Recession, and stats folks (well, 99.44% of them anyway) who blindly believed that "yesterday looked pretty much like the day before, today looks like yesterday, so tomorrow will look like today". The inherent folly of MCMC. In Nature, where God imposes immutable rules of behavior, that's reasonable. In human Nature, where some humans get to change the rules to suit themselves whenever they feel like it, the resulting (historical) data is always suspect. That data stream, after all, wiggles around in response not to just perturbations of Mother Nature, but to the whims of human rule makers. There's not the slightest reason to assert that this combinations of influence will continue as-is in the future.

By coincidence, the Gesman piece references a Charpentier presentation, and I'm considering musing on a recent posting by him. One that's not overtly Bayesian, as it happens. We'll see.

No comments: