Discover and read the best of Twitter Threads about #hidethedecline
Most recents (1)
another sleight-of-hand in new PAGES2K paper. They say tht they use CPS (which had the advantage of using straightforward average of standardized series - no ex post sign change/flipping.) But in fine print, something different: their "CPS" weight proxies by correlation to target 
2/ mathematically, this is identical to Partial Least Squares regression of the target onto the proxies. (Stone and Brooks have interesting analysis of path from these coefficients to OLS coefficients in coefficient space.)
3/ if, as in this example, there are 200+ proxy series and a calibration period of 100 years or so, you're going to get a pretty good fit through PLS regression. D'oh.
