Anyways, I finally got around to finishing the wine project I've been dabbling in: reclassifying left-bank Bordeaux using scores from Cellartracker, and the results are interesting if you enjoy both simple linear regressions and spending your money on dumb shit like wine. Thread.
Caveats apply blah blah blah, this analysis only works for liquid markets with lots of producers in a small area, CT scores have variance, stylistically wines differ greatly even within appellations, not investment advice, but this is a decent first-order approximation AFAICT.
Data spanned 60 left-bank (i.e. cab-heavy) chateaus, considering only their grand vins, from 1990-2016. Scores were predicted from a three-factor log model of vintage, chateau, and improvement over time: PredScore = 100* (1- exp[-vtg*name-improv*year]), with an R^2 of 76%.
Results (which I've partly posted before) are in line with what's expected: first growths are the best and most improved; 00/05/09/10/15/16 are the top vintages; global warming and improvement in technique means that even mediocre vintages today would be excellent 30 years ago. 
Unsupervised clustering selects the following six classes, corresponding to 1-5 plus a declassified rump. I did not select any of the cutoffs, it was purely algorithmic- the data above appears to have a clear 1.5th growth of Mouton, MHB, and LLC between 1st and 2nd:
Using Liv-Ex price data from their 2011 reclassification, we can use this to identify value picks and overpriced producers, since the chateau scores correlate extremely well to case prices (R^2 of 89% for a log-linear regression)
The biggest outliers are basically what you'd expect: S-M is the top insider value buy by a mile, followed by LLC and Montrose; while any Rothschild-adjacent winery is bid up way above its innate quality.
That list of value picks spans a wide price range: recent Sociando-Mallet is about $50/bottle, Domaine du Chevalier about $100, Leoville Las Cases about $250, and Haut-Brion about $600. Value vintages include 06, 08, 12, and 14, which are all very good albeit on the younger side.
Anyways, just a fun toy model I've been playing around with to make my buying decisions, I might write this up in a longer-form blog post on Medium if there's interest.
Minor correction; 100*(1-exp[-vtg-name-improv*year]. As an example, Latour has a Château score (name) of 1.37, 2005 a vintage score of 1.82, and an average improvement of 0.03, so it’s predicted score is 100*(1-exp(-1.37-1.82-0.03*(2005-2003)))=96.1, vs 96.9 on CT.
Gotta eat my own cooking on this now that it’s blown up; this is lovely and eminently drinkable wine, lots of fresh red fruit but still recognizably Bordeaux. I do prefer the austerity of a Sociando, but with spicy pork dumplings this works well.
