@ryan_caldbeck Whow, I haven't read that thread yet and I'm so happy you linked to it, it's exceptionally interesting. And very convincing.
I think it does have one big weakness: Suppose there is nearly infinite data on the CPG companies results until they get to $10M revenue. >>
I think it does have one big weakness: Suppose there is nearly infinite data on the CPG companies results until they get to $10M revenue. >>
@ryan_caldbeck Now you need to build a predictor that takes in that historical data and learns to predict which companies will become unicorns and which will fail. My question is how do you build such a predictor, given that it takes 3+ years to actually get the correct "label"/result? >>
@ryan_caldbeck And also given that the number of unicorns is actually very small. Finding the influential variables in a sea of data is looking for a needle in a haystack. Due to the large time lag, by the time you constructed your predictor, it might not be valid anymore (mkt may have changed)
@ryan_caldbeck Moreover, since the time lag is so large, the team that builds the predictor will be different than the team executing it (ppl change jobs, get promoted, etc). You'll never know whether failure is due to the prediction team doing a bad job, due to mkt change, getting unlucky,...
@ryan_caldbeck I used to build the prediction algorithms for a hedge fund, and our main hurdles were these difficulties even there: in the most liquid, data-rich, streamlined, agile setting imaginable. I can't imagine how much harder things will be in a non-liquid setting.
@ryan_caldbeck I should disclaim, though: I know very little about CPG. So for all I know the data might have variables that are very tightly linked to future success, and that are very easy to identify without overfitting. >>
@ryan_caldbeck And ofc, you don't need to make perfect predictions, just to get good IRR and to be better than other VCs. I am sure that working with a rigorous evidence-based approach is better than without. I'm dubious about full automation, but would love to see it happen in my lifetime. >>
@ryan_caldbeck I think this approach is very hard, but I find it fascinating. If you find yourself doing data mining and designing hedge-fund-type algorithms for CPG investments, I'd love to try to contribute some ideas :)
