macrocephalopod

Looks like you want to use this for Kelly sizing. In practice, your problem there is mu. Your return estimate will be a ton worse than even the very dumbest vol estimator

Well, this gets into a few other aspects I'm working on / been thinking about for years. There's quite a bit of evidence that S&P 500 *long-term* (10-yr) geometric returns are predictable based on valuation metrics. Can be modeled as Ornstein-Uhlenbeck process.

Don't disagree on the effect, generally. But there's probably a lot less evidence than you might think. Think about how far you're looking forward and back with feature and target calc windows there. How many independent observations do you actually have?

Yes, totally agree. We seem to think a lot alike! Are you familiar with John Hussman? He blocked me on twitter because I talked too much about that issue. He even wrote a blog post referring to me and this issue as "zombie troll bait from hell". :-)

Hahaha. Badge of honor.https://twitter.com/therobotjames/status/1372384046349623301?s=21 …

Amazing!!! You're the first person (besides myself) I've seen who talks about that "trails" issue! Bugs the crap out of me that almost nobody notices how strange those scatterplots look.

AQR (I think?) had a paper basically debunking the cape ratio for market timing by pointing out that using overlapping 10Y returns on the lhs of a regression and trying to adjust t-stats with newey-west is hopelessly optimistic

Instead you should use non-overlapping 1M returns on the lhs but smooth the regressors with a 10Y moving average, which gives the exact same regression coefficients but now you have non overlapping samples so you can do a normal t-test