Learn · Tests
Test H12 · 2016–2023
Machine learning with astro features Beat 0% of shifted-calendar feature sets
Claim: Adding planetary features improves ML forecasts more than shifted-calendar features do.
How it was tested
Two standard models (ridge regression and LightGBM, fixed settings, no tuning) forecast each stock's next-week return and volatility from market features, with and without astrological features, and with five sets of shifted astrological features taken from shifted calendars. The question is whether the real planets help more than the shifted ones. Full details and charts are on the Models page.
Result
H12: Adding planetary features improves ML forecasts more than shifted-calendar features do
Beat 0% of shifted-calendar feature setsVerdict: the real sky beat 0% of 2,000 shifted calendars.
real astro features ranked against 5 placebo sets in 4 model/target pairs; best pair lgbm return IC (rank p 0.50), Bonferroni over the 4
Development data 2016–2023; the 2024–26 data stays sealed until it is opened once. Effects are averages over the period, before any trading costs unless stated. How we test · all results