Learn · Models
Test H12 · machine learning
Do planets help a forecasting model? Beat 0% of shifted-calendar feature sets
If astrology carried information about markets, a machine-learning model given planetary features should forecast NSE stocks better than the same model without them, and better than one given shifted-calendar features.
The data
One row per stock every fifth trading day, 2016–2023, for every NSE stock with a year of history. Each model forecasts two things for the following five trading days: the return (scored by rank correlation with what happened, the IC) and the volatility (scored by QLIKE, a standard loss for variance forecasts).
Three kinds of features
- Market baseline: momentum over 20, 60 and 250 days, volatility over 20 and 252 days, the Nifty's 20-day return and India VIX.
- Astro: the sign of every planet, retrograde flags for Mercury, Venus, Mars, Jupiter and Saturn, the Moon's nakshatra, the tithi, whether an eclipse, Amavasya or Purnima falls in the coming five sessions, and the stock's own Gochar: Jupiter's and Saturn's houses from its listing Moon and whether the H11 rule favours it. All are known in advance.
- Shifted astro (placebo): exactly the same astro features taken from five calendars shifted by at least a year. They look and behave like astrology but point at the wrong dates.
Models and validation
Ridge regression and LightGBM, each with one fixed configuration chosen before any result, so there was no tuning to flatter a feature set. Walk-forward validation: train on two years (504 trading days), test on the next six months (126), roll forward, with a 10-day embargo and labels purged so no training row overlaps a test week. That gives 288 test dates.
Results
| Target | Model | Market only | + real astro | + shifted astro (range) | Shifted sets as good or better | |
|---|---|---|---|---|---|---|
| return IC | ridge | 0.0536 | 0.0591 | 0.0588 – 0.0628 | 4 of 5 | higher is better |
| return IC | lgbm | 0.0356 | 0.0457 | 0.0337 – 0.0516 | 2 of 5 | higher is better |
| volatility QLIKE | ridge | 0.9220 | 1.1407 | 0.9786 – 1.1437 | 4 of 5 | lower is better |
| volatility QLIKE | lgbm | 0.8654 | 0.8679 | 0.8216 – 0.8568 | 5 of 5 | lower is better |
Adding the real planets raised the return IC a little, but adding shifted planets raised it as much or more in most runs, so the gain is just what extra features of this shape give any flexible model, not information in the planets. For volatility, the real astro features made the ridge forecasts clearly worse and left LightGBM unchanged, while several shifted sets did better.
The model leaned on shifted planets about as much as on real ones: around half of its attribution for returns and a quarter for volatility either way. A model will always find some use for extra inputs; the placebo sets show that this use is not astrological.
Verdict
The real astro features never beat the shifted ones by a margin chance could not explain. The single p-value for H12 (ranking real against shifted-calendar features in all four model and target pairs, with a Bonferroni correction) is 1.00: the real sky beat 0% of 2,000 shifted calendars.
Development data only; the 2024–26 holdout stays sealed. H12 summary · How we test · All results