Lab · Moontrade · Parked
The one that did not work, measured properly
A currency-trading prototype that asked an unfashionable question: do lunar phase, geomagnetic activity and weather carry any signal a market has not already priced? We built it properly so the answer would be worth having. The answer is no.
What the numbers said
Predicting direction
Five-minute direction, from 319,336 rows of five-minute bars covering 2020 to 2025. Long calls were right 2.24% of the time, short calls 3.66%.
2.24% precisionPredicting volatility
Forecasting how much the price would move, from price history alone. An R² of 0.0078 means it explained almost none of it.
R² 0.0078The celestial signals
Adding lunar, geomagnetic and climate features scored R² 0.6614 against a baseline of 0.6812. They did not help. They made it slightly worse.
worse than baselineWhy this is on the site
Because a measured failure is worth more than a flattering story. The repository still contains a folder named for a profitable system, with a Sharpe ratio to go with it; the data behind it is a 0.54% return over 264 days from fourteen trades, in a backtest. Fourteen trades is a sample size, and this page treats it as one.
It only ever ran in paper trading. It was parked in November 2025, and the clear-out that day deleted the failed models rather than keeping them around to be rediscovered as promising.
What we kept
The method. A question worth asking, a baseline to beat, labelled data, and a result recorded whichever way it fell. That is the same discipline the Decision Bus now applies to AI work, where a model's claim is far harder to settle than a currency price and the temptation to accept a flattering number is correspondingly greater.