Research

We audit our own systems and publish what we find — including the parts that are unflattering — with the method and the raw data attached. Every number traces to a source quote or an exchange candle; datasets are CC BY 4.0, so cite freely with a link.

August 2026

1,581 crypto predictions from YouTube, scored against the market

Bullish calls — 84% of everything said — were made at worse moments than random dates on the same assets. The one stable split: people with positions scored at baseline; people with theses scored reliably below it.

1,581verified calls
−9.8ppbullish edge vs random, 30d
0 / 100>3x targets reached
Read the study
August 2026

Anatomy of a memecoin pump: early wallets vs chasers

The same tokens paid wallets copied ≤1h behind the earliest buyers a median +62% in 24 hours — and cost buyers of the pump’s close a median −18%. The side of the trade is decided by when you act, not what you pick.

4,769early entries measured
−18.2%chaser median, 24h
9.9 daysmedian lead time
Read the study

The live experiment: our own predictions, scored in public

The same scoring rules these studies apply to others run continuously against our own AI: every directional call is written to an append-only log before the outcome is known, then scored at 24h, 7d and 30d. Hits and misses both stay on the board.

See the track record