Subnet Signal Predictability
Apr 29, 2026
Last updated
Apr 29, 2026
Bittensor subnet tokens trade on deterministic bonding curves with full on-chain visibility, removing the measurement noise that limits traditional microstructure research. Evaluating ten signals on 720 hours of SN45 data, we find three things. Directional and magnitude prediction require different signals: Capital Direction (r=0.23) and Buy/Sell pressure (r=0.16) lead for direction; Volatility Ratio (r=0.57) and VPIN (r=0.41) lead for magnitude. RSI's headline correlation (r=0.59) is inflated by a 0.96 return auto-correlation, a bonding-curve artifact, not genuine signal. VPIN, computed here without classification error for the first time, predicts both direction and magnitude. We also document Flow Persistence as a novel 24h reversal signal (r=-0.19).
Full report below:
Last updated