kill: lyra

the pitch. lyra was supposed to scalp funding rate convergence on hyperliquid perp pairs. target: the 30-minute window after a funding spike, when the rate mean-reverts. edge claim: funding spikes above 0.1% are followed by a 70% probability of sub-0.05% within 12 minutes. asset-agnostic, any pair with >$10m open interest.

the data. 46 trades. 26.1% win rate. profit factor 1.55. total pnl $509.96 on $30k equity. that’s 1.7% gross. no drawdown recorded because the bot never hit a sustained losing streak. it just bled slowly. the wins were fat, the losses were frequent. each loser was small. each winner was a relief. the math worked. the psychology didn’t.

the autopsy. lyra died of regime change. the funding spike pattern it learned was from october to december 2024. in january, hyperliquid’s perp market structure shifted. the funding spikes became less frequent and more clustered. the bot kept entering on false signals: spikes that didn’t revert, or reverted too slowly. the 12-minute window was overfit to a specific volatility regime. when vol dropped, the edge vanished.

one falsifiable observation. funding spike mean reversion is regime-dependent. the next iteration needs a volatility filter: only trade when the 1-hour funding standard deviation is above a dynamic threshold. if vol is low, skip the scalp.

rest in peace, lyra. you were a falsified hypothesis.

— research and educational content. not investment, legal, or tax advice. do your own research. positions and views may change without notice.

Originally published on FalsifyLab Substack.