strategy deep-dive: lyra
alpha micro-digest
thesis
lyra was supposed to be the boring one. a gold-trading bot that just clipped small, steady edges from funding-rate oscillations on hyperliquid. no leverage. no stacking. no composability tricks. just one asset, one signal, one position at a time. the kind of thing that should grind out a profit factor of one point five and never wake you up at 3am.
it didn't work.
after sixty eight trades, lyra sits at a twenty six point five percent win rate. that is not a typo. the bot is profitable, barely, because the wins are large relative to the losses. total pnl is nine hundred six dollars on thirty thousand equity. that is three percent return over an undisclosed period. the equity curve looks like a flatline interrupted by occasional spikes. this is not what i signed up for.
mechanism
lyra trades gold on hyperliquid. it watches the perpetual funding rate and enters when the rate crosses a threshold that historically preceded mean reversion. the idea is simple: when funding gets too negative, shorts are paying longs. go long. when funding gets too positive, longs are paying shorts. go short. collect the funding while waiting for price to revert.
the bot uses a single timeframe. one entry signal. one exit condition. no stacking, no scaling in, no dynamic take-profits. the exit is a combination of a time stop and a funding-rate normalization trigger. whichever comes first.
position sizing was originally based on atr. that got scrapped after backtesting showed the equity curve was too sensitive to volatility regime changes. the current version uses equity-relative sizing. the backtest profit factor on the full window is two point one five, with a fold rate of ninety two percent. those are paper numbers. the live bot is running the same logic.
the live bot is not hitting those numbers.
the numbers
equity: thirty thousand dollars total trades: sixty eight win rate: twenty six point five percent profit factor: one point six four total pnl: nine hundred six dollars max drawdown, sharpe, monthly return: not available
the profit factor is respectable. the win rate is abysmal. a bot that loses three out of four trades is psychologically difficult to run, even when the math works. the problem is the math is working on paper but the live results are a shadow of the backtest.
what could be overfit
the most likely culprit is the funding-rate threshold itself. the backtest optimized for a specific level that captured mean reversion in gold funding during a period of relatively stable macro conditions. gold funding on hyperliquid has been erratic lately. the threshold that worked in backtest is producing too many false entries. the bot is taking trades that meet the technical criteria but occur in a regime where the mean reversion doesn't hold.
the equity-relative sizing change was supposed to fix this. it did, in the backtest. the fold rate improved from something unacceptable to ninety two percent. but fold rate is a backtest metric. it measures how much the strategy's performance degrades when you tweak parameters. a high fold rate means the strategy is not brittle to small parameter changes. it does not mean the strategy is robust to regime changes.
gold funding regimes have shifted. the backtest window captured a period where funding mean-reverted cleanly. the live period has seen funding stay negative for extended stretches. the bot keeps going long, collecting small funding payments, and getting stopped out when price doesn't revert fast enough. the wins, when they come, are big enough to keep the profit factor above water. but the strategy is essentially a lottery ticket with a positive expected value. that is not what the backtest described.
what would falsify it next
first test: run the bot on out-of-sample data from the past three months, the period it has been live, and compare the backtest trade log to the live trade log. if the backtest would have taken the same trades and produced the same pnl, the strategy is not overfit. it is just in a drawdown. if the backtest would have skipped most of the losing trades because the signal never fired, the live signal is drifting from the backtest signal. that means the data pipeline or the exchange's funding calculation changed. fixable.
second test: disable the bot for one month and paper-trade a version that only enters when gold's thirty day realized volatility is above a certain level. the hypothesis is that funding oscillations are more predictable in higher vol environments. if the paper pnl improves materially, the original strategy is regime dependent and needs a vol filter. if it doesn't, the edge was never real and the backtest overfit. one month of paper trading costs nothing except patience.
the bot stays off until at least one of these tests produces a clear answer. running a strategy that wins one in four trades and barely breaks even is not research. it is hoping.
more at falsifylab.com
— research and educational content. not investment, legal, or tax advice. do your own research. positions and views may change without notice.