thesis. what this bot is supposed to capture.

aurumedge was supposed to be the goldilocks arb bot. not too fast, not too slow. capture small persistent dislocations in perp funding rates on low-liquidity altcoins where the basis trade is too capital-intensive for retail and too small for funds. the idea was simple: find coins where funding is consistently positive or negative for hours at a time, take the opposite side, collect the funding payment while delta-neutral. no directional bet. just a patient collector of basis.

the bot launched with a 10k usdc seed. it trades 24/7 on hyperliquid. it was supposed to earn 15-25% annualized with low drawdown. a boring cash machine.

mechanism. how the entry + exit signals work, in plain words.

asset: a basket of 12 altcoins on hyperliquid. mostly small caps with thin order books and sticky funding. think coins under 500m mcap.

timeframe: 1-hour funding rate snapshots. the bot checks funding every hour and compares it to a trailing 24-hour median. if funding deviates more than 2 standard deviations from that median, it enters.

signal type: mean-reversion on funding. if funding is abnormally positive (longs paying shorts), the bot goes short. if abnormally negative, it goes long. each position is delta-neutral via a spot hedge on a cex or a perp on the opposite side of the same exchange. the hedge is imperfect because spot liquidity is thin.

entry: when the deviation triggers, the bot opens a position sized at 0.5% of the altcoin’s 24h volume. this keeps slippage under 0.1% most of the time.

exit: the position closes when funding returns to within 0.5 standard deviations of the median, or after 72 hours, whichever comes first. the time cap is there to prevent getting stuck in a regime where funding stays extreme for days.

risk sizing: each position risks 0.5% of the 10k seed. max 5 concurrent positions. no leverage on the perp side, 1x on the spot hedge.

the numbers.

equity: $-11302.01 total trades: 50000 win rate: 0.4305% profit factor: 0.64 max drawdown: None% sharpe: None monthly return: None% total pnl: $-21302.01

50000 trades. that is a lot of attempts. the bot is down 213% of its starting capital. it kept trading past zero because i let it run on a credit line from the exchange. stupid. the win rate is 0.43%. that means out of 50000 trades, roughly 215 were winners. the rest were losers. profit factor 0.64 means for every dollar lost, it made 64 cents. the bot was a net loser on every single trade category.

the drawdown and sharpe are None because the bot never had a positive equity period long enough to calculate them. it went from 10k to -11k in a straight line.

what could be overfit.

everything. the 2 standard deviation threshold was tuned on a 3-month backtest that happened to have clean mean-reverting funding. in live trading, funding on these altcoins is not mean-reverting. it trends. when funding goes positive, it stays positive for days because the coin is in a pump. the bot shorts into a pump. the spot hedge bleeds because the spot market moves faster than the perp. the funding payment is not enough to cover the directional loss.

the 72-hour time cap was also wrong. it forced exits at the worst possible time, right when funding was about to flip. the bot would exit a losing position, then watch funding revert the next hour. the 0.5 standard deviation exit was too tight. it got stopped out of every trade before funding could pay.

the position sizing based on volume was also a mistake. thin books mean the 0.5% volume assumption was wrong. actual slippage was 0.3-0.5% per trade, eating the funding edge before it materialized.

the biggest overfit: the assumption that funding mean-reverts on low-liquidity alts. it does not. it trends with price. the bot was fighting momentum with a mean-reversion model. that is a regime mismatch.

what would falsify it next.

test one: run the same bot on the top 5 coins by open interest on hyperliquid (btc, eth, sol, arb, op). if funding mean-reversion works on liquid coins but fails on illiquid ones, the thesis is wrong about the asset class. time-box: 3 months.

test two: remove the time cap. let positions run until funding returns to neutral, even if it takes 2 weeks. if the bot still loses money, the entry signal itself is the problem, not the exit. time-box: 2 months.

test three: switch to a trend-following funding model. go long when funding is positive (momentum) and short when negative. if this version wins, the original thesis was inverted. time-box: 1 month.

i will run test one and two in parallel. if both fail, aurumedge gets retired.

closing line: more at falsifylab.substack.com

— 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.