soulz: the liquidity sniffer that hasn't died yet
thesis
most onchain alpha decays because it's too obvious. everyone can see the mempool. everyone can watch the orderbook. the edge that lasts is the one hiding in microstructure noise, the stuff that looks like random jitter until you stare at it long enough.
soulz is a bet that there's a repeatable pattern in how liquidity appears and disappears on a five-minute timeframe. not momentum, not mean reversion, not funding rate arbitrage. something quieter. a specific sequence of orderbook events that, when it occurs, predicts a small directional move over the next few candles.
the hypothesis: market makers have a rhythm. they reload inventory in predictable ways. if you can detect the reload before it completes, you can front-run the spread compression by about ninety seconds. the edge is thin. the win rate is supposed to be thin. the whole thing works or fails on whether profit factor stays above one point two after fees.
mechanism
soulz watches a single perp pair on hyperliquid. timeframe is five-minute closes, but the signal itself comes from tick-level orderbook snapshots aggregated into five-minute windows. it's not looking at price. it's looking at the shape of the book: depth concentration, imbalance skew, and a derived metric i'm calling "liquidity refresh rate" because i don't have a better name for it.
entry triggers when the refresh rate crosses a threshold and the skew confirms direction. no indicator stack. no confirmation candle. the signal either fires or it doesn't.
exit is a fixed-bar hold. soulz doesn't trail, doesn't scale out, doesn't use a take-profit limit. it enters at market, holds for a predetermined number of five-minute bars, and closes at market. the hold period was chosen by scanning one thousand two hundred historical instances and finding the horizon where edge decay crosses zero.
risk sizing is flat. every trade is the same notional. no volatility adjustment, no kelly fraction, no dynamic anything. this is deliberate. if the edge is real, it should survive dumb sizing. if it needs clever sizing to show a profit factor above one, the edge probably isn't there.
soulz is paper-only. it's been running in a soak environment for a while now, generating fills against live orderbook data with simulated latency and a conservative fee assumption. the 100-trade tracker is live. two candidate parameter sets produced eval packs from historical replay. that's all. no real capital has been touched.
the numbers
equity: $11,796.45 total pnl: +$1,796.45 total trades: 1,236 win rate: 46.9% profit factor: 1.25 max drawdown: not tracked yet sharpe: not tracked yet monthly return: not tracked yet
these are paper numbers. simulated fills, simulated fees, no slippage surprises. the profit factor of 1.25 is above water but not by much. a few bad months of slippage could push it below 1.1. the win rate of 46.9% means this bot loses more trades than it wins. that's expected for a strategy that relies on a handful of outsized winners to carry the PnL. the average win is larger than the average loss, but not by a comfortable margin.
1,236 trades is a decent sample size for a five-minute strategy. it's enough to say the pattern isn't pure noise. it's not enough to say the pattern will persist through a regime change.
what could be overfit
the liquidity refresh metric has three parameters. they were chosen by scanning two years of historical tick data and picking the combination that maximized profit factor. that's a textbook overfitting risk. the parameter set might be describing noise in the 2022-2023 market microstructure rather than a durable behavioral pattern.
the exit hold period was optimized on the same dataset. if market-maker inventory cycles change speed, the fixed-bar exit becomes wrong-sized. too short and you leave pnl on the table. too long and you give back gains to mean reversion.
the strategy assumes hyperliquid's fee schedule stays constant. a fee increase of even one basis point would kill it. the edge is thin enough that fee assumptions are load-bearing.
the paper simulation doesn't model the liquidity impact of the bot's own orders. at small size this is fine. at any meaningful scale, the entry would push the book in the wrong direction and erode the edge. this is a small-capacity strategy by nature, but the paper results don't quantify where the capacity limit sits.
what would falsify it next
first test: run the 100-trade tracker to completion and compare live fill prices against the paper simulation's assumed fills. if realized slippage exceeds assumed slippage by more than fifty percent of the average win size, the strategy fails. this is a clean gate. no judgment call. the numbers either work or they don't.
second test: take the two candidate parameter sets from historical replay and run them both on out-of-sample data from the last three months. neither parameter set has seen this data. if profit factor on the out-of-sample period drops below 1.1 for both candidates, the signal was overfit and soulz gets killed at the CI gate like most of the others.
twenty-three bots have died in this pipeline. most of them never made it past the bootstrap confidence interval check. soulz is still breathing because the numbers haven't falsified it yet. but the bar is low right now. paper profit factor of 1.25 on historical data is not a victory lap. it's an invitation to test harder.
more at falsifylab.com
ā 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.