LYNXPerp anomaly scanner

Lynx

A screener for crypto perps that hunts volume moving before the news does — and grades itself honestly on whether that means anything.

Where this actually stands. Lynx is good at noticing that something unusual is happening on a specific coin. It has not been shown to know which way to trade it.

Measured over 246 flags across 25 coins and 120 days: following the detected flow returns +0.00% at 1h and −0.10% at 24h; fading it is the mirror. The t-statistics are 0.03 and −0.20 — anything under about 2 is indistinguishable from random. After fees, every cell is negative or zero.

So the engine reports both directions with a t-statistic every day and says “no directional bet is justified yet” when the number can't support one. Below a confidence floor it goes further and suggests no position at all. It is allowed to say it doesn't know.

The north star

Lynx exists to do three jobs, in a loop that tightens every week:

1
Suggest trade ideas with a confidence score.It combines multiple indicators — volume anomalies first, more leading indicators over time — into a single confidence that an idea turns a profit. Suggested size is proportional to that confidence.
2
Track the realized performance of every idea.Each alert becomes a paper trade. Forward returns at 1h, 4h and 24h settle its P&L — hit rate, calibration, and per-indicator edge are all measured, not assumed.
3
Adjust itself from those results.Which indicators actually predict profit is deliberately learned from outcomes, not hand-set. Weights bootstrap uniform, then a periodic refit re-learns them from realized returns — so the engine gets smarter every week.

What it watches

Every 15 minutes Lynx screens low-float and privacy-coin perps on Hyperliquid for one specific fingerprint:

→
An hourly volume spike far above the coin's own baselinemeasured with a robust z-score (median/MAD), so past spikes can't poison the baseline — plus a hard $1M hourly floor to ignore dust.
+
Rising open interestnew positioning, not churn. This is the filter that matters most: volume tells you how much trading happened, but OI tells you whether positions were opened or closed. Two people closing out create volume and lower OI.
−
No market-wide explanation, no public catalystBTC/ETH aren't spiking, breadth is narrow, and a news scan finds nothing. Signal = anomaly − explanation.

Only coins that could clear those bars are examined at all: a coin whose typical hour is $5k would need a 200× hour, so it never competes for one of the 16 candle fetches a free-tier run can afford. Among those, the ones running hottest right now go first.

The motivating case: in May 2026, ZEC perp volume spiked with rising OI and zero public news — days later the Orchard disclosure landed. Lynx was built to surface that shape early. Worth knowing: across 120 days and 25 coins, the specific price-down-plus-OI-up pattern that case represents fired 10 times. 96% of flags are new longs instead.

How it learns

Every alert is scored by a weighted blend of its indicators — volume z, ratio to median, OI rise, funding size and trend, isolation, absence of news, repeat flags, basis, persistence and measured taker flow — eleven in all. The weights are re-fit periodically from the realized returns of past ideas by ridge regression, so indicators that predicted profit gain weight, ones that didn't lose it, and counter-predictive ones get flipped. A confidence-calibration curve and an equal-weight benchmark keep the sizing honest.

Right now that benchmark is winning: out of sample the learned weights lose to the uniform bootstrap and the calibration curve is inverted — higher confidence has predicted worse outcomes. Nothing has been promoted to production as a result. The engine is designed to notice that about itself and refuse, which it did.

What it is not

Lynx never trades. It suggests; a human decides and executes. That's a design constraint, not a roadmap gap.

Honest caveats, kept on purpose: the thesis is hindsight-selected from one motivating case; most flags will be false positives; volume-spike detection is inherently noisy; and acting on information that looks like a tip can carry real legal risk. A flag from Lynx earns a manual look — nothing more.

Two ideas have already looked significant in testing and turned out to be noise — a “fade the spike” edge, which vanished once the test universe was sampled properly, and a dedicated accumulation detector, killed by a train/holdout split. That is the method working, and it is why nothing here claims an edge it can't show you the t-statistic for.