I Analyzed Hundreds of \"Crime Tokens\" In Perp And Uncovered The Only Winning Strategy
Source: http://x.com/i/article/2042252055902048256
TL;DR
After studying 220+ Binance perp pairs, several hundred manipulation events, and 60+ data dimensions, we found one positive-EV strategy that actually survives on crimes.
The data is unambiguous: predicting the pump and "picking the top" are both unworkable.
The only viable approach: short into the post-pump retracement, with strict trailing exits.
The only indicator that survived: Kline-only.
Get in early. Hold short. Exit fast.
Part I: The Coin Flip That Wasn't a Coin Flip
Last week's
@coinglass_com report told you two things.
First: The only one after Binance, is still Binance
Second: 90% of trading volume sits in perps. That number stops being a statistic and starts being a confession: "Gambling" is no longer a fringe behavior in this industry, but a consensus.
Saying that out loud will get me yelled at. Fine.
But if you're going to gamble, stop pretending it's value investing. Gamble big. Gamble fast. Gamble where the volatility actually lives.
Binance's "perp crime tokens" — are one of the only places left in this market where retail can show up, take a real position, get a real result, and find real flow. They are Alpha in a market that has very little of it.
Plenty of righteous influencers will tell you crimes are mere extractions and slowly drain the retail population. They're not wrong about the drain. They're wrong about who's funding it. The capital that MMs bring in to run these tokens, plus the capital that shows up to trade them, is one of the few non-trivial sources of fresh money still arriving at the secondary market.
And crimes have three properties almost nothing else does:
- Non-quantitative.
- Directional.
- Volatile.
Trading traditional markets means out-fighting Capitol Hill and Wall Street's insiders. Trading crimes means out-fighting one MM who isn't necessarily smarter than you.
So the question is simple: is there a valid +EV strategy to rob the MMs?
I used AI plus personal experience and got somewhere.
PS: The title is, of course, clickbait. You wouldn't be reading otherwise.
Part II: Defining the Animal
A "crime," as I'm using the term, is not just a token that goes up fast. It's a specific kind of asset:
- Holder concentration is 96%+.
- A Binance perp exists. Spot listing is optional.
- The market maker (MM) uses off-exchange financing to run violent pump / dump waves on short timescales, pulling in liquidity and counterparty flow
- The endgame is to harvest liquidations and funding fees from both sides while exiting spot inventory.
It is, quite literally, an art of manipulation. The MM has to understand and orchestrate perps, cross-venue spot, on-chain flow, marketing, and human psychology — all at once.
Part III: The MM Is Not Invincible
Most retail traders assume the MM always wins. It's simply not true.
Who are on the table for the game:
- The MM
- Insiders
- Retail
- The exchange and its insurance fund
- Other whales
This is not a one-sided beating
The MM usually isn't self-funded. Even a project that raised $15M, even that's a "famous" MMs — at the size required to move a Binance perp this hard, their own balance sheet is a rounding error. They are levered.
Leverage costs money. And MMs are here to make money. Simple logic
Even with enough token holding, "just crime it up" isn't a strategy. The MM has real problems:
- What if retail doesn't follow?
- What if retail follows but in the wrong direction and timing?
- What if a bigger whale shows up specifically to hunt them?
- What if they cut so deep they blow through the insurance fund and trigger risk control and get their money frozen? (Crypto bros in Singapore know what I'm talking about. You can't withdraw a dollar after that.)
So the MM's decision-making is brutally simple:
As long as the counterparty flow profit I can take right now exceeds my marginal cost of continuing, I keep pumping, dumping, and harvesting. The moment that flips, I walk away.
Part IV: Industrializing the Hunt
If the question is "how do you defeat the MM," then the answer has to be AI.
Tooling. Modern problems get modern tools. I borrowed @karpathy 's autoresearch loop idea: give the agent a clear objective, hard constraints, and a methodology, and let it run until the data has nothing left to give.
- LLM: Opus 4.6. My 20x Claude Max tier survived the workload.
- Sandbox: an idle iMac as a remote experiment box.
- Control: VSCode on a Windows workstation, connected over Tailscale.
Data. The single biggest contribution came from
@coinglass_com, with API sponsorship from
@AlbertCoinGlass. Klines, order books, OI, funding, liquidations — all of it.
I also pulled from:
- Binance API
- Skill Hub (manual hat-tip to@0xOar— genuinely useful)
- Etherscan V2 for historical on-chain activity
The feature space.
Twelve major categories, 60+ sub-dimensions, including:
- Funding rate
- Open interest
- Long/short ratios (retail / top traders / position / account)
- Taker buy/sell ratio
- Liquidation volume
- Order book
- On-chain transfers
- Klines
The starting universe was 16 tokens I'd already flagged from experience —
,
,
$MMT, and others.
Part V: Phase One. Trying to Predict the Pump.
I started where every retail trader fantasizes about starting: predict the pump before it happens.
"There must be a signal before manipulation. Funding anomalies, OI buildup, on-chain activity. Find the signal, front-run it, print money."
This is the fastest way I know to lose money.
My definition of "manipulation" was loose. I hand-picked the most obvious MM events from
,
,
$MMT, looked for shared metrics, and projected onto the other 16 tokens to form a training set.
Autoresearch did the time-split discipline for me:
- Train on early data.
- Hold out late data.
- Holdout strictly invisible during training.
Methodology was crude on purpose: start with the extreme value of one signal (e.g. funding rate), stack indicators one at a time, watch F1.
Result:
- Training F1: 0.72. Looked promising.
- Holdout F1: ~0.1. Total collapse.
Translation: predicting when manipulation starts is not a tradeable problem.
Part VI: The Causality Was Backwards
After the first screwup, I noticed the actual mistake.
Crimes do not become crimes because they exhibit certain features. They exhibit those features because they are already being crimed.
This matches what every MM-watcher already feels in their gut. The BTC regime can be terrible and a crime will still go vertical on its own. These tokens do not negotiate with the broader market. They negotiate with exactly one variable: is there an MM on this token, yes or no?
So the question isn't "when will the MM start." The question is: given that the MM has clearly already started, can I identify the state and trade off it?
Whole framing changed. I introduced a hard definition of a "manipulation cycle":
A fast pump followed by a fast dump, end-to-end, counts as one cycle.
Which immediately raises two sub-questions:
- How much pump and how much dump constitutes a cycle?
- Once we lock the cycle, how do we trade it?
Both got handed to the AI.
Sample size went up:
- 16 tokens→ 415 cycles
- Expanded to 55 market-recognized crime tokens
- Final labeled set: 1447 cycles
Sample size finally stopped looking like astrology. And then I started failing in interesting new ways.
Part VII: Three Strategies, Three Beatings V1: Short the high.
The AI's first idea was a top-picker strategy. Backtest Sharpe: +0.72. Reasonable on paper.
Holdout: a different universe entirely.
The post-mortem was simple. I'd given Opus too much freedom and it had self-defined "the high" however it wanted. I had built a quantum framework for an urban legend.
V2: No vibes allowed. Worse result.
V2 added constraints:
- No vibes-based definitions.
- Every indicator must be data-grounded.
- Distinguish MM styles (fast-up-fast-down, slow-up-fast-down, fast-up-slow-down, etc.).
The plan was to find each MM's fingerprint. Opus very rigorously returned a Sharpe of -0.28.
Then I asked it to explain V1 and V2's decision logic in plain language and it hit me: both versions were doing the same thing.
Both were top-picking.
That's not a strategy. That's the same trade as every top-picker who hangs themselves on funding fees, except mine was hanging itself on autopilot.
The lesson wasn't "method too primitive." The lesson was "framing wrong."
(I considered long strategies too. The problem is symmetric: pumps are not really predictable, even though some of them share signals like the "fake retracement washout." Too much noise that the pump is not distinguishable from baselines)
Part VIII: V3. Think As An MM
Back to the MM's first principle: profit first.
The MM will always take the path of least resistance. Concretely:
- If sell pressure is heavy, let it dump. Maybe pile on.
- If selling is exhausted and shorts are stacking, crime it back up.
- Criming spot doesn't have to cost much especially without a spot listing.
- Shorts either burn on funding or get blown out.
So somewhere along the path, there is a point where defending the price stops being worth it. After that point, the MM stops defending and lets the market fall, because the marginal return on protection has gone negative.
What we should be looking for is not the top.
It's the capitulation point for the MM.
Once you have that, you can build entry and stop logic that doesn't get knocked out by a routine bounce back and doesn't blow up when you're wrong.
Initial results were too good:
- Logic ran on the 1H timeframe.
- Two consecutive 1H closes -5% body each.
- 3% trailing stop.
- Average PnL: +3%+.
Sharpe 15+. Overfit test "passed."
Anyone with a pulse knows that's a flag, not a result.
Part IX: V4. Optimize for "Can I Actually Run This."
Three things were now suspect.
First: probably overfit, and "crime" doesn't even have a clear definition — it was fitted on manipulation cycles inside 55 hand-picked symbols.
Second: the overfit test itself was probably contaminated.
So I changed the loss. Instead of asking "is this profitable in theory," I asked "is this profitable after the order book hits back." I matched book depth and historical funding to every trade timestamp and tried to reconstruct real fill cost.
Result, immediately:
This strategy does not make money.
Reason, also immediately:
- Orderbooks are paper-thin in the dead zones.
- Slippage starts at 2%.
- Real position size caps out below $200. A nano position with leverage.
Then two larger problems showed up.
1) How do I know where the support level is in real time?
In a real trade I do not see future candles. So how exactly is the agent computing the "1H support level"?
I checked. It was using the highest peak in the next 1H window to compute the breakdown. Classic look-ahead bias. AI was using future data to make present decisions. Of course it made money. Real trading does not come with that cheat code.
2) Order book depth doesn't match what your eyes tell you.
My computed median depth was around 70K. Anyone who has actually traded crimes knows that during an active manipulation cycle volume is enormous. Something was wrong.
Two candidates:
- Median depth as I was computing it is the wrong statistic.
- My definition of "manipulation cycle" is too loose and I'm averaging in noise.
I went after the second one: redefine the cycle. I let the data tell me which combinations of pump/down magnitude and duration actually capture meaningful MM events.
Findings:
- Significant manipulation zones cluster in the 20%–50% magnitude band.
- Once duration crosses 96 hours, win rate stops being significant.
Final cycle definition:
Pump + dump within a 20%–50% range, completed inside 96 hours.
Higher: too few samples. Lower: too much noise.
Part X: Redefine "Crime," Expand the Pool, and Retrain
I stopped staring at the 55 tokens. New pool:
Every non-TradFi perp listed on Binance after 2025-03-01: 221 tokens.
This is the cohort that was born inside the Binance Alpha + perps playbook that now defines new listings.
Then:
- Statistically define a manipulation zone as "a pump and a dump of XX% within XX hours" (specific thresholds withheld for strategy reasons).
- Count how often each token contains zones matching that definition.
- Bucket all tokens into four tiers by frequency: ultra-high / high / mid / low manipulation.
- From 217 tokens, 70 land in ultra-high or high.
Now retrain on those, ignore MM style, look for the common pre-top signal.
The results were almost insulting:
- Volume: useless.
- Order book: useless.
- Pump speed: useless.
- High amplitude + lower high: useless.
- Waiting for 4H confirmation: actively destructive.
The only thing that survived was Kline-only.
That gave us two distinguishable signals: V4A and V4B.
Part XI: V4A and V4B
V4A: Get in early.
- Early entry.
- Ignore volume and amplitude.
- Top confirmation triggers the first time instantaneous sell pressure exceeds buy pressure.
- Support break still uses 1H closes.
- Lower break threshold than V4B.
- Shorter post-confirmation search window.
Who gets in early takes the whole pie.
V4B: Wait for confirmation.
- Wait until price has already dropped some distance from peak.
- Market is still volatile, but the downtrend is confirmed.
- Longer post-confirmation search window.
- More trend confirmation per signal.
Slower. Steadier.
Both strategies share an exit framework: trail + SL.
- If the trade goes your way and then reverses by more than X%, exit.
- If you're wrong from the start, hit the stop.
The stop level is also test-derived. Across 166 backtested trades:
- Stops triggered single-digit times.
- Average loss just over 1%.
- Max drawdown: -1.87%.
R:R per trade is below 1:1. The strategy doesn't win because each trade is fat. It wins on win rate × frequency.
After re-running the overfit test with new slippage, fresh funding, and the new dataset: the order book depth concern was real. But it also turned out that manipulation zones are exactly where the order book is not thin. Volume concentrates inside the MM window.
Of the two strategies, only V4A survived.
The reason: on crimes, entering early matters far more than confirming carefully.
Part XII: Live Testing
Now put the strategies into testing. What I built:
- A scanner script using the Binance API, deployed on a VPS, sweeping every 60 seconds.
- A dashboard pushing signals via webhook into a Discord bot.
Live runtime is short, sample size is small, confidence intervals are wide. Take this as direction, not conclusion. With that disclaimer:
- V3: Most frequent signal (~70%). Win rate 50%. PnL slightly negative.
- V4A: Moderate frequency (~26%). Current win rate 100%. PnL ~25%. Still need more samples to commit.
- V4B: Triggered once.NGMI.
Part XIII: What I've Learned
- Hold short.
V4A's median holding time is 1 hour.
- Enter early. Don't over-confirm.
Confirmation is how you watch the fattest part of the move walk away.
- Exit with no hesitation.
First sign of reversal, leave. Protect the zero, then chase the win.
- Don't FOMO.
They aren't going extinct. With 70 monitored symbols, V4A produces 2–3 signals a day. You don't need this trade. Another one is coming.
Part XIV: Open Optimizations 1) Liquidation heatmaps are likely better top and bottom indicators
Intuition: better top/bottom signals likely sit inside liquidation heatmap data. I don't have clean enough data yet. The mechanism is direct: if there are no shorts left above, the MM has no incentive to keep pumping upside and harvesting funding. Symmetric on the downside. I've started building a unified pool — every new perp listed since March 2025 — specifically to test this.
2) The manipulation-frequency scoring is still partially eyeballed.
Right now the score is "how often did this token produce qualifying cycles in the last 6 months?" The problem: I don't actually want to know who was the most crimed — I want to know who is still worth criming right now. Many high-frequency historical names are deep into the MM's exit phase, possibly already abandoned. Watching them is worth nothing. The right answer is a time-decayed dynamic score.
Early data supports it. I just need more live samples.
3) Expand to old tokens.
My universe is currently restricted to post-2025 listings. In practice, many older tokens are better crime material:
- Has perps. Often has spot.
- Low cap.
- Project effectively dead.
- No attention.
- Unlocks finished.
These are textbook MM targets. If V4A holds on the crime-grade names inside the legacy universe, that means the strategy isn't fitted to a new-listing artifact — it has actually captured a general MM mechanic. Early data supports this. I'm not changing V4A's universe yet.
Part XV: Why Would I Research This In The First Place
The
@Hertzflow_xyz philosophy is straightforward. Any tradeable opportunity is, in the end, a directional bet on price motion. Traditional assets, majors, "Ponzi schemes" — what matters is not the asset, it's the regularity of how the price moves. If there is regularity, you can run a strategy on it.
Crimes are not an exception.
We do this to understand one of the biggest category of crypto-native trading opportunities and how can we Hertzflow accommodate it
We're going to standardize this kind of data as a service, eventually exposed publicly through
@goo_economy as skills.
The
@Hertzflow_xyz testnet goes live today with 17 ultra-high / high manipulation-score assets. Every one of them is currently being tracked by strategies we've already validated:
$0G ·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
If you're not yet confident enough to put real capital down, the HertzFlow testnet is where you test the water
Find your own playbook