Original Title: "I Used AI to Find Contract Ghost Coins with 100% Winning Rate, and What Turned Out Useful in the End Was..."

Original Author: Cryptographic Skanda | Skanda (X: @thecryptoskanda)

TL;DR:

1. By studying over 220 Binance contract coins, hundreds of manipulation event samples, and over 60 data dimensions, we found potentially viable positive EV trading strategies in "ghost coins."

2. Data proves: predicting triggers and "topping out" are both unfeasible.

3. The only viable strategy: shorting during a sharp rise and retreat, and strictly executing a rebound exit.

4. The only effective indicator: naked K.

5. Enter early, hold short positions, and exit fast.

Here is the main text:

This week’s report from @coinglass_com actually pointed out two issues:

First, Binance, Binance, it’s still Binance @binance.

Second, the fact that 90% of contract trading volume has clearly articulated a truth:

The act of "gambling" has, in fact, become a consensus among users in the entire industry.

While I say this, I will definitely get criticized.

While I say this, I will most likely get criticized.

But since it’s gambling, there’s no need to pretend it’s value investing.

To gamble, one must be bold.

To gamble, one must engage in extreme high-speed volatility games.

And Binance’s contract ghost coins are among the few cases in this boring market where retail investors can really participate, truly yield results, and genuinely attract traffic.

Many scholars have condemned manipulated coins, saying they represent "negative EV," only leading to fewer and fewer retail investors in the market.

But the reality is that the capital that enters through ghost coin manipulation, as well as the funds participating in ghost coin trading, are almost one of the few remaining substantial sources of incremental capital in the current secondary market.

Moreover, they have several key characteristics:

Non-quantitative.

Directional.

Volatile.

If you trade in traditional markets, you'd have to win against various insiders on Capitol Hill and Wall Street.

When you come to trade ghost coins, you just need to beat a contract manipulator who may not even be much more knowledgeable than you.

The problem lies here:

How to find the patterns of contending with the contract manipulators, thus snatching food from the tiger's mouth?

Using AI + personal experience, I have tentatively grasped some insights.

Of course, the title is bound to be clickbait.

Otherwise, you wouldn’t have clicked in to read.

1. First, Clarify What "Ghost Coin" Means

When I say "ghost coin," I do not refer to a coin that simply rises quickly. Essentially, by "ghost coin," I mean a type of asset as follows:

• Spot control rate basically above 96%.
• Has a Binance contract, and whether there’s spot trading isn’t as important.
• Typically through over-the-counter financing, uses violent up and down swings in a short period to aggregate massive liquidity and counterparties.
• Profits by triggering long and short liquidations, earning fee rates from the opponent, and ultimately completing the spot offloading, thus conducting the entire harvesting process.

In simpler terms, this is an art of manipulation.

The trader must understand contracts, cross-exchange spot markets, blockchain, operations, and even human psychology.

2. Manipulators Are Not Invincible

Many people believe: manipulators are invincible.

But the truth is vastly different.

In the entire ghost coin game, actual participants include:

• Manipulators (the "players")
• Mouse warehouses
• Retail investors
• Trading platforms and their insurance funds
• Other whales

The mantis stalks the cicada while the oriole waits behind; it’s not a simple "manipulators beat retail investors unilaterally."

First of all, manipulators themselves often need financing.

Whether it’s a project party that financed over $15 million, or those "famous" market makers, relying solely on their own funds to operate in the secondary market at that level is often just a drop in the ocean.

And financing comes at a cost.

Trading is for profit, not for the sake of artistic performance.

Thus, manipulators cannot just say "I have enough chips, so I can manipulate effortlessly."

They face a pile of practical issues:

• What if retail investors don’t follow?
• What if retail investors follow but in the wrong direction or at the wrong timing?
• What if a larger whale attacks specifically?
• Even if nothing goes wrong before, what if the insurance fund for the trading platform is breached, triggering an ADL?

Then your money might not be withdrawable at all, and the siblings in Singapore know who I am referring to.

Therefore, there’s a very simple iron rule for ghost coin manipulation:

• As long as the profits on the opposing side that can be taken now exceed my continued input for manipulation, I will continue to pull, continue to dump, and continue to harvest.
• Conversely, I abandon the position and leave.

The statement may be crude, but it outlines a framework of decision-making for ghost coin manipulators.

3. Scientifically "Fighting Manipulators," Starting with Experiments

Since the question is "how to fight manipulators," I tried to quantify this.

1) How the Tools Were Built

In modern times, problems should be addressed with modern methods.

Referring to @karpathy's idea of Autonomous Research Loop, I constructed my own. As long as clear objectives, constraints, and experimental methodology are provided, the agent will run continuously until the data can no longer be extracted.

LLM used is Opus 4.6.

My 20x Claude Max can still handle this task.

For the sandbox, I directly used an idle iMac as a remote experimental machine;

Then remotely controlled it using Tailscale from VSCode on a Windows workstation.

In terms of data, undoubtedly, @coinglass_com was the biggest help for this research.

Thanks also to @AlbertCoinGlass for sponsoring the API for this study.

K-lines, order books, OI, funding rates, liquidations, all available.

In addition, I also used:

• Binance API
• Skill Hub (manual @0xOar, indeed very useful)
• Etherscan V2 API to pull historical on-chain records

2) The Data I Reviewed

I finally organized 12 categories and over 60 sub-dimensions, including:

• Funding Rate
• OI
• Long/Short Ratios (Retail / Large Investors / Positions / Accounts)
• Taker Buy/Sell Ratios
• Liquidation Amount
• Order Book
• On-chain Transfers
• K-lines

The initially selected coins included 16 manipulatable coins judged by my experience, including $RIVER, $STO, and $MMT.

4. Phase One: I Initially Aimed to Predict "Signals Before Price Surge"

Then I adopted a hypothesis that retail investors love to fantasize about but usually leads to problems: predicting signals before a price surge.

"There must be signals before ghost coin manipulation. For example, abnormal FR, accumulation of OI, unusual on-chain activities. Find these signals, arrange in advance, and then print money."

It turned out that this was the fastest way to lose money.

At that time, I did not have a particularly strict definition of "manipulation."

I simply started by manually extracting several of the most obvious "manipulation events" from the K-lines of $RIVER, $STO, and $MMT, then sought commonalities from these events and expanded to another 16 coins, forming an experimental set.

To prevent overfitting, Autoresearch strictly divided the time:

• Early data training
• Later data as holdout (persistence validation)
• Holdout set completely invisible during the training phase

The experimental method was also quite crude:

Starting from the extreme values of single signals, such as funding rates,

then gradually stacking additional indicators until F1 improved.

Results:

Training set F1 was pulled up to 0.72.

It seemed like it was going to work.

Upon going to holdout, it pretty much all failed, and F1 dropped to around 0.1.

In other words:

Relying on "predicting when manipulation will happen" basically doesn’t work.

5. The Problem Is: You Have Misunderstood the Causal Relationship

After the first version failed, I realized a fundamental question:

Ghost coins do not become ghost coins because they meet certain indicators.

Rather, they possess those indicator characteristics because they are inherently ghost coins.

This logic is actually very consistent with everyone’s intuition.

Even if the overall market is bad, there will always be ghost coins that go insane individually.

Ghost coins never reason with the overall market; they are only related to one thing:

Is there a manipulator?

So we cannot predict when a manipulated coin will activate.

The truly feasible direction is:

Wait until it has already activated, identify "this is a manipulated coin, it is now being operated," and then based on this state, seek trading strategies.

So I completely changed my thinking.

This time I began to strictly define "manipulation cycles":

A complete manipulation cycle is characterized by a quick rise and then a rapid fall within a short time.

The next questions to solve became:

• How much rise and fall constitutes a complete cycle?
• Once locked into a cycle, what method is used for trading?

I let AI discover all these by itself.

Experimental samples were greatly expanded:

• 16 coins benchmarked against 415 manipulation cycles
• Later expanded to 55 "manipulated coins" acknowledged by the market
• Ultimately labeled 1447 cycles

The sample size finally appeared credible, and then I began to have a series of failures...

6. Several Strategy Versions, Consistently Hit

V1: Shorting at High Positions

The first version strategy proposed a "shorting at high positions" idea.

Backtesting Sharpe +0.72.

Sounds good.

Upon running holdout, the training set and test set were completely in different universes.

Later, I realized that the problem was:

I provided too few constraints, and Opus defined what "high position" meant based on its own assumptions.

After a while, it was just performing a double-slit interference experiment for fortune telling.

V2: After Restricting Assumptions, Results Were Worse

So, for V2, I directly added restrictions:

• No assumptions allowed
• Each indicator must have data support
• Different trading styles must be differentiated

For example: explosive rise and fall, slow ascent followed by explosive fall, explosive rise followed by slow fall, etc.

I initially wanted it to find different manipulators' "fingerprints."

As a result, it very scientifically presented me with:

-0.28 Sharpe.

Then I asked Opus to explain the decision logic of V1 and V2 to me,

and I suddenly realized that the essence of both strategies was doing the same thing:

Trying to top short.

This is essentially no different from many stubborn sharps who short at tops; the only difference is that they manually hang it on a tree while I hang it there with AI.

At this point, I finally realized:

It’s not that the methods aren’t advanced enough.

It’s that the thinking itself was wrong.

Note: Here I considered long strategies, but the problem is:

The initiation of ghost coins is not easily traceable; although some ghost coins show obvious anomalies when they start, such as "creating withdrawal washes" being a common characteristic, the problem is the same: how to differentiate the entry direction?

Going long after "topping" or in a downtrend will certainly lead to a loss. But this "false positive" signal is difficult to verify in advance, and lacks a good distinction between manipulated and non-manipulated rises, so it is unfeasible.

7. V3: Think from the Perspective of the Manipulator

Returning to the previous decision framework:

Profit First

Manipulators will definitely act in accordance with the trend, operating in the direction of the least resistance to reduce losses.

What does this mean?

• When selling pressure is high, they will let the market drop, even join in.
• When they can no longer push it down and short positions increase, they will pull it up.
• Pulling spot does not necessarily require much money.
• Short positions either get burned by the fees or get liquidated.

So there must be a point:

When the manipulator feels that continuing to support the market is no longer profitable.

After this point, the manipulator will allow the market to decline.

Because beyond that, it is no longer cost-effective.

Thus, what we should really seek is not the top.

Rather, it is this:

The point of abandonment.

Then design trading and stop-loss logic around this point, to avoid being swept away by ordinary rebounds while not suffering large losses when the direction is wrong.

The results of the experiment at that time appeared extremely good:

• Logic settled on the 1H line.
• Two consecutive 1-hour K-line bodies broke below 5%.
• Equipped with a 3% trailing stop.
• Average PNL was also above +3%.

But then a problem arose:

Sharpe was over 15+, and it surprisingly passed the overfitting test.

This type of number indicates a problem.

8. V4: Aiming for "Real Trading Feasibility"

After V3, I began to suspect several things.

First, it’s very likely that I have overfitted, and the current experiment did not truly define "manipulated coins"; it's just training on the operation cycles of 55 coins.

Second, even the overfitting test may have issues.

So this time I switched to a new approach:

Directly simulate based on real trading costs.

I matched the order book depth, funding rate history in the manipulation cycles with trading timestamps, attempting to restore the "true order placing" costs.

The results were straightforward:

This strategy simply does not make money.

And the reasons are very simple:

• Ghost coins typically have extremely low order book depths.
• Slippage often starts at 2% directly.
• In reality, you can only place a bacterial position of less than $200.

Then I realized two even larger issues.

1) How do I know where to start calculating the support level?

From the perspective of real trading, I have no knowledge of future K-lines.

So how can I know where to start calculating the "1H support level"?

Upon investigation, I indeed found a pitfall:

AI was using the maximum peak in the next 1H K-line after the current position to calculate when it would break below.

This is a standard look-ahead bias.

In simple words, it uses future data to help make decisions in the present.

This can definitely generate profit.

But in reality, you don’t have this cheat.

2) Order book depth severely misaligned with intuition

The average order book depth I calculated was about 70K.

However, anyone who has dealt with ghost coins knows that when they are "acting up," the trading volume is evidently large.

This is completely inconsistent with intuition.

Thus, there are only two possibilities:

• The calculation method for the median order book depth is incorrect.
• I defined the manipulation cycle too broadly, causing excessive noise.

I then chose to tackle from the second problem:

Redefine the manipulation cycle.

I let the data figure it out by itself:

What degree of pump/dump amplitude and what duration best represent "effective manipulation events."

The results showed:

• Significant manipulation events most frequently occurred within the 20%-50% range.
• Once the duration of manipulation events exceeded 96 hours, the win rate became insignificant.

Combining data and experience, I adjusted the new manipulation cycle definition to:

Pump + Dump within the 20%-50% range + completed within 96 hours.

If it's too high, the sample is too small.

If it's too low, the noise is too great.

9. Redefining "Manipulated Coins," Expanding Samples, Retraining

Next, I decided to no longer focus only on those 55 coins.

I expanded the range to:

All newly launched 221 non-TradFi contracts on Binance after March 1, 2025.

Which means the batch of coins that emerged once Binance Alpha + contract strategies truly began to take shape.

Then I did a few things:

• Statistically estimated the manipulation intervals defined as

"Completing a XX% amplitude pump and dump within XX hours"

(Specific thresholds are omitted here for the sake of strategy effectiveness)

• Statistical frequencies of manipulation intervals that fit this definition for each coin type.
• Then, according to frequency, divided all coins into four categories:

1. Extremely High Manipulation

2. High Manipulation

3. Medium Manipulation

4. Low Manipulation

Ultimately, I filtered out 70 "very high manipulation" and "high manipulation" types from the 217 coins.

Then based on these newly defined manipulation cycles, I no longer distinguished trading styles but directly sought commonalities to identify pre-signal indications for "tops."

The concluding observations are very counter-intuitive:

• Trading volume is useless.
• Order book is useless.
• Pump rate is useless.
• High amplitude + lower high is also useless.
• Waiting for 4H confirmation is purely a waste of trading opportunities.

The only truly useful item is naked K.

Finally, we arrived at two relatively significant signals: V4A and V4B.

10. V4A and V4B

V4A: Enter Early, Seize the First Move

The logic is:

• Early entry.
• Do not look at volume, do not look at amplitude.
• "Topping confirmation" only looks at instances where selling pressure first exceeds buying pressure.
• Still based on the 1H closing price if it breaks the support level.
• But the threshold for confirming a break is lower than that of V4B.
• The search range after topping confirmation is shorter.

Flowing water does not争, only seeks the martial prowess that is always swift and unbeatable.

V4B: Slow Down, Wait for Confirmation

The logic is:

• Wait until the price has already dropped from the peak for a while.
• The market still exhibits significant volatility, but the decline has already been confirmed.
• The search range after topping confirmation is longer.
• Thus, more trend confirmations can be obtained.

It is more stable, and slower.

This represents two completely different philosophies.

Both strategies currently utilize the same exit method:

trail + SL

In other words:

• If the direction is correct after entering, and it moves against by more than X%, exit.
• If wrong upon entry, handle it according to stop loss SL.

This SL is also derived from data.

In 166 backtested cases:

• SL triggers were in single digits.
• Average loss was just over 1%.
• Maximum drawdown was -1.87%.

This means I compressed the single trade R:R to below 1:1.

Not relying on single-shot profits, but on:

Win rate + high frequency.

After incorporating new slippage, funding fees, and conducting a re-evaluation with the latest data set, I discovered:

The concerns previously had regarding order book depth were indeed not misplaced.

But it also confirmed another matter:

The manipulation range itself is precisely the region where trading volumes are most concentrated.

Ultimately, among the two strategies, only V4A held up.

The reason is simple:

In ghost coins, the importance of early entry greatly surpasses "confirming very steadily."

11: Real-Time Testing

To avoid having the entire research stagnate at the "backtesting genius, real trading ghost" stage, I did two more things:

• Built a scanner script using the Binance API to scan data required for the strategies.
• Deployed it on a VPS to scan every 60 seconds.

At the same time, I also made a dashboard to push signals directly to a Discord bot through a WebHook. Although real-time testing has not been running for long, and the samples remain small, the confidence intervals are very wide; strictly speaking, it can only be used as directional reference.

However, the general results align with the research phase.

Currently, it’s about as follows:

• V3: Most frequently triggered (around 70%), with a winrate of 50%, but slightly negative PNL.
• V4A: Moderately triggered (around 26%), currently a 100% win rate, with a PNL around 25%.
• V4B: Triggered only once, and incurred a loss simultaneously while building a dashboard to connect signals through a WebHook to a Discord bot.

Though the practical testing hasn’t been long, the results are basically consistent with the research findings.

V3: Triggered most frequently (70%), win rate 50%, but PNL slightly negative.

V4A: Moderately triggered (26%), currently win rate of 100%, PNL around 25%, though the sample size is too small to draw conclusions, ongoing testing.

V4B: Triggered only once, and incurred a loss.

12. Up to now, I have summarized several key points for profitability

1.Positions must be short.

The median holding time for V4A is only 1 hour.

2.Enter early, do not wait for confirmation.

Waiting for confirmation often results in missing the most profitable段.

3.Exiting must be decisive.

Once signs of reversal emerge, exit immediately.

4.First protect against losses, then strive for profit.

Do not treat ghost coins as value investments.

5.Ghost coins do not stop; there are plenty of opportunities.

With 70 monitoring groups, we can basically scan out 2–3 V4A signals a day.

Future Optimizations

I am currently still running real-time tests, but there are already several optimization directions worth considering:

1) Support levels and topping signals may be related to liquidation heatmaps.

Intuitively, I believe that

Better top and bottom signals might correlate directly with liquidation heatmaps.

Unfortunately, I have not yet obtained sufficiently good data.

The principle is also quite simple:

• If there are no more short counterparties above.

The manipulators won't have the motivation to continue pushing down, eating fees.

• Similarly, the same applies below.

Therefore, I have started building a comprehensive pool to gather all newly launched contract coins after March 2025, preparing to conduct tests specifically for liquidation heatmaps.

2) The scoring system for manipulation frequency is still somewhat arbitrary.

Currently, I am using the frequency of manipulation cycles occurring over a past period to score coins, distinguishing "extremely high manipulation" from "high manipulation."

But this system has an obvious problem:

It’s based on frequencies from the past 6 months.

In reality, for trading, we seek not "who was the most manipulated in the past,"

but rather "who is still worth manipulating now."

Many coins that were frequently manipulated in the past

Could very likely have entered the later stages of manipulation by the manipulators or may have already been abandoned.

Continuing to monitor them is not particularly meaningful.

Moreover, this "scoring" is also artificially defined.

Thus, a more reasonable direction should be to create a:

Dynamic scoring system decaying over time.

Initial data already supports this direction; now, we just need more real-time testing samples.

3) The range should be expanded to older coins.

Currently, my coin selection range mainly focuses on new contract coins after 2025.

But in reality, many older coins are even more suitable for manipulation:

• Have contracts, and sometimes even spot.
• Market cap is sufficiently low.
• The projects are basically concluded.
• No one is paying attention.
• Unlocking has been completed.

Such coins, in some sense, are practically templates for manipulation.

If V4A proves equally effective on major manipulated coins among these old coins, it would indicate that this approach is not just overfitting to new coin samples, but truly captures a more universal manipulation mechanism.

Currently, the data temporarily supports this direction.

However, I do not plan to change the selection of coins for V4A at this moment.

13. "Why are you studying this as a trading platform?"

According to our philosophy at @Hertzflow_xyz, any trading opportunity for an asset is essentially just about the ups and downs of a betting game, whether it’s traditional assets, mainstream coins, or even "Ponzi schemes." The focus lies not on the asset itself but on the movement laws of its price. As long as there are laws, strategies can be made.

Ghost coins are no exception.

A trading platform is not merely a place for assets to be listed but a venue for running strategies. Compared to researching asset fundamentals, what we need to do more is to investigate "what strategies can be run with this trading pair, and are there trading opportunities."

If so, we need to provide that to traders.

With AI in hand, you too have the chance to contend against the dog manipulator. We will also normalize such data services and will offer public services in forms such as @goo_economy skill in the future.

@Hertzflow_xyz testnet launched today with 17 "extremely high/high manipulation scored" assets, all of which are being tracked in the strategies we've tested. Starting here, find your "technique for capturing manipulators." The 17 are:

$0G

$AKT

$ARC

$F

$H

$HEMI

$HYPER

$MMT

$MOODENG

$PARTI

$PROMPT

$SOON

$STBL

$SWARMS

$TAC

$VINE

$ZEREBRO

If you’re not yet confident enough to put real money on the line, then the HertzFlow testnet is your experimental ground to test the waters.

Royalty and nobility have no special seeds! You could also be tomorrow's "manipulator!"

Original link

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。