Source-linked AI summary
The Anatomy of a Cryptocurrency Pump-and-Dump Scheme
Jiahua Xu, Benjamin Livshits
TL;DR
Cryptocurrency pump-and-dump activity had not been comprehensively studied despite regulatory concern and growing use of anonymous communication platforms. The paper analyzes organized events, models which coins are likely to be pumped, and combines those predictions with a trading strategy that achieves returns as high as 60% over two and a half months. Its findings provide a proof of concept for strategic crypto-trading and machine-learning-based crime detection, while the strategy is limited to small retail investments and sensitive to threshold and market-impact assumptions.
Problem
The paper addresses the lack of comprehensive empirical evidence on organized cryptocurrency pump-and-dump schemes despite regulatory concern and their prevalence on communication platforms.
Method
The authors trace Telegram messages identifying 412 events, analyze coin and market features, and train predictive models estimating each listed coin’s pre-pump likelihood.
Results
An out-of-sample trading strategy using the prediction models achieves returns as high as 60% over two and a half months under conservative assumptions.
Takeaways & Limitations
The study provides a proof of concept for strategic crypto-trading and applies machine learning to detecting pump-and-dump activity.
Takeaways & Limitations
The trading strategy applies only to small retail investments because large orders can move liquid markets and prompt organizers to cancel or switch the pump.
Abstract
from arXiv · showhide
While pump-and-dump schemes have attracted the attention of cryptocurrency observers and regulators alike, this paper represents the first detailed empirical query of pump-and-dump activities in cryptocurrency markets. We present a case study of a recent pump-and-dump event, investigate 412 pump-and-dump activities organized in Telegram channels from June 17, 2018 to February 26, 2019, and discover patterns in crypto-markets associated with pump-and-dump schemes. We then build a model that predicts the pump likelihood of all coins listed in a crypto-exchange prior to a pump. The model exhibits high precision as well as robustness, and can be used to create a simple, yet very effective trading strategy, which we empirically demonstrate can generate a return as high as 60% on small retail investments within a span of two and half months. The study provides a proof of concept for strategic crypto-trading and sheds light on the application of machine learning for crime detection.
1 Introduction
The paper addresses the lack of comprehensive empirical evidence on cryptocurrency pump-and-dump schemes by tracing organized events and associated market behavior. It also develops predictive models and a trading strategy based on pre-pump market movements.
- Empirical study: 412 pump events were identified by tracing message histories from over 300 Telegram channels.The events span June 17, 2018 to February 26, 2019.
- Research gap: The study examines routinely organized cryptocurrency pump-and-dump events, addressing a phenomenon previously lacking comprehensive research.It traces records across multiple exchanges from June 17, 2018 to February 26, 2019.
- Empirical study: Around 100 organized Telegram channels coordinate an average of 2 pumps daily, generating 6 million USD in aggregate artificial trading volume monthly.The paper also reports that some exchanges actively participate in pump-and-dump schemes.
- Prediction: The random forest model predicts each coin’s pre-pump likelihood with an AUC of over 0.9.The model uses pre-pump market movements and is reported to perform strongly both in-sample and out-of-sample.
- Trading strategy: A calibrated prediction model combined with a simple trading strategy demonstrates a 60% return over eleven weeks under strict assumptions.The strategy is presented as applying to small retail investments.
2 Background
The paper describes pump-and-dump schemes as coordinated cryptocurrency demand manipulation organized through chat platforms and exchanges. It outlines the actors, sequence of actions, rapid price reversal, insider advantage, and regulatory ambiguity surrounding these events.
- Concept: A pump is a coordinated, intentional, short-term increase in cryptocurrency demand that leads to a price hike.Telegram and Discord facilitate these activities through encryption and anonymity features.
- Actors: Organizers coordinate events through chat applications, while participants buy the announced coin and often purchase at inflated prices.Organizers possess insider information and are described as the ultimate beneficiaries, whereas participants are the scheme’s victims.
- Actors: A target exchange is selected for the event and may benefit from higher prices, increased transaction fees, or access to order information.The paper identifies Yobit as an exchange that openly organized pumps multiple times.
- Process: The typical sequence is set-up, pre-pump announcement, pump, dump, and post-pump review.Organizers recruit members, announce the time, exchange, and pairing coin, issue countdowns, reveal the target coin, and later report selected price points.
- Market dynamics: Pump attempts can generate rapid price surges followed by declines as participants panic-sell, while organizers selectively present favorable results afterward.Selective disclosure of volume and duration can give newcomers the impression that pumps are highly profitable.
- Process: During a failed attempt, administrators abandoned the pre-announced pump because the target coin’s price moved unexpectedly.The event indicates that administrators knew the coin choice before public announcement, giving them time to accumulate it before participants acted.
- Regulatory context: Cryptocurrency pump-and-dumps operate in a setting where securities-law applicability may be ambiguous and regulation remains weak.The paper states that information asymmetry and incomplete disclosure can constitute intentional investor deception.
3 A Pump-and-Dump Case Study
The BVB case study traces a pump targeting a dormant, low-rated coin, followed by an exceptionally rapid price spike and collapse. Trading activity showed inflated buy volume, limited participation, and investors potentially left holding illiquid coins.
- Target and announcement: The pump was organized across at least four Telegram channels and targeted one of Cryptopia’s BTC markets.The largest channel, Official McAfee Pump Signals, had 12,333 members.
- Target and announcement: The pump targeted BVB, a dormant coin associated with an unserious project and rated 1 out of 5 on Cryptopia.BVB was not listed on CoinMarketCap, and its associated project activity had largely ceased.
- Price and volume dynamics: Within the first 15 minutes, BVB’s volume rose from nearly zero to 1.41 BTC while its price increased from 35 Sat to 115 Sat.The price reached roughly three times its starting value during the event.
- Price and volume dynamics: The first buy order completed within 1 second, and the price reached its peak after 18 seconds, before one channel announced the coin.The authors conjecture that the first rapid order may have been automated, while later-announcing participants could not profit from the peak.
- Price and volume dynamics: The price fell below its opening level three and a half minutes after the pump began, after which transactions became sporadic.The rapid reversal followed the brief peak reached during the buying wave.
- Aftermath: Buy volume exceeded sell volume, totaling 1.06 BTC versus 0.58 BTC and 1,619.81 thousand BVB versus 1,223.36 thousand BVB.The discrepancy indicates greater buying aggressiveness and suggests some holders were unwilling or unable to liquidate after the collapse.
- Aftermath: Only 322 trading transactions accompanied an announcement viewed 1,376 times by a channel exceeding 10,000 members.The authors interpret this low participation ratio as consistent with observers or members aware of the difficulty of profiting.
4 Analyzing Pump-and-Dump Schemes
The study combines Telegram and exchange data to characterize pump-and-dump activity, its market effects, and features associated with pumped coins. It finds substantial concentration across exchanges, abnormal volume and returns, cross-exchange price discrepancies, and a preference for small-cap coins.
- Data and approach: The study combines Telegram channel histories with exchange data to analyze and model routinely organized pump-and-dump events.The analyzed event pattern is set-up, pre-pump announcement, pump, dump, and post-pump review.
- Activity distribution: 412 events were distributed unevenly across exchanges: 51% occurred on Cryptopia, 27% on Yobit, 17% on Binance, and 5% on Bittrex.The selected coin had previously been pumped on the same exchange in 35% of events.
- Trading volume: 8,793 BTC of trading volume occurred during pump hours, compared with 943 BTC before pumps, making pump-period volume 9 times higher.Binance accounted for 93% of the pump-period volume, roughly equivalent to 50 million USD.
- Price effects: Pumps on Yobit and Cryptopia produced larger price increases and sharper subsequent declines than pumps on Bittrex and Binance.Binance generated more pump-hour volume, but its price increases were generally smaller, possibly because of substantial bid and sell walls.
- Cross-exchange effects: Pump activity created cross-exchange arbitrage opportunities because the same coin could have different prices during the pump hour.The pumped exchange was not always the venue with the highest price.
- Coin characteristics: Pumped coins ranged from 1 BTC to 27,600 BTC in market capitalization, and half had market caps below 100 BTC.The preference for small-cap coins is consistent with evidence that smaller market capitalization is associated with more successful pumps.
- Pre-pump signals: Abnormal return signals appeared before announcements, were most concentrated one hour before pumps, and were strongest on Cryptopia.In numerous Cryptopia pumps, the hourly return before the pump exceeded the hourly return during the pump, supporting prediction before announcement.
5 Predicting Pump-and-Dump Target Coins
The paper predicts which coins will be pumped from pre-pump market information and evaluates a trading strategy based on those predictions. Random-forest models outperform GLM models, while threshold choice affects both returns and portfolio diversification.
- Feature selection and sample: Market movements before a coin announcement can predict which coin will be pumped using coin features and pre-pump data.
- Feature selection and sample: 53,208 pump-coin observations contained 180 pumped cases, making the sample heavily imbalanced toward unpumped coins.On average, each event included 296 candidates and one actual pumped coin.
- Feature importance: Market capitalization and last-hour return were the most important random-forest features, while short-term movements mattered more than longer-term movements.Return features were generally more important than volume or volatility features, and return1h had the highest explanatory power when used alone.
- Feature interpretation: Positive return coefficients indicate that higher pre-pump returns are associated with greater pump likelihood, while previously pumped coins are more likely to be pumped again.
- Model evaluation: Random forests achieved ROC AUC values above 0.94, compared with 0.63–0.88 for GLM models, and retained strong performance out of sample.The authors therefore eliminated GLM models from further analysis.
- Investment strategy: The trading strategy invested one hour before announcements in proportion to each coin’s random-forest pump likelihood, using a 0.37 BTC market-depth baseline.Theoretical returns reached 140% in training and 80% in validation under model-specific thresholds; the test achieved 60% over two and a half months.
- Investment strategy: High thresholds can improve precision but may exclude all purchases or create an undiversified portfolio, and validation returns eventually decline as thresholds rise.
- Investment strategy: The strategy purchased nine test-sample coins, all ultimately pumped, producing a 60% return with negligible transaction-fee effects.The result was similar to training and validation results using RF1 with threshold 0.3.
6 Related Work
The paper extends emerging research on cryptocurrency pump-and-dump activity by pursuing prospective prediction with a homogeneous Telegram-based dataset. It also connects to broader work on market manipulation, crypto-market arbitrage, and cryptocurrency price dynamics.
- Pump-and-dump research: Earlier studies examined cryptocurrency pump-and-dump activity, while this paper distinguishes itself through prospective prediction rather than retrospective investigation.
- Pump-and-dump research: The study uses clearly announced Telegram events from June 17, 2018 to February 26, 2019, with Discord announcements overlapping the Telegram data.
- Broader market-manipulation literature: Related literature addresses manipulation methods, regulatory deterrence, Bitcoin price manipulation, crypto-market arbitrage, and transaction-exposure risks.
- Cryptocurrency market research: Other cryptocurrency research studies market movements through Bitcoin time-series models, bubbles, asymmetric volatility, and the relationship between discussion quality and price volatility.
- Comparison of studies: The paper’s prior-work comparison is summarized in a table covering studies of cryptocurrency pump-and-dump activity.
7 Conclusions
The paper concludes that cryptocurrency pump-and-dump activity is persistent and that pre-pump market movements contain information useful for predicting target coins. Its models and trading strategy provide a proof of concept for strategic trading and possible market-abuse detection.
- Pump-and-dump activity persists in cryptocurrency markets and drives tens of millions of dollars in phony trading volume each month.
- Pre-pump market movements frequently contain information about which coin will be pumped, enabling prediction with LASSO-regularized GLM and balanced random-forest models.
- Out-of-sample tests show that the trading strategy can exploit returns as high as 60% over two and a half months under conservative assumptions.
- The authors propose that machine-learning prediction could support strategic trading, reduce victimization, and help regulators detect market abuse and criminal behavior.
Appendix
The appendix documents feature-importance and coefficient visualizations used to inspect the models. These figures respectively show mean-decrease-in-Gini importance and unstandardized GLM coefficients.
- Figure 17 ranks feature importance using mean decrease in the Gini coefficient, with darker cells indicating higher importance.
- Figure 18 reports unstandardized GLM variable coefficients and marks variables not selected by the model with “-”.