# 🛡️ AI vs. Memecoin Manipulators: The Multi-Agent Trading Strategy ## Tags: #CryptoTrading, #ArtificialIntelligence, #Memecoins, #MultiAgentSystems, #Cybersecurity ## 📋 Overview - **Type**: Academic Research Paper / Strategic Technical Framework - **Main Topic**: Utilization of a multi-agent LLM system to detect manipulative bots and execute profitable "copy trading" strategies in the Solana memecoin market. - **Authors**: Researchers from University College London (UCL) and Nanyang Technological University (Singapore). ## 🎯 Core Purpose & Context The explosive rise of memecoins (spearheaded by platforms like **Pump.fun** on Solana) has popularized "Copy Trading"—where retail investors automatically mimic the trades of successful wallets. However, this market is plagued by **manipulative bots** and **rug pulls**, making naive copy trading highly unprofitable. This research aims to solve two critical problems: 1. **Bot Detection**: How to filter out scams and coins artificially pumped by bots. 2. **KOL Selection**: How to differentiate between a truly skilled trader (Key Opinion Leader) and a lucky gambler or a wash-trading bot. The researchers propose a **Multi-Agent System** that mimics a professional hedge fund team, using Chain-of-Thought (CoT) reasoning to analyze on-chain data, visual charts, and social sentiment simultaneously. --- ## 🧠 Key Concepts & The Ecosystem ### 1. The Battlefield: Pump.fun & Bonding Curves - **Launch Mechanism**: Anyone can launch a coin for free. Prices are determined by a "bonding curve" (mathematical price increase based on demand). - **Migration**: Once a coin reaches a market cap threshold (selling ~800M tokens), it "graduates" to a decentralized exchange (Raydium). - **The Danger Zone**: The period before migration is where most scams occur. ### 2. The Enemy: Types of Manipulative Bots The paper identifies four specific automated threats used to scam traders: * **Bundle Bots**: The creator buys a huge chunk of supply in the very first block (concealing true ownership) to dump later. * **Sniper Bots**: Algorithms that buy immediately after launch to front-run retail traders. * **Bump/Wash Bots**: Bots that buy and sell the same amount repeatedly to artificially inflate volume and get the coin on the "Trending" page. * **Comment Bots**: Scripts that spam generic "To the Moon!" comments to create fake social proof. ### 3. The Solution: Multi-Agent Framework Instead of one AI doing everything, the system splits tasks among specialized agents: * **🕵️ Meme Evaluation Agent**: Analyzes the project. *Is this a scam?* * **👤 Wallet Evaluation Agent**: Analyzes the trader to copy. *Is this guy actually good?* * **💰 Wealth Management Agent**: Decides allocation. *Can we afford this?* * **⚡ DEX Agent**: Executes the trade. --- ## 🧭 Strategic Analysis & "Game Changers" ### 🕵️‍♂️ From Technical Analysis to "Forensic Blockchain Analysis" The most significant shift in this paper is the move away from traditional price prediction (RSI, Moving Averages) toward **forensic analysis**. The AI isn't predicting if the line goes up; it is investigating **crime scenes**. It looks for "Bundle Bots" (hidden supply control) and "Wash Trading" to disqualify assets. * **The Insight**: In memecoins, *security* is the primary alpha. If you avoid the scam, the volatility takes care of the profit. ### 🤖 The "Hedge Fund in a Box" Architecture This paper validates the **Multi-Agent** approach over single LLMs. A single GPT-4 instance struggles to process charts, sentiment, and transaction hashes simultaneously. By assigning specific roles (Analyst vs. Trader vs. Risk Manager), the system achieves **73% precision**, significantly outperforming single models. This suggests the future of Fintech AI is **modular**, not monolithic. ### 💡 The "So What?": Solving the Lemon Market The memecoin market is a classic "Market for Lemons" (where bad products drive out good ones because buyers can't tell the difference). This framework proves that AI can act as the **verification layer**, restoring trust. The agents identified KOLs (traders) that generated **$500,000 in profit**, proving that "smart money" does exist on-chain—you just need AI to find it amidst the noise. --- ## 📊 Detailed Breakdown & Algorithms ### Phase 1: Detecting the Manipulation (The "Scam Filter") The researchers developed specific algorithms to flag bad actors before the AI even makes a decision. **Algorithm 1: Bundle Bot Detection** * **Logic**: It scans the very first block of the coin's existence. * **Red Flag**: If multiple wallets buy in Block 0 and are funded by the creator (directly or indirectly), it is a **"Creator-Funded Bundle."** This is a high-probability rug pull setup. **Algorithm 2: Bump/Wash Bot Detection** * **Logic**: It calculates a "Wash Trading Score." * **Formula**: Ratio of "matched buy-sell pairs of identical amounts" to the trader's net position. * **Red Flag**: If a trader flips the same amount repeatedly without holding, they are faking volume. ### Phase 2: Agent Workflow 1. **Meme Evaluation Agent (The Gatekeeper)** * **Inputs**: Candlestick charts (visual), Transaction history (bundles), Comment history (text). * **Process**: Uses Few-Shot Chain-of-Thought. It identifies if the chart looks "organic" (healthy volatility) or "manufactured" (few large green candles followed by silence). * **Output**: Boolean (Good Farming Potential: Yes/No). 2. **Wallet Evaluation Agent (The Headhunter)** * **Task**: Identify "Smart Money" to copy. * **Key Metrics**: * *Profitability*: Must have consistent wins. * *Experience*: Avoiding wallets that only traded once (luck). * *Bot Check*: Ensures the wallet isn't just a sniper bot. * **Output**: Boolean (KOL: Yes/No). 3. **Performance Results** * **Dataset**: 1,000 memecoins launched post-$TRUMP creation (Jan 2025). * **Meme Selection Precision**: **73%** (The AI correctly identified profitable coins). * **KOL Identification Precision**: **70%** (The AI correctly identified profitable traders). * **Comparison**: The Multi-Agent system significantly outperformed traditional Machine Learning (LASSO, Random Forest) and Single-Agent LLMs. ## 🗞️ Key Facts & Timeline (Case Study: $MAO Coin) *To illustrate the speed of manipulation, the paper tracked a coin called MAO:* * **15:06:24**: Creator deploys MAO. Instantaneously, **Launch Bundle** bots buy supply in the same block. * **15:06:25**: **Sniper Bots** buy in immediately (1 second later). * **15:10 - 16:26**: **Comment Bots** spam "SEND IT" and **Bump Bots** fake volume. * **16:40:54**: **Rug Pull**. Creator dumps everything. Price crashes. * **Takeaway**: The entire lifecycle was less than 2 hours. Humans cannot process the forensic data fast enough to spot the bundle; AI can. ## 🔑 Key Takeaways 1. **Copy Trading is Dangerous**: Without filtering, you are likely copying a bot or a lucky gambler who is about to lose everything. 2. **Multimodal Analysis is Mandatory**: You cannot look at price alone. You must look at the **Blockchain** (bundles), the **Chart** (visual patterns), and the **Socials** (bot detection) together. 3. **Chain-of-Thought Works**: Forcing the AI to explain its reasoning (e.g., "I see a creator bundle, therefore this is risky") drastically improves performance over simple prediction. 4. **Bot Indicators**: High volume does not equal interest. If the volume is comprised of perfectly matched buy/sell orders, it is a **Bump Bot**. ## ❓ Unresolved Questions / Follow-up - **Adversarial Evolution**: If this AI becomes popular, creators will stop using "same-block" bundles to hide. How will the AI adapt when creators use delayed funding or CEX-funded wallets to hide their tracks? - **Execution Speed**: The paper analyzes data *up to 12 hours* post-migration. Can this system run in real-time (sub-second latency) to catch opportunities before they expire? ## 🕰️ Detailed Chronological Walkthrough Based on the provided text segment, here are the detailed notes and important points: ### **Publication Details & Overview** * **Title:** Resisting Manipulative Bots in Memecoin Copy Trading: A Multi-Agent Approach with Chain-of-Thought Reasoning. * **Context:** The paper addresses the surge in meme coin investment sparked by the launch of the **\$TRUMP** coin on **January 17, 2025**. * **Problem:** * Copy trading on platforms like **GMGN** is popular but risky due to manipulative bots, entry lag, and unpredictable "Key Opinion Leader" (KOL) performance. * KOLs often use copiers as "exit liquidity" (buying low, waiting for copiers to inflate price, then dumping). * Single Large Language Models (LLMs) struggle with asset allocation and lack domain-specific data for cryptocurrencies. * **Proposed Solution:** An explainable **multi-agent system** using few-shot Chain-of-Thought (CoT) prompting, modeled after an asset management team. * **Performance Results:** * Outperformed traditional Machine Learning (ML) and single LLMs. * Achieved **73% precision** in identifying high-quality meme coin projects. * Achieved **70% precision** in identifying high-quality KOL wallets. * Selected KOLs generated a total profit of **$500,000**. ### **Methodology: The Multi-Agent Framework** The system decomposes trading into four specialized subtasks handled by distinct agents: * **Meme Evaluation Agent:** * Identifies meme coins with growth potential and long-term viability. * Analyzes candlestick patterns, trading metrics, and comment sentiment. * **Trader Evaluation Agent:** * Selects KOL wallets to follow. * Assesses candidates based on historical trading performance. * **Wealth Management Agent:** * Allocates capital across various copy trading opportunities. * **Order Execution Agent:** * Responsible for submitting buy orders on the **Pump.fun** platform. ### **Dataset Discrepancies in Text** * The **Abstract** states the empirical evaluation used a dataset of **1,000** meme coin projects. * The **Introduction** states the data used covered **4,000** meme coin projects. ### **Background: Meme Coins & Pump.fun** * **Pump.fun Model:** * The largest meme coin crowdfunding platform on **Solana**. * **Creation:** Users upload an image, name, and ticker to create a coin with a **1 billion** total supply. * **Distribution:** 800 million coins are tradeable; 200 million are locked. * **Social Features:** Functions like StockTwits; trades "bump" coins to the front page; flags potential bot activity. * **Lifecycle Stages:** 1. **Launch Stage:** "Primary market" where traders trade with the issuer/contract via a bonding curve. Subscription withdrawals are allowed. 2. **Migration:** Triggered when all 800 million tradeable coins are purchased. The project is listed on a Decentralized Exchange (DEX) like **Raydium**. ### **Bonding Curve Mechanism** * **Function:** Defines the price relationship between SOL deposited and meme coins received. * **Formulas:** * **Relation:** $y = y' - \frac{k}{x + x'}$ * $x$: SOL deposited. * $y$: Meme coins received. * $k$: Constant product. * $x', y'$: Virtual reserves (Constants: $x' = 30$, $y' = 1,073,000,191$). * **Price:** $P = -\frac{(x + x')^2}{k}$ * ** dynamics:** Price rises as demand (SOL deposited) increases. * **Fees:** Traders pay a **1% transaction fee**. ### **Market Actors** * **Pumpfun:** Charges 1% trade fees and a 0.015 SOL migration fee. * **DEX:** Uses Automatic Market Makers (AMM) for secondary market liquidity. * **Meme Coin Creator:** Initiates the contract. Often acts as a strategic manipulator using bots to hide ownership. * **Traders:** Participants in launch or DEX stages; vulnerable to manipulation. * **Bot Providers:** Rent/sell scripts to creators/traders to facilitate market manipulation. ### **Manipulative Tactics & Bots** * **Rug Pulls:** Sudden exit scams where creators/early holders cash out. * **Common Bots:** * **Bundle Bot:** Hides ownership. * *Launch Bundle:* Creator buys immediately at token creation (lowest price) to hoard supply. * *Heuristics:* Identified when Creator + Wallets A & B buy in the **same block**. * *Creator-Funded Bundle:* Creator limits funding to Wallet A & B, who then buy in the same block. * **Volume Bot:** Simulates liquidity/activity to attract naive traders (Wash Trading). Or "Bump bots" that buy and sell repeatedly to keep the coin on the front page. * **Sniper Bot:** Buys tokens immediately after liquidity is added. * **Comment Bot:** Fabricates social sentiment. Here are the detailed notes and important points extracted from the provided text segment regarding meme coin manipulation and the proposed multi-agent copy trading framework. ### **3.4 Manipulative Bot Mechanisms on Pump.fun** * **Launch Bundle Bot (Counteracting Visibility)** * **Problem:** Creator holdings are public on the Pump.fun dashboard; heavy purchasing by the creator signals concentrated ownership and fear of a "rug pull," deterring investors. * **Solution:** Creators use "launch bundle bots" to generate, fund, and control multiple fresh wallets. * **Action:** These wallets simultaneously buy the coin within the exact same creation block. * **Goal:** Masks centralized ownership and creates an illusion of organic demand (analogous to splitting orders in stock markets). * **Detection:** * Pump.fun flags transactions in the same block as potential bot activity. * Only the coin creator can insert transactions into the creation block, making this suspicious. * **Bump Bot (Visibility Manipulation)** * **Mechanism:** Every transaction updates a token’s attributes (name, price, activity) on the platform front page. * **Action:** Bots repeatedly execute offsetting buy and sell orders. * **Effect:** Does not alter actual holdings but artificially inflates trading interest. * **Goal:** Keep the token on the front page to attract potential traders (inflating perceived popularity). * **Comment Bot (Social Manipulation)** * **Definition:** Automated scripts designed to fabricate user engagement. * **Action:** Disseminates brief, context-free, positive messages (e.g., “To the moon!”, “Don’t miss out!”) via multiple controlled wallets. * **Goal:** Mislead genuine users into perceiving strong community backing/social validation. --- ### **4. Case Study: MAO Token Lifecycle** * **Subject:** MAO meme coin (selected because it exhibits all four bot types). * **Timeline & Actions:** * **Stage 1: Token Creation & Launch Bundle** * **[2025-01-17 at 15:06:24]** (Block 314596960): Creator wallet `7xA7A` creates MAO. * In the same block, the creator uses a script to generate fresh wallets (e.g., `712nX`, `6f Yzn`, `4hZpo`) to purchase MAO. * *Result:* Artificially inflated price and concealed creator position. * **Stage 2: Sniper Bot Front-Running** * **[4 Blocks + 1 Second after Launch]**: Sniper wallet `EW6Dk` front-runs retail traders. * *Result:* Sniper secures a low entry price via speed advantage; slightly inflates coin price. * **Stage 3: Comment Bot Activity** * **[15:10:17 – 16:26:42]**: Bots post fabricated messages (e.g., "SENDOOR") to suggest active community communication and lure uninformed traders. * **Stage 4: Bump Bot Activity** * **[15:37:36 – 16:40:45]**: Bump bot `4h7Lk..` repeatedly buys and sells the exact same amount of MAO. * *Result:* Each transaction bumps MAO to the front page of Pump.fun. * **Stage 5: Rug Pull** * **[16:40:54]**: Creator (`7xA7A`) and launch bundles sell holdings for significant profit. * Price drops sharply within one minute. * Sniper detects the drop and exits with moderate profit. * Retail traders mostly close with a loss. --- ### **5. Analysis of Bot Activity Impact (Figure 7 Data)** * **Launch Bundles:** Projects generally show slightly higher maximum returns but **shorter dump durations** (creators dump quickly to profit). * **Sniper Bots:** Most projects have them; performance negligible between those with/without. * **Bump & Comment Bots:** Projects with these show **significant increases** in both maximum returns and dump duration (due to increased exposure and fabricated community). --- ### **6-8. Proposed Multi-Agent Framework** **System Overview:** A framework allowing agents to learn via few-shot Chain-of-Thought (CoT) prompting to make decisions on wallet and coin selection. **The Four Agents:** 1. **Trader Evaluation Agent:** Identifies KOL (Key Opinion Leader) wallets. * *Criteria:* Consistent profitability, low-frequency trading. * *Metrics:* Total Profit, Profit Std, Average Transaction Number. 2. **Wealth Agent:** Manages cash allocation and decides if a copy trade is feasible based on balance. 3. **Meme Evaluation Agent:** Identifies high-potential "farming" coins. * *Inputs:* Transaction indicators, candlestick charts, user comments. * *Assessment:* Looks for absence of creator bundles, presence of bump bots (visibility), and high organic activity. * *Visualization:* Prefer "healthy" charts with fluctuations (organic) over charts with few large green candles (pre-dump). 4. **DEX Agent:** Executes trades via smart contracts. **Detection Algorithms:** * **Algorithm 1: Bundle Bot Detection** * Aggregates transactions by block. * *Launch Bundle:* Creator transfers funds to txn makers in the launch block. * *Buy Bundle:* All transactions in a non-launch block are buys. * *Creator-Funded Buy Bundle:* Buy bundle where creator funded the wallets. * *Sell Bundle:* All transactions in a non-launch block are sells. * **Algorithm 2: Bump Bot Detection (Wash Trading)** * Calculates "Wash Trading Score": Ratio of matched buy-sell pairs of identical amounts to the trader's net position. * Threshold ($c$): If score $\ge 50$, the trader is flagged as a bump bot. --- ### **9. Experiment Methodology & Performance** * **Data Source:** Flipside (Solana historical data). * **Dataset:** First 1,000 meme coins migrated from Pump.fun following the launch of `$TRUMP` on **[2025-01-17 at 14:01:48]**. * **Scope:** Data includes transaction/transfer records from launch up to 12 hours post-migration. * **Agent Performance:** * Agents using transaction data achieve **high precision** but **low recall** (conservative approach). * Predictive power peaks around the **30-minute** mark. * Agents relying only on technical or comment data are also overly conservative. Based on the provided text segment, here are the detailed notes and important points regarding the study on the multi-agent framework for meme coin copy trading: ### **Performance Analysis** * **Overall Agent Performance:** * The agent that combines **all data sources** achieves the strongest results across accuracy, precision, recall, and F1 score. * It effectively filters out poor opportunities while capturing promising ones. * Performance remains stable across various post-migration time intervals. * **Wallet Evaluation Agent Performance (Section 9.3):** * **Focus:** Greater emphasis is placed on **precision** rather than recall due to the massive volume of traders involved in meme coins. * **Precision Rate:** Approximately **70%** (specifically 4,773 correct identifications out of 6,879 predicted). * **Conclusion:** The agent successfully predicts future wallet profitability using historical trading features. * **Classification Confusion Matrix (Table 3 Data):** * **True Positives (Predicted True / Actual True):** 4,773 * **False Positives (Predicted True / Actual False):** 2,106 * **False Negatives (Predicted False / Actual True):** 238,169 * **True Negatives (Predicted False / Actual False):** 369,282 * **Total Wallets Analyzed:** 614,330 * **Meme Evaluation Agent Performance (Table 2 Data):** * Performance is broken down by data modality (Transaction, Technical, Comment, All) and time intervals (1 min to 10 hours). * **"All" Modality (Combined):** Achieves the highest scores. * *Peak Precision:* **73.28%** at the 1-hour interval. * *Peak F1 Score:* **0.6197** at the 1-hour interval. * **Comment Modality:** Shows the lowest performance (F1 scores between 0.08 and 0.16), indicating less predictive power on its own compared to transaction or technical data. ### **Framework Overview (Conclusion)** * **Methodology:** An explainable, multi-modal, multi-agent framework utilizing **few-shot Chain-of-Thought (CoT)** prompting and algorithmic bot detection. * **System Architecture:** Decomposes trading into four specialized agents: 1. Meme Evaluation 2. Trader Evaluation 3. Wealth Management 4. Order Execution * **Goal:** Mimic the workflow of professional asset managers. * **Empirical Results:** * Tested on **1,000 meme coin projects**. * Consistently achieved **70–73% precision**. * **Financial Impact:** Selected Key Opinion Leader (KOL) wallets generated over **$500,000 in profit**. ### **Prediction Model Features (Table 1)** The model utilizes specific features categorized by notification and description: * **Average Return:** Mean return across previous 'x' participated meme coins. * **Number of Trades:** Total executed trades. * **Return Standard Deviation:** Deviation across all participated coins. * **t-stat:** Statistic of the trader's mean return. * **Time Since Last Trade:** Seconds elapsed since the trade prior to the current coin. * **Time Since First Trade:** Seconds elapsed since the trader's very first trade. * **Bot Detection:** A dummy variable (0 or 1) indicating if a bot (Bundle, Sniper, or Bump) is detected. * **Candlestick:** Chart data at the time of the first trade. * **Comments:** Comment history at the time of the first trade. ### **System Instructions & Prompts (Appendix A)** **Meme Evaluation Agent** * **Role:** Professional Meme Coin Analyst. * **Task:** Assess "farming potential" (sustainable upside potential) based on candlesticks, transactions, and comments. * **Inputs Provided to Agent:** * Bundle types: Launch, Creator-funded, Buy, Sell. * Bot presence: Bump Bot. * Metrics: Pre-migration duration, Unique traders, Transaction count, Holding centralization. * Comment History. * **Output Format:** JSON containing "reasoning" and a boolean "good_farming". **Wallet Evaluation Agent** * **Role:** Professional Meme Coin Wallet Analyst. * **Task:** Assess if a wallet is appropriate for copy trading based on the past **50 migrated meme coins**. * **Inputs Provided to Agent:** Total Profit, Total Profit Std, Average Transaction Number, Transaction Number Std, Total Tokens Participated. ### **Chain-of-Thought (CoT) Reasoning Examples** **Wallet Evaluation Logic:** * **Positive Evaluation (Good Wallet):** High total profit + High standard deviation (multiple profitable trades) + High transaction count (active) + High token count (diversified). * **Negative Evaluation (Bad Wallet):** Low profit + Zero standard deviation + Low transaction count + Low participation (lack of diversification). **Meme Evaluation Logic (Specific Scenarios):** * **Scenario 1 (Bad):** * *Condition:* Single green candle from launch to migration. * *Reasoning:* Suggests early participants plan to dump post-migration. * **Scenario 2 (Good):** * *Condition:* Healthy pre-migration duration, fluctuations, no creator-funded bundles, positive community sentiment (e.g., "Everyone's shipping"). * *Reasoning:* Indicates active trading and lack of manipulation. * **Scenario 3 (Bad):** * *Condition:* **Creator-funded Bundle: Yes**. * *Reasoning:* Creator accumulated tokens before migration; signals intent to dump, undermining sustainability. * **Scenario 4 (Bad):** * *Condition:* High "Buy Bundle" + Presence of "Bump Bot". * *Reasoning:* Significant bot activity suggests a price spike followed by an immediate decline. * **Scenario 5 (Bad):** * *Condition:* Low pre-migration duration, low unique traders, medium centralization, **Empty** comment history. * *Reasoning:* Lack of community interaction and poor distribution leads to poor farming potential.