Skip to content
Crypto.world

Research

MELT: A Behavioral Trace Dataset for High-Risk Memecoin Launch Detection

Launchpads like Pump.fun let anyone create a memecoin in minutes, and existing rug-pull detectors do not work there because the launchpad — not the creator — owns the liquidity pool.

MELT: A Behavioral Trace Dataset for High-Risk Memecoin Launch Detection
arXiv
2602.13480
Published
2026-02-13
Authors
Sihao Hu, Selim Furkan Tekin, Yichang Xu, Ling Liu

AI summary

Overview

Research area: Blockchain security and machine learning — specifically memecoin launch analysis and rug-pull-adjacent threat detection on Solana.

Technical level: Intermediate. The paper assumes some familiarity with blockchain concepts (accounts, transactions, DEX liquidity pools) and standard ML benchmarking, but explains launchpad mechanics from first principles.

Scope: The paper introduces MELT, a dataset of 41,470 Solana memecoin launches with over 200 million parsed transactions, bundled-account traces, 122 behavioral features and risk labels, and benchmarks ML models on the task of predicting high-risk launches before DEX trading begins.

What This Paper Is About

Launchpads like Pump.fun let anyone create a memecoin in minutes, and existing rug-pull detectors do not work there because the launchpad — not the creator — owns the liquidity pool. The real risk instead comes from insiders who buy tokens at the cheapest bonding-curve tiers during the launchpad sale and dump them after migration to a DEX, often hiding their holdings across multiple coordinated accounts. The authors build a dataset that reconstructs these "behavioral traces" (how insiders accumulate, coordinate and unwind positions) and use it to train models that flag high-risk launches before public trading starts.

Key Contributions

  1. MELT dataset: The first behavioral trace dataset for high-risk memecoin launch detection on Solana, covering 41,470 memecoins that completed a four-stage lifecycle (creation → launchpad sale → migration → DEX trading), with 218,530,716 transactions parsed into typed behavioral records (create, buy, sell, wash trade, transfer).
  2. Bundle-trace data: A method for linking accounts likely controlled by the same entity, built from three sources — multi-account co-purchase in a single transaction, shared fund-flow (funder addresses, excluding centralized exchanges), and crawled Jito bundle IDs — covering 22,668,222 pre-migration buy/sell transactions.
  3. Feature and label suite: 122 behavioral features in five groups (contextual information, holding concentration, market activity, bundle statistics, time-series), all computed strictly from pre-migration data, plus a configurable two-layer risk annotation scheme (a statistical threshold on post-migration price plus manual inspection of ambiguous cases).
  4. Benchmark and downstream evaluation: A comparison of tabular models (LR, RF, LightGBM, XGBoost, MLP) and time-series models (TCN, LSTM, GRU, Transformer), an ablation study, and a simulated memecoin-selection strategy showing that model-guided selection reduces investment loss by up to 34 percentage points.

Main Findings

  • Concealed ownership is pervasive: On average, 36.5% of token supply appears held by independent accounts but is actually controlled by coordinated bundled accounts; 28.13% of holders fall into bundles. Per source, fund-flow contributes the most (22.57% of holders, 28.22% of supply), followed by Jito bundles (10.39% / 15.96%) and co-purchase (6.46% / 9.16%).
  • Bundling discriminates risk: After merging bundled accounts, the top-10 holder share rises by 24 percentage points for high-risk memecoins versus 6 points for low-risk ones (median increases of 24, 9 and 6 points for high-, medium- and low-risk tokens respectively).
  • High-risk launches look different during the sale: They have shorter launchpad sale durations, fewer holders at migration, fewer buy transactions but larger average buy volumes, and more concentrated early accumulation — the first 10 and 20 buyers of high-risk tokens hold 17 and 19 percentage points more supply than those of low-risk tokens.
  • Developer pre-acquisition is near-universal: 98.7% of "create" events co-occur with a developer buy in the same transaction.
  • Transaction mix: Of 30,833,503 pre-migration transactions, 73.5% are normal buys or sells, 21.4% are wash trades and 4.9% are transfers. The post-migration one-hour window contains roughly six times as many transactions as the entire pre-migration period.
  • Risk labels are skewed: 84.13% of launches (34,890) are labeled high-risk, 11.34% (4,702) medium-risk and 4.53% (1,878) low-risk. Within a 20-minute post-migration window, about 73% of memecoins drop below 40% of their migration price; the largest bucket, 24,988 memecoins (60.26%), falls in [0.0, 0.2).
  • Tabular models beat time-series models: MLP is the best single model (AUPRC 0.5729, macro F1 0.6981) versus Random at AUPRC 0.2589. Simple ensembles help slightly (MLP+LSTM reaches AUPRC 0.5827). Among time-series models, LSTM performs best, while the Transformer does not, which the authors attribute to short, noisy sequences lacking long-range dependencies.
  • Market activity features matter most: Removing the market activity group causes the largest ablation drop (AUPRC 0.5729 → 0.5369). Removing bundle statistics (0.5451) hurts more than removing holding concentration alone (0.5693), and removing both together drops AUPRC to 0.5268.
  • Practical loss reduction: Random selection loses 60.71% (top-100) and 64.84% (top-200); MLP-guided selection loses 26.64% and 31.47%, a reduction of 34 percentage points, with precision of 0.8000 and 0.7550 on the selected sets.

Methodology in Plain English

The authors use an event-anchored collection window from December 1, 2024 to March 1, 2025, centered on the $TRUMP launch of January 17, 2025, so the data spans a growth phase, a peak speculation phase (including $MELANIA on January 19 and the inauguration on January 20) and a cooling phase (accelerated by the $LIBRA collapse on February 14, 2025).

They first find memecoins that migrated by tracking Pump.fun creator-account transactions that send SOL to the DEX fee account, then pull token metadata via the Metaplex Token Metadata program-derived address. For each memecoin they collect transactions from creation to migration, plus the first hour after migration.

Raw Solana transactions are low-level program calls with no user-level meaning, so the authors write a parser that identifies the interacting program and counterparties, computes each account's net balance change, and classifies the transaction as create, buy, sell, wash trade or transfer.

To unmask coordination, they cluster accounts into "bundles" using three signals: accounts signing the same purchase, accounts funded by the same address (excluding exchange addresses), and accounts whose transactions share a Jito bundle ID crawled from the Jito Explorer. Overlapping bundles are merged into entities.

Features are engineered only from pre-migration data to avoid label leakage. Labels come from post-migration data in two layers: a statistical rule based on min_price_ratio (the minimum price within 20 minutes of migration divided by the migration price), and manual inspection on GMGN for memecoins with min_price_ratio ≥ 0.5 to catch slow, bot-controlled manipulation. The default scheme labels launches high-risk if min_price_ratio < 0.3 or flagged as manipulated, low-risk if min_price_ratio ≥ 0.7 and not manipulated, and medium-risk otherwise.

For the detection benchmark they drop launches with sale duration under one minute or fewer than 100 holders, which removes 19,835 memecoins (47.8%), 95% of them high-risk, leaving 16,048 high-risk and 5,587 normal examples. They train with class-weighted binary cross-entropy, split chronologically 7:3, tune with five-fold cross-validation and report averages over five runs, using precision, recall, F1 and AUPRC because accuracy is unreliable under imbalance. Finally, they simulate investing in the top-k tokens ranked by predicted risk in descending order, buying at the migration price and averaging losses across 100 randomly sampled sell times within the following hour.

Why This Matters

Impact on research: The paper reframes memecoin fraud detection away from DEX pool events (burns, liquidity withdrawals) toward launchpad-stage behavior, and supplies a public dataset and pipeline for a problem that previously had only economic or sociological descriptions. It also demonstrates that bundle-trace linkage — normally studied for Sybil or money-laundering detection — carries predictive signal for financial risk.

Real-world applications:

  • Pre-trade screening dashboards that warn retail buyers about a memecoin before it starts DEX trading.
  • Wallet or trading-bot integrations that filter or size positions using predicted risk scores.
  • Launchpad and DEX compliance tooling that flags coordinated insider accounts at migration.
  • Blockchain-forensics workflows that group addresses into real-world entities.

Industry relevance: With Pump.fun reported to have issued 12.8 million memecoins by October 2025 — roughly half of all tokens created on Solana — and only about 2% of launches reaching migration, automated early risk filtering addresses a large-scale, high-volume market problem that manual review cannot cover.

Future Directions

  • Extending MELT to unmigrated memecoins, which the authors identify as a limitation; these can still harm retail buyers during the launchpad sale at smaller scale.
  • Applying the bundle-trace methodology to related problems such as Sybil detection and money laundering across the wider Solana ecosystem.
  • Operating the released pipeline continuously to ingest newly launched memecoins rather than a fixed historical window.
  • Improving detection beyond the reported AUPRC ceiling (best result 0.5827) — for instance by exploring representations that go beyond distilled memecoin-level bundle statistics, or by optimizing models directly for loss minimization rather than classification.

Target Audience

Researchers and practitioners in blockchain security, financial fraud detection and applied ML who work with on-chain data; dataset and benchmark builders interested in adversarial user behavior; and quantitatively minded traders or platform engineers who need automated risk screening for launchpad-issued tokens. Readers without blockchain background will need to pick up bonding-curve and DEX liquidity concepts, but the paper defines them.

Authors’ abstract

Launchpads have become the dominant mechanism for issuing memecoins, exposing investors to a new class of high-risk launches that existing rug-pull detection methods cannot capture. We argue that detecting these threats requires structured behavioral traces that underlie raw heterogeneous blockchain data, i.e., how insiders accumulate, coordinate, and unwind positions. To enable such analysis, we introduce MELT (MEmecoin Launch Trace, the first behavioral trace dataset for…

The full abstract on arXiv
Read the original paper