On decentralized exchanges, anyone can become a market maker by depositing two tokens into a pool. The pool trades automatically against all comers, and depositors earn a fee on every trade. The catch is inventory risk: when traders mostly sell one token into the pool, depositors end up holding more and more of an asset that is falling in price. In newly launched tokens this risk is often engineered. Insiders spread trades across many wallets they control and trade with themselves ("wash trading"). This makes volume and fees look healthy and the price look strong. Then they sell into the liquidity that the fake activity attracted. Liquidity-provision models today cannot see this. They read price and volume at face value and assume prices follow a random walk that nobody is pushing.
This note proposes a simple test of a specific idea. A liquidity provider should measure order flow per resolved economic entity, meaning clusters of wallets linked by shared funding, bundled transactions, and identical transfers, rather than per address. Fed into an otherwise standard risk-averse policy, a few such entity-level signals should cut the losses that come from insider dumps. They should do this without giving up the real fee income that wash traders, ironically, pay to the pool. We test this on real transaction replays of tens of thousands of Solana memecoins and on Ethereum pools, not on simulated random-walk prices.
Core Claim
Fabricated volume has a structural signature that price and volume alone cannot reveal. In healthy markets, trade count grows in proportion to the number of distinct counterparties. Collusive trading packs many trades into a small, closed set of partners. Degree-Strength Correlation Reveals Anomalous Trading Behavior found that on Chinese equities, trade count scales roughly linearly with counterparty count in healthy stocks but superlinearly (exponent about 1.122) in manipulated ones. Distinguishing manipulated stocks via trading network analysis found that the correlation between returns and the seller-to-buyer ratio falls below 0.2 under manipulation. On blockchains, wallets can be split almost for free, so these statistics only mean something after wallets are merged into entities. MemeTrans does this merging using atomic transactions, funder flows, and transaction bundles from the Jito block engine. It finds true top-10 concentration about 24% higher than it appears. Enhancing Meme Token Market Transparency does the same with community detection over fund-flow and behavioral links.
Our claim is that these entity-resolved breadth statistics, computed causally (using only data available at each moment), are the missing state variable for liquidity provision. They should predict the reversal that follows wash-driven run-ups, which Crypto Wash Trading documents on centralized exchanges. They should also let a provider weigh wash-fee income against the risk of a dump.
The new contribution is the coupling of these two pieces. Manipulation detectors so far produce ex-post labels or static risk scores, and never a quoting or withdrawal policy. Concentrated Liquidity Provision: a Reinforcement Learning Perspective and Strategic Liquidity Provision in Uniswap V3 optimize liquidity ranges, but only under exogenous random-walk prices with synthetic order flow. We deliberately treat "multimodal" narrowly. The transaction graph is the added modality alongside the price series. Text and social signals enter only as an ablation arm, because the evidence in hand does not yet justify making them central.
Research Program
The first component is a replay environment built on real swaps. For Solana, we use the post-migration exchange trading in MemeTrans: 187 million of its 218 million transactions, across 41,470 tokens from December 2024 to March 2025. For Ethereum, we reconstruct pool states from Uniswap V2 "Sync" events, following Wang et al., and from Uniswap V3 pool records, using extraction tooling like XBlock-ETH. The provider is assumed to be small relative to the pool. Historical swaps then replay against the pool math, and the provider's share determines its fees and inventory. This keeps the real, manipulated order flow that random-walk simulators throw away. The cost is that counterparties cannot react to the provider, a limitation also acknowledged for historical replay in Latent Market Dynamics. We report results only where the provider's share is small enough for this assumption to be plausible.
The second component is a causal entity resolver. We reimplement the MemeTrans linking rules (funder, bundle, and atomic-transaction links) and the fund-flow community clustering. We then recompute clusters on a rolling basis using only transactions up to each decision time. This guards against an easy-to-miss leak: clusters built from the whole history quietly encode who later dumped.
From these clusters, each decision window yields a small set of features. We compute the degree-strength exponent of the entity-level trading network and the share of volume and fees that comes from entity self-trades or closed cycles. We also compute the return-versus-breadth correlation and net inventory flow from the largest entities.
The third component is the policy. We reuse the risk-averse proximal policy optimization (PPO) setup of Concentrated Liquidity Provision, which handles range resets with gas costs and a risk-aversion penalty. We change only its state, adding the entity features, and we train it on real replays. For constant-product pools the actions are entry, sizing, and exit. For concentrated pools they are range width and center.
A central design point is decomposing fees by entity. Wash trades pay fees to the pool, so blindly withdrawing whenever manipulation appears forfeits real income. The policy is rewarded on net profit and loss: fees plus inventory value change minus gas, with a penalty on conditional value at risk (CVaR, the average loss over the worst tail of outcomes). It therefore has to learn when wash-fee income outweighs the coming dump.
Baselines
The naive reference is passive provision. The provider deposits at migration (or at a fixed date on Ethereum) across the full price range and holds until the end of the window. Every gain is measured against this. A second simple reference is the optimal static allocation and uniform reset rules from Strategic Liquidity Provision in Uniswap V3. These are tuned on the training period and run unchanged on the test period.
The strongest learned competitor for inventory control is the same risk-averse PPO agent with only price, volume, volatility, and inventory in its state. This isolates whether risk aversion alone already handles manipulated pools: such agents widen ranges when risk rises, and could plausibly back off from pump volatility without any graph information. Its reported result is an improvement of about 27 USDC in CVaR at a cost of about 4 USDC in mean profit and loss, in the medium-volatility simulated scenario (σ=0.02, g=4) of Concentrated Liquidity Provision. That figure comes from random-walk simulation. We found no reported result on Lacuna for liquidity-provider profit and loss on real manipulated pool replays, so we rerun this agent ourselves.
The strongest competitor on the manipulation side is the MemeTrans multilayer perceptron (MLP) classifier, trained on its 122 engineered features. We use it as a gate: the provider never enters, or exits immediately from, any token flagged high-risk. On its own token-selection task it cut simulated losses from 60.7% to 26.6%, a 56.1% reduction (MemeTrans). That is a different metric from ours, so we re-evaluate the gate in our replay. The classifier labels 84.13% of migrated tokens high-risk, which means the gate also gives up most of the available fee income. That trade-off is precisely what we measure.
We add the online-learning market maker with a hard inventory cap from Adaptive Market Making with Inventory Constraints via Online Learning. It needs no assumptions about the price distribution, which makes it the natural model-free competitor.
The key ablation computes every graph feature on raw addresses instead of resolved entities. If entity resolution is the active ingredient, the address-level version should lose most of the gain, since splitting wallets is exactly what defeats address-level statistics.
All methods share the same token-time splits: train on December 2024 through January 2025, validate on February, test on March. They also share the same gas and fee model, the same number of training environment steps for the learned agents, and the same hyperparameter search budget. Beating a baseline means better test-period CVaR at matched or better mean profit and loss, with bootstrap confidence intervals over tokens that exclude zero.
Why not just…?
Why not just gate liquidity with the MemeTrans classifier? This is the real threat, and the experiments are built to settle it. The classifier scores tokens around launch, so it cannot react to manipulation that starts weeks after migration. It also cannot price the wash-fee income a provider gives up by staying out. We evaluate the dynamic policy on tokens split by whether the gate lets them through, and on Ethereum pools where launch-time features do not exist. If the gate matches our policy on CVaR and profit and loss in both groups, we will conclude that a static launch classifier is enough. The research question then narrows to timing within tokens the gate admits.
Why not just rely on risk aversion, since risk-averse agents already widen ranges in volatile conditions? Pump-and-dump price paths often look like ordinary rising markets with moderate volatility until the moment of the dump. A volatility-driven agent will tend to treat a wash-driven rise as real demand, which is the tension flagged between Crypto Wash Trading and random-walk liquidity models. The price-only PPO baseline tests this directly. We additionally report losses in the 24 hours before the largest entity-level sell-off in each token. That window is where the two agents should differ if the claim is right.
Why not just apply a fixed rule, such as withdrawing when the return-breadth correlation drops below 0.2? Thresholds like this came from 2004 equity data in Distinguishing manipulated stocks via trading network analysis. Conservative motif thresholds only lower-bound wash activity, as Unveiling the Paradox of NFT Prosperity notes. We include the rule, applied to entity-level features, as a baseline. If it matches the learned policy, that is a cheaper and more transparent result worth reporting as the main finding.
Why not just fuse news and social text, as in multimodal trading agents? Financial Forecasting from Textual and Tabular Time Series shows that text helps on earnings with dense disclosures. MemeTrans found engineered tabular features beat Transformers on short, noisy memecoin sequences. Agents like FinAgent report sensitivity to tools and model randomness in high-frequency crypto trading. We therefore add text only as an arm, using on-chain token metadata and launch text where available, and report its marginal contribution rather than assuming one.
Proposed Experiments
The primary experiment compares all baselines on March 2025 Solana replays. It reports net profit and loss decomposed into genuine fees, wash fees, inventory loss, and gas, plus 5% CVaR and peak inventory held in tokens that later fall more than 80%.
Four decisive checks follow. The first is the address-versus-entity ablation: the claim fails if address-level features recover most of the gain. The second is the classifier-stratified analysis described above. The third checks the mechanism: entity-level self-trade share in a window should predict the following window's return sign and the provider's inventory loss beyond what price and volume predict. We test this with out-of-sample regressions, and the claim fails if the incremental explained variance is negligible. The fourth is a transfer test. The policy is trained on Solana and evaluated on Ethereum Uniswap V2 and V3 pools without retraining the entity resolver's rules. This checks whether the signal generalizes or just memorizes launchpad-specific bundle patterns.
As a sanity check on the labels, the entity-level degree-strength exponent should be higher on MemeTrans high-risk tokens than on low-risk ones, echoing the 1.122 versus 1.0 gap in equities. If even that fails, the signal is not there to exploit.
References
- Degree-Strength Correlation Reveals Anomalous Trading Behavior
- Distinguishing manipulated stocks via trading network analysis
- MemeTrans: A Dataset for Detecting High-Risk Memecoin Launches on Solana
- Enhancing Meme Token Market Transparency: A Multi-Dimensional Entity-Linked Address Analysis for Liquidity Risk Evaluation
- Crypto Wash Trading
- Concentrated Liquidity Provision: a Reinforcement Learning Perspective
- Strategic Liquidity Provision in Uniswap v3
- Cyclic Arbitrage in Decentralized Exchanges
- XBlock-ETH: Extracting and Exploring Blockchain Data From Ethereum
- Latent Market Dynamics: A World Model Framework for Agentic Prediction Markets
- Adaptive Market Making with Inventory Constraints via Online Learning
- Unveiling the Paradox of NFT Prosperity
- Financial Forecasting from Textual and Tabular Time Series
- A Multimodal Foundation Agent for Financial Trading: Tool-Augmented, Diversified, and Generalist