TL;DR
This paper introduces a dynamic model to determine the "sweet spot" for AMM trading fees by analyzing the competition between decentralized and centralized exchanges. The authors demonstrate that liquidity providers (LPs) maximize profits by slightly undercutting centralized exchange (CEX) costs during normal periods while using high-fee "thresholds" to survive bouts of extreme volatility.
What Problem Does This Paper Address?
Passive Liquidity Providers (LPs) in Automated Market Makers (AMMs) like Uniswap face a persistent problem: they often lose money. While they earn trading fees, they frequently suffer from Loss-Versus-Rebalancing (LVR)—a form of adverse selection where arbitrageurs exploit "stale" prices on the AMM before the pool can update to reflect the global market price (usually discovered on a CEX).
Existing research often assumes CEXs are frictionless or that AMM traders are "noise traders" who don't care about price. This paper fills the gap by modeling a more realistic world where AMMs compete directly with CEXs for "fundamental" traders who route their orders to whichever venue offers the best total price (including fees and slippage). The core challenge is finding a fee that is low enough to attract this profitable volume but high enough to offset the inevitable losses to arbitrageurs.
Methods
The authors develop a dynamic reduced-form model of CEX-DEX interaction. Key components include:
- The Model: A Constant Product Market Maker (CPMM) operating alongside a CEX. The CEX price follows a geometric Brownian motion.
- Agent Types:
- Arbitrageurs: Profit-seekers who align AMM prices with CEX prices.
- Fundamental Traders: Traders with exogenous demand who "smart-route" orders to the cheapest venue.
- Simulation Framework: A large-scale, parallelized Monte Carlo simulation that explores a "hypersurface" of market conditions (volatility, demand, and fees).
- Hedged PnL: To isolate the effectiveness of the fee structure, the authors focus on "hedged PnL," which removes the risk of the underlying asset's price movements, leaving only the profit/loss from trading activity and fees.
- Calibration: They use a grid-search optimization (via the Optuna library) to fit their model parameters to real-world price ratio distributions.
Datasets
The study uses two primary sources of high-frequency data:
- Uniswap v2 (DEX): Minute-by-minute reserve balances for the ETH/USDC pool extracted via Dune Analytics for January 2025.
- Binance (CEX): Minute-level historical price data for the ETH/USDC pair for the same period.
By comparing these, they calibrate the "effective" CEX trading cost ($\eta^0$), which includes not just nominal fees but also bid-ask spreads and slippage.
Key Insights and Ideas
- The Undercutting Rule: The optimal AMM fee is almost always strictly lower than the effective CEX trading cost. If the AMM fee is too high, fundamental traders stay on the CEX, and the AMM is left only with "toxic" arbitrage flow.
- Fundamental Flow as a Stabilizer: Unlike "noise" traders, fundamental traders in this model are price-sensitive. Their activity creates a "reverting drift" that helps keep the AMM price aligned with the CEX, reducing the window of opportunity for arbitrageurs.
- The Volatility Threshold: As volatility increases, the "profit region" for LPs shrinks. At a certain point, the only way to protect the pool is to set a prohibitively high fee—effectively a "kill switch" that halts trading during market chaos.
- Robustness of Constant Fees: While a perfectly dynamic fee is ideal, the authors find that a well-chosen constant fee (around 70-80% of CEX costs) is remarkably robust across a wide range of "normal" market conditions.
Key Results
- Calibration: For the ETH/USDC pair in early 2025, the authors found the effective CEX cost to be between 9 and 12 basis points (bps).
- Theoretical vs. Actual: While Uniswap v2 uses a standard 30 bps fee, the model suggests the optimal fee for that environment should have been closer to 5.42 bps.
- PnL Improvement: In historical backtests, switching from the standard 30 bps fee to the optimized 5.42 bps fee turned a negative hedged PnL into a significantly positive one. This is because the lower fee "unlocked" a massive amount of profitable fundamental volume that previously stayed on the CEX.
- Dynamic Fee Schedule: The authors propose a "threshold-type" dynamic fee: a low fee (e.g., 5 bps) for standard volatility and a high fee (e.g., 100+ bps) for extreme volatility.
Limitations and Future Work
- Model Parsimony: The study uses a Uniswap v2-style CPMM. While the authors argue the core trade-offs apply to concentrated liquidity (Uniswap v3), the specific math for "tick-based" liquidity is more complex and not explicitly modeled here.
- Gas Fees: While adjusted for in calibration, gas fees are not a first-class citizen in the core simulation, which might affect results for smaller trade sizes.
- Future Directions: The authors suggest that Uniswap v4 "hooks" provide the perfect infrastructure to implement the threshold-based dynamic fees suggested by their research. They also plan to explore a rigorous continuous-time limit for trading frequency in future work.
Summary updated