CLMM-Bench: out-of-sample audit of Uniswap v3 LP strategies

What it is

A benchmarking framework for concentrated-liquidity market-making (CLMM) strategies on Uniswap v3, built to answer one question honestly: does strategy selection on historical on-chain data survive out-of-sample?

• Data: on-chain swap and liquidity events for five pools (including ETH/USDC at the 0.05% and 0.30% fee tiers) pulled via Dune SQL, stored as Parquet.
• Architecture: modular strategy plug-in interface; every configuration runs through the same walk-forward harness.
• Validation: walk-forward selection, probability of backtest overfitting (PBO), Deflated Sharpe ratio (DSR), regime-aware evaluation.

Findings

• Walk-forward-selected configurations from a 40-point grid lost 6.08% and 2.93% per three-month window in ETH/USDC — none beat a never-rebalanced position.
• Overfitting diagnostics are not cost-invariant: raising the reset charge from 0 to 0.25% of capital moved PBO from 0.454 to 0.114 while DSR collapsed from 0.557 to below 0.001. The “less overfit” reading and the “no skill” reading arrive together.

Why it matters

Most LP-strategy papers report in-sample Sharpe. This shows what happens when you hold the selection procedure itself to account, and why cost assumptions have to be reported alongside any overfitting statistic.