Any backtest can be made to look good after the fact. This page explains the process a strategy has to survive before it manages real money, and shows two candidates that did not survive it.
Give someone a price history and enough parameters, and they will find a strategy that would have performed beautifully on it. That is not a discovery, it is curve fitting: the parameters memorised the noise of that specific period. The strategy then meets new data and falls apart.
This is not a hypothetical failure mode in DeFi liquidity management, it is the normal one. A strategy tuned on the period that made it look best is the default output of an unstructured research process.
Our test method is walk-forward validation. The history is cut into consecutive folds: the strategy is tuned on roughly a year of data, then judged only on the following quarter, which it has never seen. Then the window slides forward and it happens again. Across markets from 2018 onward, that produces 29 independent out-of-sample test periods per pool.
What gets reported is the performance on the unseen quarters only. A strategy cannot look good here by having memorised the past, because the past it was tuned on is excluded from its score.
Every candidate is measured against two mandatory baselines, not against our own strategy alone:
The champion is whatever currently manages capital. Beating it is necessary but far from sufficient. A 50/50 baseline is a deliberately unclever strategy that simply re-centres a position. If a sophisticated candidate cannot beat that, the sophistication is decoration. We also report buy and hold, because a liquidity strategy that loses to holding the tokens has no reason to exist.
Each candidate is tested across several fee regimes, not just the one that flatters it, because a pool that pays generously and a pool that pays poorly are different problems.
The pass and fail conditions are written down before the test runs, along with the hypothesis and what would falsify it. Changing them afterwards requires writing down what changed and why.
The reason is uncomfortable but well established: with results in hand, it is close to impossible not to rationalise. A losing result becomes "the market was unusual that period", a winning one becomes proof. Pre-registration is standard practice in clinical trials and almost absent from trading products, which tells you something about trading products.
The same discipline governs the live pilot: its pass criteria and decision date were fixed before the first cycle ever ran.
In July 2026 we ran this process on two candidates. Both are respectable designs, used by other protocols. Neither was promoted.
Base plus limit.A split position: a main range earning fees, plus the surplus token working as a one-sided limit order. Across 29 out-of-sample folds it beat the strategy that runs our live capital, and still lost to the plain 50/50 baseline at the pool's real fee level. It cleared the bar in only one of three fee regimes tested, which is not enough. Re-running it with real on-chain swap costs did not change the conclusion, so this is structural rather than a modelling artefact: the surplus sitting in the limit leg is capital not earning fees.
Rebalance cooldown. Waiting before re-placing a position that has drifted out of range. As an edge strategy it showed no improvement over the baselines, and was not promoted as one. It is worth being precise here: a cooldown guard does exist in the live path, but for an operational reason, to avoid churning gas on repeated re-placement attempts, and never as a bet on returns. The test confirmed it costs no edge, which is what an operational guard has to prove.
Both remain in the shadow fleet, where the point is to observe the machinery rather than the returns. Base plus limit is genuinely better at staying in range during trends, and that is worth measuring on live data even though it did not win on edge.
Passing the statistical gate is not enough on its own. A candidate then runs in shadow: it decides and records every cycle against live market data, and signs nothing. This stage tests the machinery, whether it behaves sanely, whether its rebalance rate is what was predicted, whether its decisions survive human review.
Only then does anything reach real capital, and it starts deliberately small. Today the shadow fleet runs well over a hundred strategy configurations across dozens of pools. The number promoted to live so far is zero. The bar existing and never being met is the bar working.
Honesty about the limits, because a methodology page that only lists strengths is marketing:
Out-of-sample validation reduces the risk of curve fitting, it does not remove it. Enough candidates tested against the same history will eventually produce one that passes by luck. Historical fee data is itself an estimate, which is why we measure fees on-chain rather than trusting advertised rates. And no amount of backtesting says anything about a market regime that has never happened.
Past performance, including out-of-sample performance, does not guarantee future results. See the risks page for what can go wrong.