Lab: real data, both comparisons
Backtest summary
P* = 30%, floor 5 bp. Two fair comparisons against a static fee: at equal time-average fee, and at equal cost to traders (volume rises with volatility).
| Period | ARB change, equal time-average fee | ARB change, equal cost to traders | P_trade predicted / observed | Blocks |
|---|---|---|---|---|
| Feb 2026 storm | -15.3% | -2.9% | 0.299 / 0.279 | 28,716 |
| Oct 2026 calm | -5.5% | -2.3% | 0.156 / 0.171 | 21,557 |
| Oct 2025 to Oct 2026 (1-min data bridged) | -26.9% | -12.7% | 0.243 / 0.236 | 2,627,970 |
- LP gain against a static 5 bp pool in the one-year backtest (1-minute data bridged to 12 s blocks), full-range ETH, across 5 retail scenarios: -0.10% to +0.70% of capital per year (-$981 to $7,024 a year per $1M of liquidity). In the main scenario (an aggregator routes retail between clim and a static 5 bp pool four times deeper), ETH gains +0.39% a year, about 53% of it earned in the five most turbulent weeks; the same scenario on an asset twice as volatile (the same year with every return doubled) gains +0.40% a year.
- The model predicts how often arbitrage happens; it underestimates how much it takes: observed ARB/LVR is 1.29 to 1.33 times the model.
- A volatility computed inside the pool gets 54% to 103% of the same gain (above 100%: it did slightly better in one sample): Chainlink CRE is here for robustness (four venues must agree), not accuracy.
Raw data: /data/lab/summary.json
Replay: 2026-02-04 12:00-16:00 UTC
12:00 to 16:00 UTC on 4 February 2026, P* = 30%, pool S fixed at 15.03 bp, V's average over the window.
Pool V's loss to arbitrage, against pool S
18.6% less
other 4 h windows of the storm: 18.4% less to 3.1% more
σ
34% to 296%
Fee on V
5.0 to 32.4 bp
pool S: 15.03 bp
Blocks arbitraged on V
29.8% / 28.5%
predicted / observed
We chose this window while designing clim, around the sharpest rise in volatility. None of the storm's 92 four-hour windows does better (best: 18.4% less); the median is 2.1% less, and V lost less in 66 of 92.
ETH/USD
σ from the desk (annualized %)
Fee (bp): pool V follows the storm, pool S stays at V's average
Lost to arbitrage, cumulative, net of the fees arbitrageurs paid (USD per $1M of liquidity, lab)
Raw data: /data/lab/replay-2026-02-04.json