CLRE Concentrated-Liquidity Rebalancing Study (v3)

TL;DR — should you auto-rebalance a tight Uniswap v3 position? Mostly no.

• At ±5% width, exit-triggered re-centering needs a 111.6% fee APR just to break even on a trending 90-day window (108.0% on a disjoint oscillatory one). Very few pools ever pay that.
• Explicit costs (gas + swap fees) are only 3–4% of the damage. The rest is the policy itself: re-centering at its own trigger scale is structurally short realized variance — a discrete LVR maximiser (Milionis et al. 2022), invariant to path structure. (v2's "short mean-reversion" label was falsified by the OU control and retracted; the log keeps the retraction.)
• The thing that IS regime-dependent is the static tight range (33.5% trending → 1.2% oscillatory), not the re-centering decision.

Question: how should a Uniswap-v3-style LP position be rebalanced, and what does each policy actually cost?

Full result: threshold ±5% loses on 200 of 200 pure mean-reverting (OU) synthetic paths, median break-even 116.5% [5–95pct: 89.7–141.3].

Read first: final/CLRE_v3.1_Rebalancing_Study_Marco_Amendola.pdf (4 pages).

Reproduce

python3 tests.py              # 9/9 sanity suite — required before any run
python3 reproduce_gate.py     # replays published numbers from archived inputs
python3 adversarial_check3.py # every PDF number verified against runs/*.json

Archived hourly inputs for every window ship in runs/ — the gate runs offline. Synthetic study: seed 42, parameters logged before the run.

Structure

• clre.py — engine: exact v3 position math, policies, costs, evaluation
• tests.py — nine-test validation suite
• reproduce_gate.py, adversarial_check3.py — gates
• run2.py, scan_windows.py, build_pdf3.py — runs and artifact build (superseded scripts in archive/)
• make_figure.py — README figure
• runs/ — archived inputs + full outputs (run002/003/004/004b, gate, scan, fee anchor)
• research_log.md pre-registered predictions (including the failed ones), two v2 errata, belief updates. The credibility artifact.
• adversarial.md both adversarial passes, findings triaged, repairs
• brief.md origin and acceptance criteria

Method discipline

Predictions pre-registered before every run (Run 002: 2/4 failed — the failures produced the main finding; Run 003: 5/5; Run 004: 4/4). Reproduction gate before any new run. Fee income inverted (break-even APR), not assumed; pool-level anchor measured (DeFiLlama series; Dune query 7923963 published). Known limitations and the production path per gap are stated inside the PDF.

What this means in practice

The break-even numbers above are a bar. A pool clears it only if its sustained fee APR exceeds the break-even for your width and regime — and pools paying 100%+ sustained are rare and usually paying you for a risk. Before enabling any auto-rebalancing policy on a position: (1) measure the pool's realised fee APR over a window at least as long as your expected holding period (anchor query: Dune 7923963), (2) compare it to the break-even for your width from this study, (3) if it doesn't clear, a static range — re-chosen only on regime change — dominated every re-centering policy tested here.

Live monitor: CLRE Break-Even Monitor on Dune — which top pools currently clear the bar, refreshed weekly.

Marco Amendola · July 2026 · on-chain research under Xpertknight · not investment advice