LST yield spreads are a window artifact — measure them at a year, not a month

TL;DR — the decision rule

A 30-day realized-yield spread between two large ETH LSTs is not evidence of a yield edge. weETH appeared to out-yield wstETH by 25.6bps on a trailing 30-day read; at 180 days the gap is 5.1bps and at 366 days it inverts to −2.5bps. The lead changed hands 52 times in a year with a median regime of 3 days, and cumulatively wstETH paid more over the full year. Never size an LST switch on a spread measured over less than ~180 days, and quote the 1-year number alongside any short-window figure. Where a small persistent edge does exist, check first whether it is a redemption-driven transfer that will fade: a venue whose exit queue pays leavers the lesser of their request-time amount and their share value hands in-queue yield to stayers, and that flatters the rate exactly while outflows are large.

Question

Refreshing protocols/etherfi-weeth.md, weETH showed 2.48% realized against wstETH’s 2.22% — an apparent 26bps pickup on a token whose restaking sleeve is under 1% of the pool and pays nothing. Either there was an unexplained yield source, or the measurement was wrong. The answer generalises: we hold and compare LSTs routinely, and a 20bps phantom edge is more than enough to justify a bad rotation.

Methodology

Findings

1. The spread decays monotonically with window length

2. It is the peer declining, not the subject rising

Non-overlapping 30-day buckets, first vs last:

Network staking yield is compressing for everyone. weETH’s own rate is flat (2.40–2.66% every month, no trend); it ran below the peer median through 2025 and crossed over in March 2026.

3. The lead changes hands constantly — and wstETH wins the year

Over 122 three-day periods spanning 2025-08-03 → 2026-08-06 (data/weeth-wsteth-3day-1y.csv):

The noise is ten times the mean. Split in halves, the first six months had weETH ahead 39% of periods (mean −9.6bps) and the second six months 72% (mean +4.6bps) — a real recent tilt, but one that still leaves wstETH ahead cumulatively over the full year.

4. The short-window signal is inside its own noise

Daily annualised rates over 150 days: weETH mean 2.451%, stdev 0.400pp; wstETH mean 2.411%, stdev 0.243pp. weETH > wstETH on 83 of 150 days (55%). Standard error on the difference is ±8.5bps at 30 days but ±3.8bps at 150 — so the 150-day estimate of +4.0bps is the defensible one, and the 30-day +25.6bps is mostly one weak month at the peer.

5. Ruled out: fees, cash drag, and one-off accounting

• Fees: Lido getFee() = 999bps at every sampled block over 12 months. Unchanged.
• Cash drag: Lido’s buffered (unstaked) ETH is 0.01–0.13% of pooled ETH; ether.fi’s is 0.34–1.76%. ether.fi carries more idle drag, so this works against the observed spread.
• Lumpiness: if the excess came from one-off accounting during ether.fi’s restaking wind-down, it would arrive in spikes. weETH’s top-5 days are 5.0% of its 150-day total (uniform = 3.3%); wstETH’s are 4.6%. Essentially the same — the excess accrues smoothly.

6. The exit queue transfers yield from leavers to stayers

ether.fi’s WithdrawRequestNFT records both an ETH amount fixed at request time and a share count; getClaimableAmount pays the lesser of the two. Yield earned while a request sits in the queue is therefore forfeited by the leaver and accrues to remaining holders. Live open requests at block 25,695,733:

Small in a calm queue (hours of accrual). At the spring run’s scale — ~1.4M ETH exiting at a ~4.9d median wait — the transfer is worth roughly 5bps annualised to stayers. It appears exactly when outflows are large and switches off when they stop.

Decision framework

1. Quote the 1-year spread beside any short-window spread. If they disagree in sign, the short window is noise.
2. Require ~180 days before acting on an LST spread. Below that, the standard error on the difference exceeds most plausible real edges.
3. Attribute before rotating. Work through fees, idle-ETH drag, lumpiness, and queue mechanics before concluding “better operators.” Only the last of those is durable.
4. Treat redemption-driven yield as temporary. A lesser-of-amount-or-shares queue rule flatters the rate while a venue is bleeding deposits — the yield is being paid by the people leaving, and it ends when they finish.
5. Read the field, not the pair. A stable cross-venue ordering (here cbETH > weETH > wstETH > osETH ≈ rETH) is operator-quality signal; a pair that crosses over is usually one side moving.

This stops applying when: a venue changes its fee schedule or queue accounting; a protocol adds a genuine non-staking revenue leg (real AVS fees, restaking that actually pays); or Ethereum issuance changes enough that cross-venue comparison needs re-basing.

Caveats

We measured exchange-rate growth, which is net of each protocol’s fees — so this compares what a holder receives, not validator-level gross performance, and cannot separate “better operators” from “lower fees.” Attributing the residual ~4bps to operator quality is inference from the stable cross-venue ordering, not a direct measurement of attestation effectiveness or MEV capture; a validator-level study using beacon-chain data would settle it. The 5bps queue-transfer estimate assumes the eval’s reported 4.9d median wait and uniform exit sizing, neither of which we re-derived here. osETH, rETH and cbETH were sampled only on the monthly grid, not daily.