The question
Our first pass at this market, over roughly one week, found three things: the BTC and ETH perpetual bases behaved like efficient random walks, SOL’s basis genuinely mean-reverted, and on all three the funding rate rose when the perpetual traded rich to spot. We flagged the SOL result openly as single-window — “genuine in this window, replication ongoing, not a settled law.” This note keeps that promise. The sharp question: was SOL’s mean-reversion a property of SOL — a less-perfectly-arbitraged perpetual — or a property of that one week?
Data
Six one-week windows (Monday 00:00 → Monday 00:00 UTC) spread across January–August 2026, so they span different volatility and funding regimes, for BTC, ETH and SOL. Each window uses public one-minute perpetual and spot prices plus the eight-hourly funding settlements from public exchange APIs; both legs are quoted in USDT, so no de-margining is needed. The windows were fixed by a mechanical, outcome-blind rule — the most recent completed week as the canonical window, then five more at six-week steps back — chosen before any basis data was examined, so none was hand-picked on volatility or result.
Method
We re-ran the identical study: reconstruct the basis in basis points, guard it with the same robust outlier filter, and run the same increment-shuffle surrogate test per asset, with thresholds that auto-scale to each asset’s basis size and every parameter frozen in advance (read from the original study’s code, not re-tuned). We added one pre-registered check against a specific confound: because an eight-hourly funding settlement can induce a mechanical oscillation in the basis that mimics mean-reversion, we also ran the reversion test on a settlement-masked basis — dropping every minute within ±30 minutes of a funding stamp — so that any surviving reversion is reversion beyond the funding clock. The whole harness was self-tested on synthetic data, including a control proving the mask removes settlement-only reversion, before any live window ran.
Findings
Two of the original three findings replicate cleanly; the flashy one does not.
The funding-tethering replicates everywhere. Across all six windows and all three assets — eighteen asset-windows in total — the funding rate correlates positively with the basis (+0.14 to +0.77): funding rises when the perpetual is rich to spot, exactly as the tethering mechanism predicts, in every regime we sampled.
The majors are efficient in every window. The BTC and ETH bases are random walks in all six windows — no mean-reversion beyond chance anywhere. This doubles as an internal control: a pipeline that “found structure” everywhere would be suspect, and this one does not.
SOL’s mean-reversion is window-specific. SOL reverts only in the one window that overlaps the original study’s week (real 11.6% convergence versus a 2.7% surrogate) — and in none of the five independent windows, where its basis is an ordinary random walk. The apparent structural edge was a property of that week, not of SOL.
| Window | Asset | Basis σ (bps) | Mean-reversion vs surrogate | Funding ↔ basis |
|---|---|---|---|---|
| W1 (Aug 17–24) | BTC | 2.3 | none — random walk | +0.54 |
| ETH | 1.5 | none — random walk | +0.48 | |
| SOL | 3.0 | genuine — 11.6% vs 2.7% surrogate | +0.71 | |
| W2 (Jul 6–13) | BTC | 0.8 | none — random walk | +0.36 |
| ETH | 0.7 | none — random walk | +0.58 | |
| SOL | 1.4 | none — random walk | +0.52 | |
| W3 (May 25–Jun 1) | BTC | 0.8 | none — random walk | +0.47 |
| ETH | 0.7 | none — random walk | +0.41 | |
| SOL | 1.5 | none — random walk | +0.59 | |
| W4 (Apr 13–20) | BTC | 0.9 | none — random walk | +0.56 |
| ETH | 1.2 | none — random walk | +0.77 | |
| SOL | 1.7 | none — random walk | +0.63 | |
| W5 (Mar 2–9) | BTC | 0.9 | none — random walk | +0.14 |
| ETH | 1.1 | none — random walk | +0.68 | |
| SOL | 1.5 | none — random walk | +0.58 | |
| W6 (Jan 19–26) | BTC | 0.7 | none — random walk | +0.54 |
| ETH | 0.9 | none — random walk | +0.58 | |
| SOL | 1.2 | none — random walk | +0.47 |
Two details sharpen the reading. SOL’s reverting window survives the settlement mask (still about 12% versus a ~3% surrogate with settlement-adjacent minutes removed), so the original result was not a funding-clock artifact — it was a real effect that simply does not recur. And SOL’s basis is materially tighter in the five independent windows (1.2–1.7 bps) than in the overlapping one (3.0 bps): the “less-perfectly-arbitraged” condition was itself specific to that week.
What this means
We pre-registered a replication of our own most edge-like finding, and it did not replicate. That is the result worth publishing. Single-window structure in a market this liquid is fragile, and the disciplined way to find out is to fix the windows and the test in advance and report whatever comes back — which is what happened here. The engine said “efficient” where the market is efficient, confirmed the funding-tethering across every regime we sampled, and correctly flagged the one flashy result as window-specific rather than defending it. The value of this note is the downgrade, not a rescue.
None of this is a trading claim. The basis is a few basis points; the funding-tethering is a structural correlation, not a costless trade; and the carry it implies flips sign across windows, so there is no standing effect to capture. As with everything here, this is monitoring-grade measurement, not investment advice.