# Retrospective freeze: results (2026-10-15)

Rule declared in [PREREG.md](PREREG.md) and committed (220254d) before anything below was computed. This is a declared analysis, not a blind pre-registration: the archived boards were public, the author had worked with Hyperliquid's board before, and the three snapshot dates were chosen (from the Wayback list) when the rule was written. All three outcome windows had already ended. The only blind test is the live frozen cohort (first readings 2026-12-09). Script: `tools/retro_freeze.py`. Raw output: [results.json](results.json).

## Setup

Three archived copies of Hyperliquid's public leaderboard (Wayback Machine: 2025-11-10, 2026-01-26, 2026-04-03). For each, the cohort is picked with the same rule as the 2026-10-14 frozen cohort: all-time PnL top 100 plus 30-day ROI top 100 (account ≥ $10k). Each wallet's portfolio was fetched on 2026-10-15; `t_before` scores the history up to the snapshot, `t_after` the history after it, with the estimator behind the wallet check (at least 8 weeks each side).

## Result: the declared rule says **win**, on a weak margin

| Snapshot | Cohort | Scored both sides | ρ(t_before, t_after) | ρ(all-time PnL, t_after) |
|---|---|---|---|---|
| 2025-11-10 | 199 | 134 | 0.274 | 0.227 |
| 2026-01-26 | 199 | 143 | 0.190 | 0.173 |
| 2026-04-03 | 200 | 128 | 0.226 | 0.072 |
| Pooled | | 405 | **0.204** | 0.134 |

- Rule: ρ_t > ρ_pnl in at least two of three snapshots (it is higher in 3 of 3, by 0.05, 0.02 and 0.15), and the pooled ρ_t's 95% bootstrap interval excludes 0 ([0.116, 0.296]).
- **The rule was too easy to pass: the comparison that matters, t against PnL rank, is not significant (below).**

## What it does and does not show

- **t read at a past date predicts the record after it, weakly.** The same wallet appears in several snapshots, so the pooled pairs are not independent; resampling by wallet (269 wallets) gives ρ_t a 95% interval of **[0.085, 0.313]**, still above 0. Wallets' returns are also correlated with each other in the same weeks, which this resampling does not handle, so the interval is likely too narrow. A rank correlation of 0.20 is below the +0.30 threshold Convexly set for its own score.
- **t is not shown to predict better than ranking by profit.** The difference has a wallet-resampled 95% interval of [−0.083, 0.220], which includes 0; two of the three snapshot margins are 0.05 and 0.02.
- **It orders traders; it does not find stars.** After the freeze, 1.5% to 2.1% of each cohort reached t ≥ 2, about what chance gives (2.3%). Wallets with t_before ≥ 2 had a higher median t_after (0.82 to 1.13) than the rest (0.59 to 0.80).
- **Coverage.** 405 of 598 cohort slots (68%) had 8 weeks or more on both sides; the rest were too new at the snapshot or stopped trading after it. Who drops out is not random, and the direction of that bias is unknown.
- t_after measures consistency (risk-adjusted), not the size of the return.

For comparison, the closest product (Convexly, Polymarket) has published two pre-registered forward tests (convexly.app/research/forward-validation, read 2026-10-15): its Edge Score reached an out-of-sample Spearman of +0.11 against a +0.30 threshold it set, which it reports as not cleared; and a separate persistence test on wallets that cleared its bar, closed 2026-08-30, read "persistence_confirmed" (excess forward edge +2.93 pp), with 5.28 effective independent clusters by Kish and one wallet supplying 41.2% of positions.