A decile-study protocol for testing whether a frozen Fool Score predicts future negative alpha in Hyperliquid wallets, and whether the inverse strategy remains positive after fees, spread, slippage, funding and missed fills.
Version
v0.1, July 2026
Status
Pre-registered test protocol
Claim
Predictive study, not marketing backtest
Abstract
Main claim
The central question is not whether high-score wallets lost in the past. The question is whether a score computed before the validation window predicts future underperformance. The cleanest public answer is a decile study: sort wallets into ten buckets by pre-trade Fool Score, then measure future original-direction markout and future net fade return.
This is the same logic that makes the day-trading skill literature useful for whosthefool. Barber, Lee, Liu and Odean sort traders by prior performance and test whether the ranking has out-of-sample content [1]. Research Note 03 adapts that design to public perpetual-futures wallets, with cost-aware execution accounting and multiple-testing discipline.
The study passes only if the pre-trade score orders future outcomes. Equation (3) is the product quantity: inverse return net of all execution costs.
1. Design
Frozen score, future returns
A decile study is attractive because it is easy to audit. At time t0, freeze every scored wallet's Fool Score and assign it to D1 through D10. From that moment onward, the validation data are future data. No feature weight, data filter or horizon can be changed after observing the outcome.
The main output is the slope in Figure 1. A valid score should generate two mirrored curves: original markout becomes worse as Fool Score rises, while net inverse markout becomes better. A profitable top decile is not enough if the middle deciles are noise; the slope is what shows predictive structure rather than lucky selection.
Figure 1
Expected decile slope after costs
Illustrative target result. The study only passes if the slope is visible before looking at any single horizon: higher pre-trade Fool Score should imply higher future net fade return.
2. Sample
Walk-forward sample construction
The unit of sampling is a wallet-window. Each window has a trailing scoring period and a non-overlapping future validation period. Wallets that stop trading are not silently removed: inactivity is an outcome and remains in the report as lower opportunity count. This prevents the most dangerous survivorship bias, where the worst wallets disappear before their losses can be measured.
Figure 2
Walk-forward validation clock
The score may use only information available before the validation window. Every change to the score recipe restarts the clock.
T-90 to T0
1. Score
Build episodes and freeze score before validation.
T0 to T+30
2. Validate
Measure future original and inverse markout.
T+30
3. Report
Publish decile tables, costs and failures.
T+30 to T+120
4. Repeat
Roll forward with no overlap in validation data.
For each validation trade, the study records coin, direction, notional, entry time, markout horizon, realized fills, estimated spread/slippage, funding, failed-order cost and whether the product would have been in watching or pending execution. Hyperliquid's public information endpoints provide the raw data surface needed for this reconstruction [7].
3. Tests
Statistical tests
There are four pre-registered hypotheses. The study reports means, medians, confidence intervals and wallet-level clustered statistics. Because markouts overlap across horizons and wallets trade in bursts, naive independent episode t-tests are not sufficient. The protocol therefore uses wallet clustering, Newey-West style horizon corrections where horizons overlap [3], and block bootstrap confidence intervals[4].
Figure 3
Pre-registered pass criteria
A paper-like research page should state what would count as evidence before users see the chart.
Table 2 · Hypothesis matrix
| Test | Hypothesis | Pass criterion | Why it matters |
|---|---|---|---|
| H1 | Original-direction markout decreases monotonically with Fool Score. | Spearman rho < -0.70 across deciles at two or more horizons. | Rejects the idea that score is only descriptive. |
| H2 | Net inverse return increases monotonically with Fool Score. | Spearman rho > 0.70 and D10 net edge > 0 after costs. | Establishes that the weak-tail signal is tradeable. |
| H3 | D10 outperforms matched controls and random fade. | Top-minus-control confidence interval excludes zero. | Separates Fool Score from generic contrarian exposure. |
| H4 | The effect survives multiple horizons and testing correction. | Benjamini-Hochberg q <= 0.10 after 1h, 4h, 24h and 7d tests. | Prevents cherry-picking a single lucky horizon. |
Multiple horizons create multiple chances to fool ourselves. A one-hour result that disappears at four hours or seven days is still interesting, but it is not enough for a product claim. The study therefore reports raw p-values and Benjamini-Hochberg q-values across the pre-declared horizon family [5].
4. Regression
Controls and cross-sectional slope
The chart is the communication layer; the regression is the robustness layer. A Fama-MacBeth style cross-sectional regression estimates whether score rank predicts next-window net fade return after controls for account value, turnover, coin mix, liquidity bucket and market regime [2].
The product thesis needs beta to be positive after controls. If beta vanishes once turnover, coin mix or liquidity are controlled, the score is not adding enough information.
5. Baselines
Not simply fading losers
The strongest criticism of whosthefool is that it might be doing hindsight selection: find people who lost, then narrate why they were fools. The decile design answers that by comparing Fool Score against baselines that a user could have implemented without behavioral research.
Table 1 · Required baselines
| Baseline | Construction | What it rules out |
|---|---|---|
| Random fade | Same timestamps, same coin universe, randomized wallet labels. | Tests whether the edge is simply being short noisy flow. |
| Recent loser fade | Fade wallets selected only by trailing PnL. | Tests whether Fool Score adds behavioral information beyond losses. |
| Matched controls | Nearest wallets by account value, turnover, coin mix and liquidity bucket. | Controls for the fact that some markets are structurally easier to fade. |
| No-trade cash | Zero exposure with funding and opportunity cost reported separately. | Keeps the product honest when execution cost overwhelms signal. |
A useful product edge requires Fool Score to beat the recent-loser baseline. If recent losers and high Fool Score perform the same, the product should simplify its claims. If Fool Score beats recent losers after costs, the research can say something stronger: the behavioral decomposition matters.
6. Publication
What the page should publish
The published result should feel closer to an academic appendix than a marketing chart. For each horizon, show the decile mean, median, confidence interval, number of active wallets, number of inactive wallets, total notional, realized fees, funding, slippage and rejected-order cost. Users should see the cost bridge before they see any headline return.
Minimum evidence for a green label
- D10 net fade return is positive after all costs in at least two independent windows.
- D10 beats random fade, recent-loser fade and matched controls.
- The decile slope is monotone enough to be visually and statistically clear.
- Multiple-testing corrected q-value is no greater than 0.10 for the pre-declared horizon family.
- Capacity is reported: the edge cannot require more notional than the market can absorb.
7. Falsification
What would disprove the score
A serious research page should be willing to lose. Fool Score is unproven if the D10 confidence interval includes zero after costs, if D10 does not beat recent-loser selection, if the slope only appears after removing inactive wallets, if the effect exists only in one cherry-picked coin or horizon, or if execution costs consume the entire gross markout. In those cases, the correct product behavior is to downgrade the score label, not to rewrite the backtest.
That discipline is precisely what can make the page persuasive. Users do not need a promise that every fade wins. They need evidence that the platform knows what would count as being wrong.
References