Can Weak Wallets Generate Alpha?

The claim is not that every losing wallet should be copied in reverse. The defensible claim is narrower: if a wallet can be classified ex ante as a member of a persistent negative-alpha cohort, then the inverse of its future trades may have positive expected value after execution costs.

This hypothesis is grounded in a large behavioral-finance literature. Active retail trading has been shown to reduce net returns ^[1], investors display a tendency to realize winners and retain losers^[2], aggregate individual trading losses can be economically large ^[3], and day-trader skill is highly dispersed with a reliably weak lower tail ^[4]. The open question is whether similar weak-tail behavior is measurable and tradeable in Hyperliquid wallets.

The estimand is the out-of-sample net fade edge of wallets selected by a pre-trade score: the inverse markout net of all execution costs, written π in the equation block. This note uses that one term throughout. The unit of observation is a closed episode, not an account screenshot. An episode starts when net exposure leaves zero and ends when exposure returns to zero or flips.

A positive research result requires that top-score wallets have worse future original-direction markout than matched controls, and that the inverse strategy remains positive after all costs. This separates behavioral selection from execution quality. The horizon parameter t is not free: Figure 1 sketches the target term structure, in which gross inverse markout must clear the all-in cost stack before some horizon for a tradable window to exist at all.

The sample is built by a two-stage public funnel. A daily discovery pass screens the venue's leaderboard for wallets that are active enough to score and not structurally unfadeable; survivors enter a scoring queue, where each wallet's trailing 90 days of fills are reconstructed into closed episodes and scored. The top 1,024 wallets by Fool Score form the deck — simultaneously the tradable cohort and the pre-registered test cohort. Every input is a public API endpoint ^[7]: fills by time, account snapshots and L2 book. Anyone can rebuild the sample.

The screening floors and ceiling are not neutral: they define the population the thesis speaks about. The volume and account-value floors remove wallets that cannot be scored or faded at size; the turnover ceiling removes high-frequency and market-making flow, whose fills reflect inventory management rather than belief. All claims in this note are therefore claims about the screened cohort, and the published audit must report the funnel count at every stage — wallets seen, wallets screened, wallets scored, wallets decked — so that selection is visible rather than silent.

Because episodes within one wallet are correlated, wallets — not episodes — are the clustering unit for every standard error in Section 5. Wallets that go inactive during a validation window remain in the denominator until their exit is observed, which is what the exit column of Figure 2 makes explicit.

The score is deliberately boring. Each wallet-window is measured on five factor families, each anchored in a documented behavioral regularity; family measurements are winsorized, z-scored against the screened cohort, and combined with equal weights into a single monotone score. Higher means more exploitable. There is no fitted model and no interaction terms in v0.3 — a score this simple is harder to overfit and easier to falsify.

Two discipline rules protect the score from its own authors. First, the recipe is frozen and hash-stamped before each validation window opens; the walk-forward test in Figure 6 is only meaningful if the score cannot move after the fact. Second, any change to a feature, weight or winsorization bound re-registers the study and restarts the validation clock. The features themselves are chosen to be expensive to fake: the only way for a wallet to stop looking foolish on these measures is to stop trading foolishly.

Barber and Odean document that households with the highest turnover materially underperform after costs^[1]. Their overconfidence study also links higher trading activity to lower net returns^[5]. Odean's disposition-effect paper shows that investors are more willing to realize gains than losses, and that the losers they keep do not subsequently justify that choice ^[2].

The most relevant empirical bridge is the day-trading skill paper. Barber, Lee, Liu and Odean sort traders by past performance and test the next period; the bottom-ranked tail continues to lose after fees, while only a very small top tail earns reliable positive abnormal returns ^[4]. This is exactly the shape a fade engine needs: avoid the skilled tail and target the persistently weak tail.

Crypto markets do not remove behavioral bias. Schatzmann and Haslhofer find evidence of the disposition effect in Bitcoin transaction data, with varying intensity across regimes ^[6]. That does not prove whosthefool's edge, but it supports the premise that on-chain markets can contain measurable behavioral regularities.

What the literature predicts, concretely, is a sticky transition matrix. If skill and its absence persist, a wallet's score decile in one window should forecast its decile in the next; Figure 2 shows the target shape, including the exit column that makes survivorship an explicit object of the audit rather than a silent bias.

The central test is a decile sort (Figure 3). Freeze the score before the validation window, sort wallets into deciles, and measure the next-window net fade edge. The expected result is a monotonic relation: higher Fool Score implies lower original-direction markout and a larger net fade edge. The four pre-registered hypotheses, each with a numeric pass criterion, are collected in Figure 4.

Signal quality is not sufficient. The edge must survive exchange fees, builder fee, bid-ask spread, slippage, funding and rejected-order opportunity cost. Hyperliquid's public API surfaces the relevant evidence: user fills, fills by time, order status, L2 book snapshots and builder-fee fields ^[7].Figure 5 makes the requirement explicit as a bridge from gross markout to realized PnL, with a stress control on every execution cost.

The audited deliverable is the walk-forward equity curve in Figure 6: each window re-freezes the score on trailing data only, and the fade portfolio is compared against a matched baseline, with the drawdown path reported alongside the average.

A wallet entering watching is not a failed trade. It means the target wallet has no open exposure at that moment. The strategy has joined the portfolio and waits for a future tradable event. If the target already has exposure, execution can move to the queue immediately (Figure 7).

A pre-analysis plan is only credible if it names the ways it can be wrong before a critic does. Five threats dominate here; each carries a mitigation that is part of the design, not an afterthought.

The thesis is pre-registered to be rejected, not defended. Each hypothesis in Figure 4 carries a numeric criterion, and any of the following outcomes falsifies the thesis or marks it unproven: top-decile wallets fail to underperform matched controls at one-sided p < 0.05 with wallet-clustered standard errors; the block-bootstrap 95% confidence interval of the top-decile mean net fade edge includes zero after all costs; the decile slope loses monotonicity, with Spearman ρ below 0.8 or insignificant; no test survives Benjamini-Hochberg control at q = 0.10; or fewer than two of the three walk-forward windows deliver a positive top-decile edge.