Behavioral Signals in Perpetual Futures Wallets

A research note on the behavioral signatures that make a perpetual-futures wallet fadeable: excessive turnover, loser retention, loss-driven size escalation, poor timing and persistence across independent windows.

Version

v0.1, July 2026

Status

Behavioral taxonomy

Claim

A wallet-level signal, not a moral label

Abstract

Main claim

Dumb money should not mean a wallet that lost yesterday. It should mean a wallet whose future trades can be classified ex ante as low-quality because the account repeatedly exhibits documented behavioral mistakes. The distinction matters for whosthefool: the product should fade behavior, not embarrassment.

The behavioral-finance literature gives this claim a disciplined vocabulary. High-turnover retail accounts tend to underperform after costs [1]; investors sell winners more readily than losers [2]; overconfidence is associated with excessive trading [3]; aggregate individual-investor losses can be large [4]; and the lower tail of day-trader skill is persistent [5]. The crypto question is whether those same mistakes leave measurable traces in public perpetual-futures wallets.

DEi,w=PGRi,wPLRi,w(1)\mathrm{DE}_{i,w}= \mathrm{PGR}_{i,w}-\mathrm{PLR}_{i,w} \tag{1}
ADRi,w=E[ΔNi,eri,e<0]E[ΔNi,eri,e>0](2)\mathrm{ADR}_{i,w}= \frac{\mathbb{E}[\Delta N_{i,e}\mid r_{i,e}<0]}{\mathbb{E}[\Delta N_{i,e}\mid r_{i,e}>0]} \tag{2}
LDRi,w=median(τi,eri,e<0)median(τi,eri,e>0)(3)\mathrm{LDR}_{i,w}= \frac{\mathrm{median}(\tau_{i,e}\mid r_{i,e}<0)}{\mathrm{median}(\tau_{i,e}\mid r_{i,e}>0)} \tag{3}
Bi,w=z(TURN)+z(DE)+z(ADR)+z(LDR)z(MARKOUT)(4)B_{i,w}= z(TURN)+z(DE)+z(ADR)+z(LDR)-z(MARKOUT) \tag{4}

The behavioral score B is intentionally a measurement spec, not a fitted black box. Higher values mean more costly turnover, stronger disposition behavior, more averaging down, longer loser duration and worse original-direction markout.

1. Literature

Why behavior can be alpha

The point of a fade engine is not that retail traders are universally wrong. The point is that repeated errors can be sorted before the next trade. Barber and Odean show that the most active households have lower net returns than less active households after trading costs [1]. Odean's disposition-effect evidence shows a second failure mode: accounts realize gains while letting losing positions continue [2]. These are not one-off mistakes; they are behavioral processes that can persist in account histories.

The day-trading evidence is especially relevant. Barber, Lee, Liu and Odean find strong cross-sectional dispersion in speculator skill: a small skilled tail exists, but the weak tail remains weak after sorting by prior behavior [5]. For whosthefool, that is the core design rule. Do not fade the entire leaderboard. Exclude the skilled tail, exclude market-making flow, and focus on wallets whose public fills reveal repeated, non-random mistakes.

2. Signals

Observable behavioral signatures

Public fills are enough to define a behavioral fingerprint. A trade becomes useful only after it is placed inside an episode: direction, entry, exit, notional, holding time and post-entry markout. Once episode accounting exists, the signals in Table 1 can be measured without knowing a wallet owner's intention.

Table 1 · Behavioral signal families

SignalObservable behaviorPrimary measurementResearch anchor
OvertradingHigh turnover and high trade count relative to account value.Turnover, trades per $1k account value, fee share of gross PnL.Active retail trading underperforms after costs [1].
Disposition gapWinners are closed quickly while losers are retained.PGR - PLR, loser holding-time ratio, realized win/loss skew.Investors are reluctant to realize losses [2].
Averaging downPosition size rises after adverse markout instead of after evidence of skill.Notional-after-loss / notional-after-win; leverage drift in drawdowns.Overconfidence predicts excess trading [3].
Poor timingEntries are followed by adverse short-horizon markout.5m, 1h and 4h original-direction markout after entry.Trading losses can be economically large in aggregate [4].
PersistenceThe same wallet remains in the weak tail across windows.Score-decile transition probability; repeat negative markout.Day-trader skill and lack of skill are persistent [5].

Figure 1

Behavioral fingerprint of a fadeable wallet

The target wallet is not merely losing. It is losing in a repeated behavioral shape: costly turnover, loser retention, size escalation, poor entry timing and persistence across windows.

TurnoverDispositionAveragingTimingPersistenceClean loser8277647169Noisy tourist6539475128High frequency9412182215Lucky winner3622251817

3. Episodes

The shape of losing

A bad wallet usually does not lose by being wrong once. It loses by being wrong asymmetrically: small realized wins, large tolerated losses and long loser duration. That structure is close to Odean's disposition-effect mechanism, but translated from stock sale decisions into perpetual-futures exposure management [2].Figure 2 shows the measurement target.

Figure 2

Small wins, large losses, long loser duration

Illustrative episode geometry. A fadeable wallet tends to harvest small wins, tolerate large losses and let losing exposure live longer than winning exposure.

Median win+18 bps
Median loss-63 bps
90p loss-185 bps

Holding time

Median win2.1h
Median loss9.4h
90p loss21.6h

The important product implication is that a wallet can look exciting while still being a high-quality fade target. A large win rate is not sufficient if the average loss is much larger than the average win, and a green leaderboard day is not sufficient if the wallet's episode history shows loss-extension behavior.

4. Measurement

From fills to score

The measurement system should be ex ante. At the start of a validation window, freeze the feature definitions, rebuild the trailing episodes, winsorize extreme observations, z-score each feature against the screened wallet universe and publish the score hash. The score is allowed to be simple because the empirical burden is carried by the out-of-sample test, not by model complexity.

Figure 3

From wallet behavior to product decision

The research object is a measurement system. The paper's value is that every product decision can be traced back to observable public behavior.

Step 1

Public fills

user fills and account snapshots

Step 2

Episodes

entry, exit, direction, markout

Step 3

Features

behavioral ratios, not raw PnL

Step 4

Frozen score

rank before validation

Step 5

Decision

fade, watch or skip

Hyperliquid's public API makes this design auditable at the wallet level: account state, fills by time and order evidence can be used to reconstruct exposure and markout without private brokerage data [7]. That is why whosthefool can present research as a product primitive rather than as an opaque marketing claim.

5. Product

What this means for whosthefool

The research direction suggests a strict product language. A card should not say "this wallet is bad" only because the last trade lost money. It should show a small number of behavior-backed reasons: overtrading, loss holding, size escalation, poor timing and persistence. A user should be able to understand why a wallet is in the deck before choosing fade or skip.

Product standard

  • Every Fool Score reason maps to a public, episode-level measurement.
  • The deck excludes high-frequency and market-making flow before behavioral scoring.
  • Watching is a valid state: the wallet is behaviorally selected, but has no open exposure yet.
  • Fade sizing should be smaller when the behavioral signal is strong but execution quality is weak.
  • Explanations should emphasize repeated behavior, not insults or one-day loss screenshots.

6. Validity

What this note does not prove

This note is a taxonomy, not a live PnL audit. It argues that the right unit of analysis is repeated behavior, and that existing literature makes those behaviors plausible predictors of future underperformance. It does not prove that a particular score is profitable after fees, funding and slippage. That is the job of the decile study in Research Note 03.

The design can fail in clear ways. If high-score wallets stop trading, the product should wait rather than force a trade. If the weak-tail signal disappears in a regime shift, the score should lose validation status. If the signal exists before costs but not after execution, it is a research signal but not a product edge.