Score-weighting vs equal-weighting: why equal allocation is usually wrong

Two ways to answer the same question

Diversifying across leaders solves one problem and immediately creates another. The first problem — single-trader concentration — is well understood: one account, one strategy, one drawdown, and nothing structural to absorb it. The second problem is quieter. Once capital is spread across several leaders, something has to decide how much each of them gets, and that decision has more influence on your outcome than most users realise.

There are two serious answers. Equal-weighting gives every leader in the basket the same share of mirrored capital. Score-weighting gives each leader a share proportional to the strength of their measured record, so the leader with the most robust evidence controls the largest slice and the leader who barely qualified controls the smallest.

Both are rules. Neither is arbitrary. But they encode very different beliefs about what you know, and only one of those beliefs survives contact with the data you actually have.

How equal-weighting works, and why it feels fair

Equal-weighting is simple to describe and simple to verify. Ten leaders, ten equal allocations. Nobody is favoured, no model sits between you and the allocation, and the composition of your exposure is obvious at a glance. If a leader performs badly, the damage is capped at their fixed share by construction. That is a real property and it explains most of the appeal.

It is also genuinely defensible in one specific situation: when you have no reliable way to distinguish between the candidates. If the leaders in front of you are indistinguishable on the evidence available, spreading capital evenly is the honest response to your own uncertainty. Equal-weighting is the correct allocation under complete ignorance about relative quality.

The problem is that this is not the situation on Hyperliquid. Positions, fills and realized outcomes are on-chain and continuously observable. The record is not equally thin across candidates — it is measurably different in length, in consistency, in how the trader sized positions, and in how close they have run to liquidation. Choosing to allocate as though those differences did not exist means discarding information you already have.

The structural problems of equal allocation

In a multi-leader system, equal-weighting produces four distinct problems. They compound, and none of them are fixed by choosing better leaders.

• It flattens real dispersion in evidence quality. Even inside a pool that clears every hard floor, the difference between the top-ranked and the last-qualifying leader is substantial: hundreds of closed trades across multiple regimes versus a shorter record in one regime; profit factor with a comfortable margin versus profit factor barely above the threshold. Equal-weighting assigns those two the same capital. That is not neutrality, it is an active claim that the gap between them carries no information.
• It funds the weakest qualifying leader at full size. Every basket has a marginal member — the leader who just cleared the bar and would be replaced first. Under equal-weighting that leader receives exactly as much capital as the strongest member. Their contribution to your drawdown is therefore identical in magnitude to the strongest leader's contribution to your gains, which is precisely the wrong symmetry.
• It fragments margin against Hyperliquid's minimum sizes. Each leader is mirrored in its own isolated sub-account with its own balance and its own margin. A slice sized purely by headcount rather than by conviction can fall below the point where mirroring is faithful: minimum order sizes and lot rounding mean small target sizes round badly, some of the leader's trades cannot be mirrored at all, and margin buffer per sub-account thins out. Tracking error rises fastest in the smallest slices, so equal-weighting spends capital on positions least likely to reproduce the leader's behaviour.
• It cannot respond to deterioration. Under equal-weighting a leader's allocation is a function of how many leaders exist, not of how they are performing. So there is no gradual response available. The weight stays fixed at full size right up to the moment a binary removal decision fires, which forces every adjustment to be an exit and pushes the whole burden of adaptation onto the replacement mechanism.

How score-weighting allocates against evidence

Score-weighting starts from the composite score that already governs qualification. Every candidate is measured on the same five factors — realized PnL consistency, win rate, profit factor, position discipline and account survivability — each with a hard floor that has to be cleared before a leader is eligible at all. Failing any single floor disqualifies a trader regardless of how strong the remaining factors look.

The step that equal-weighting skips is using the rest of that number. The score is not only a pass/fail gate; it is a graded measure of how much corroborated evidence supports the leader. Allocation weight is set in proportion to it, so a leader whose consistency holds across regimes and whose position discipline is well inside its floor receives a larger share of mirrored capital than a leader who satisfies the same criteria with less margin and a shorter record.

Two properties of the inputs are worth stating plainly. The factors are computed from realized, closed outcomes and observable position behaviour — not from unrealized marks, which can flatter an account holding a losing position indefinitely. And the score is recomputed continuously as new closed trades appear on-chain, which means the weights are not a signup-time snapshot. They move with the evidence.

We publish the factors and the logic, not the exact weightings or thresholds. That is a deliberate limit rather than marketing coyness: on a venue where positions are public, a published formula becomes a target to optimise against rather than a standard to satisfy, and the moment a rule can be gamed it stops measuring what it was built to measure. The mechanism you need in order to evaluate the system is fully stated. The parameters are not.

Why higher conviction should receive more capital

The case for scaling capital with conviction is not that stronger leaders will earn more next month. Nobody knows that, and any system claiming otherwise is selling a forecast. The case is about the reliability of the evidence, which is a different and much more tractable claim.

A leader with a long record spanning several volatility regimes, consistent sizing and no near-liquidation events has demonstrated something a leader with three profitable months has not: that the results are unlikely to be an artifact of one favourable environment, and that the risk-taking behind them is repeatable. Their forward distribution is not necessarily better, but it is narrower and better understood. Allocating more capital to the position you understand better is standard risk practice in every other part of finance, and it is the same logic here.

The asymmetry cuts the other way too. The marginal qualifying leader is the one most likely to be on probation next quarter and most likely to be replaced. Funding them at a smaller weight means the eventual replacement — including its market-exit cost and the 0.1% builder fee on the closing notional — applies to a smaller slice of your capital. You are paying less for the outcome you should expect to be least stable.

Score-weighting is not concentration, and the distinction is load-bearing. Weights are capped, so no single leader can dominate the basket regardless of how strong their score is. A cap is an admission built into the design: the score measures a past record, the record can be wrong about the future, and no amount of evidence justifies exposing a portfolio to one strategy's failure. Score-weighting tilts toward the best evidence; the cap prevents the tilt from recreating single-trader risk under a different name.

How weighting interacts with replacement and risk controls

Weighting is not a separate feature from replacement. It is what makes replacement proportionate, because a variable weight gives the system a response that sits between doing nothing and closing everything.

When a leader's composite score starts to decay, their weight falls with it. Capital flows away from a deteriorating leader while the evidence is still ambiguous — before there is enough of it to justify removal, and long before a drawdown becomes emotionally obvious. Sustained decay moves them to probation, which cuts weight further and stops new capital being routed to them, and if the trend continues they are replaced by the highest-ranking qualifying candidate outside the basket. A hard breach — position sizing abandoning the discipline they qualified on, exposure carried near liquidation, an abrupt strategy change, extended inactivity — bypasses that curve and sets the weight to zero on detection.

Under equal-weighting none of that gradation is possible. Weight is fixed by headcount, so the only available responses are keep or remove, and every adjustment has to be an exit that pays taker cost and the builder fee. Score-weighting converts most adjustments into a repricing rather than a transaction.

Sub-account isolation is what allows the weights to mean something operationally. Each leader has its own balance, its own margin and its own liquidation price, so a weight is a real quantity of capital rather than a notional share of a pooled account. Reducing one leader's weight or removing them entirely closes positions in that sub-account only; no other leader's margin, exposure or liquidation price moves. In a single netted account, where Hyperliquid holds one net position per market, changing one leader's share necessarily disturbs positions that other leaders contributed to — which means the weights would not be independently enforceable at all.

Conviction-scaled allocation

Equal-weighting is the right answer to a question nobody is actually asking on a transparent venue: how should capital be split when the candidates cannot be distinguished. On Hyperliquid they can be. The records differ in length, in consistency, in discipline and in how close each account has run to failure, and a system that allocates as though they were identical is choosing to ignore its own best input.

Score-weighting is the more defensible default because it makes allocation a function of evidence rather than headcount, caps any single leader so the tilt never becomes concentration, and gives the system a proportionate response when evidence starts to weaken. It does not predict returns and it will occasionally allocate more to a leader who goes on to underperform. What it does is keep capital and conviction aligned, continuously, without waiting for someone to notice.

If you want the layer underneath this, the scoring note sets out the five factors and floors in detail, the replacement note covers what happens when a score decays, and the documentation and risk disclosure describe the isolation model and what weighting does not fix.