How leader replacement actually works (and why it matters)

Selection is treated as permanent. It is not.

Almost every copy trading product on the market is built around a single decision: which trader do you follow. The interface is a leaderboard, the action is a subscribe button, and after that the product has essentially nothing left to say. There is no mechanism that ends the relationship. If the trader you selected stops being good, the system keeps mirroring them faithfully until you personally notice and intervene.

That is a strange way to build a system whose entire value proposition rests on the quality of one input. A trader's public record is evidence about how they behaved over a past window. It is a reasonable basis for allocating capital. It is not a property of the trader that persists indefinitely, and treating the selection decision as permanent quietly assumes that it is.

Replacement is the part of a copy trading system that decides what happens when the original assumption stops holding. It deserves more attention than the selection step, because selection happens once and replacement is the process that has to run forever.

Why edge decay is inevitable

Edge decay is not a moral judgement about a trader and it is not usually a sign of carelessness. It is the ordinary outcome of a strategy interacting with a changing market and a growing account. On Hyperliquid perpetual futures there are four mechanisms that do most of the work.

• Regime change. A trader whose approach depends on directional persistence performs differently once volatility compresses and moves stop extending. Nothing about their execution changed; the conditions the strategy monetises are no longer present. Regimes shift on their own timeline and no scoring window can anticipate the shift.
• Crowding. Profitable behaviour on a transparent venue is observable. Hyperliquid positions and fills are on-chain, so a working pattern attracts imitation, and imitation compresses the dislocation the pattern exploited. The more legible the edge, the faster this happens.
• Size growth. A strategy that worked at one account size may not survive at five times that size. Larger orders consume more book depth, entries and exits move price against the trader, and the market selection that was viable narrows to the deepest perps. Success itself degrades the conditions the record was built under.
• Strategy drift. Traders change. Holding periods shorten, leverage creeps up after a losing stretch, market selection widens into thinner perps, a directional trader starts fading moves instead. When behaviour changes, the historical record stops describing the strategy currently being run — the score is measuring something that no longer exists.

How continuous scoring detects deterioration

The mechanism that catches decay is the same one that governed selection: the composite score, recomputed continuously as new closed trades appear on-chain. Every leader in the basket is re-evaluated on the same five factors used to qualify them — realized PnL consistency, win rate, profit factor, position discipline and account survivability — against the same floors, with no exemption for leaders already being mirrored.

The important consequence is that deterioration does not have to become a crisis before it has an effect. Because allocation is proportional to score rather than equal across the basket, a falling score reduces a leader's share of mirrored capital immediately and gradually. A leader whose consistency degrades and whose profit factor compresses is receiving less capital well before any removal threshold is reached. Score-weighting means the response to decay is continuous, and removal is only the endpoint of a curve that started moving earlier.

It is worth being precise about what the score can and cannot see. It reads realized, closed outcomes and observable position behaviour. That means it detects a change in results and a change in risk-taking reliably, and it detects them with a lag proportional to how many trades the leader closes — a trader with a slow holding period generates evidence slowly, so their deterioration is confirmed slowly. The score cannot read intent. It does not know whether a leader is deliberately switching approach, temporarily reducing activity, or has stopped paying attention. It infers from behaviour, and behaviour is the only honest input available.

We do not publish the exact weightings or thresholds, for the straightforward reason that a published rule on a transparent venue is a rule that can be optimised against rather than satisfied. The logic is the part that matters and it is stated in full: five factors, hard floors, continuous recomputation, weight proportional to score.

Probation and emergency replacement are different mechanisms

Deterioration arrives at two very different speeds, and treating both the same way would produce the wrong response to at least one of them. So there are two paths.

Probation handles slow decay. When a leader's composite score trends down persistently — not a single losing week, but sustained degradation across the factors — their allocation weight is reduced and no new capital is routed to them. They remain mirrored, at a smaller share, while the system accumulates more evidence. Probation exists because a short losing sequence is statistically indistinguishable from noise for most strategies, and removing a leader on noise is itself a cost: you pay exit friction, you lose an exposure that was working, and you replace a known profile with a less-observed one. Probation resolves in one of two directions. The score recovers and the weight is restored, or the decay continues and the leader is replaced by the highest-ranking qualifying candidate not currently in the basket.

Emergency replacement handles breaches that make waiting for more evidence irrational. These are behavioural rather than performance signals: position sizing that abandons the discipline the leader qualified on, exposure carried close enough to liquidation that a single adverse move ends the account, an abrupt change in strategy that invalidates the record entirely, or extended inactivity that leaves the allocation dormant. In those cases the problem is not that recent results are poor — it is that the leader has stopped being the trader the score described. Additional data would not clarify anything; it would only be collected while capital sits exposed to a profile nobody evaluated. Removal happens on detection.

The distinction matters in practice because it separates two questions that get conflated constantly: is this leader performing worse than expected, and is this leader still running the strategy we allocated to. The first deserves patience. The second does not.

What happens to open positions

This is where sub-account isolation stops being an architectural detail and becomes the thing that makes replacement clean. Each mirrored leader trades in its own Hyperliquid sub-account, with its own balance, its own margin and its own liquidation price.

When a leader is removed, the mirrored positions in that leader's sub-account are closed at market, and the freed capital is reallocated to the incoming leader's sub-account under the current score weights. Every other sub-account is unaffected: no other leader's positions are touched, no other leader's margin changes, and no other leader's liquidation price moves. Attribution stays intact, so the outgoing leader's realized contribution is a closed, separable number rather than something entangled with the rest of the basket.

Compare that with a system that mirrors several leaders inside one netted account. Hyperliquid holds one net position per market per account, so unwinding one leader's contribution there means adjusting a net position that other leaders are also contributing to — and any margin released or consumed changes the risk profile of positions that had nothing to do with the replacement. Isolation is what allows a leader to be removed as a discrete operation instead of a portfolio-wide disturbance.

The honest part: replacement is not free. Closing at market pays taker cost, the 0.1% builder fee applies to the closing notional as it does to any mirrored volume, and the exit price is whatever the book offers at that moment rather than a price of our choosing. Sometimes a position closed on replacement would have recovered, and the removal will have cost you that recovery. There is no version of a rules-based process that avoids this — a rule that only ever fires when it turns out to have been right is not a rule. What the process can do is keep those costs bounded and infrequent, which is exactly why the default path is a gradual weight reduction rather than an immediate exit.

Why rules-based beats manual, including your own judgement

The realistic alternative to automated replacement is not a better automated process. It is you, watching, and deciding. That comparison is worth making explicitly, because manual replacement fails in patterns that are well documented and not really about intelligence.

Manual removal is late by construction. A drawdown becomes emotionally obvious well after it becomes statistically meaningful, so the decision to stop following a trader tends to arrive near the bottom of their bad stretch — which is also, for a strategy that has not actually broken, the point where expected forward returns are least bad. Continuous scoring reduces weight when the evidence starts shifting, not when the loss becomes uncomfortable.

Manual selection of the replacement is also biased. Under the discomfort of a loss, the natural move is toward whoever has performed best most recently, which systematically selects for traders whose recent numbers are inflated by a favourable regime — the most fragile moment to allocate to them. A scored candidate pool applies the same floors and the same five factors to the incoming leader as to the outgoing one, with no weight given to how recent or how salient their good period was.

And manual processes are inconsistent across leaders and across time. The threshold you apply to a leader you chose yourself is not the threshold you apply to one the system chose. The threshold you apply while up is not the one you apply while down. Rules do not have that variance, and they do not require you to be awake. Perpetuals trade continuously; a leader can breach position discipline at any hour, and a process that depends on you noticing is a process with gaps in it.

None of this makes the rules correct in every instance. It makes them consistent, and consistent application of a defensible rule is the only thing that compounds. A discretionary process that is right more often but applied unevenly does not.

Adaptive following versus static following

The practical difference between a static and an adaptive system shows up over a long enough window that most users never see it in a demo. A static system's composition is a snapshot of who looked good at signup. An adaptive system's composition is a continuously updated view of who currently satisfies the criteria, with capital flowing toward strengthening evidence and away from weakening evidence, and removal available as the endpoint rather than as an emergency.

That is the whole reason replacement logic belongs in the same design as score-weighted allocation and sub-account isolation. Weighting decides how much each leader gets. Isolation makes each leader independently accountable and independently removable. Replacement decides when the allocation should stop. Any one of the three without the other two is incomplete: weighting without replacement allocates carefully to a set that never changes, and replacement without isolation cannot be executed without disturbing everything else.

If you want the mechanics underneath this — how the five factors are computed, how score translates into an allocation weight, and how isolation is structured per leader — the scoring note and the documentation set them out, and the risk disclosure covers what replacement does not fix, including replacement lag itself.