​
The question nobody answers

A headline Sharpe ratio cannot distinguish a trader with genuine entry timing from one who held through a favourable regime. Both produce the same number. Decomposing realized P&L across time partitions separates them.

​
Two attribution modes, always both

realization

Attributes P&L to when the position closed. “When did the money land.”

exposure

Attributes P&L across the position’s life. “When was the risk held.”
They disagree for any position held across a partition boundary, and the two modes legitimately have different totals. The disagreement is informative — a trader whose realization concentration is sharp but whose exposure concentration is flat is being credited for exits, not for holding.

​
Windows

  • Partitions (session, weekpart, hour_of_day) tile the span exactly once.
  • Overlay windows deliberately do not tile, so they may overlap.
Window definitions, classification, and caveats are always returned even when you filter, because a concentration table without its window definition is a number without its method.

​
Every label needs a permutation p-value

The module refuses to classify rather than guessing. A style label requires both a magnitude and a permutation p-value — 2,000 permutations per window.A 40-trade account concentrates by accident. “Your top trader is a funding arbitrageur” is a claim about a named stranger, and it should not rest on a pattern that random reshuffling reproduces.

​
Three Hyperliquid facts that corrupt naive implementations

These were verified against live data and will affect anything else reading fills, not just this module.
Verified against live userFunding: all 24 hours present, minute == 0 on 233 of 233 events.The 8-hour cadence is Binance’s. Anything assuming three settlements a day is using a denominator 8× too small.
The reducing leg on a spot fill still carries a populated closedPnl.A dir.startswith('Close') filter silently drops them. On one live account that attributed 5,299ofa5,299 of a 515,191 book.The correct rule: a non-zero closedPnl is realized P&L, whatever the leg is called.
Oldest-first from startTime, with no cursor and no error.Four active accounts all returned exactly 2,000; one covered 96 minutes of a 180-day request. You do not get an error or an empty result — you get a correct measurement of the wrong period.The module paginates: advance to last_ts + 1, dedupe on tid/time/hash/coin, with a 40-page cap.
A fourth trap, from the same family: Hyperliquid sometimes reports an outer window’s pnl as exactly 0.0 while an inner window is non-zero. Nested-window subtraction turns that into a phantom loss equal to minus the inner window. Treat an outer 0.0 as unreported, not flat.

​
Cost and freshness

A live decomposition paginates fills and funding and runs 2,000 permutations per window — a nightly-cron cost, not a per-request one. Endpoints serve the published snapshot and ship age_s. Pass live=true on the MCP tool to force a re-run.

​
Access

Full specification: docs/methodology/temporal-alpha-decomposition-v1.md.
Scienter publishes impersonal market signals of general and regular circulation. Nothing here is investment, legal, or tax advice, or a recommendation to buy, sell, or hold any asset. Scienter is not a registered investment adviser or broker-dealer. Trading digital assets can result in the total loss of your funds — see Disclaimers for the full text.