HOKUTO

HOKUT

Built on Hyperliquid · live, real capital

We don't predict the market. We measure it and hold course.

Launch app ↗

Return on capital — —

北斗  ·  seven stars, one bearing

How it works

Three steps, and a human in none of them.

  1. STEP 01

    Engines hunt patterns.

    Deterministic engines read the market's raw microstructure on Hyperliquid — around the clock, no human hands. Every entry ships with an on-chain stop before anything else.

  2. STEP 02

    A governor grades them.

    Every closed trade is graded into a public record. Engines with positive expectancy earn real capital; engines that slip below the bar are cut off automatically and must earn their way back on paper. No favorites — it cut our best engine the morning after its best night.

  3. STEP 03

    Everything is published.

    Wins, losses, drawdowns, cuts — live at hokuto.trading, percentages only. No signals, no promises. Just the record, growing.

The seven

The four gates and the three rules.

A strategy walks the first four stars in order — measured, tried, graded, auditioned — and any one of them can end it. The last three are not gates. They apply to everything that got through, for as long as it keeps trading.

EARTH VIEWThe figure exists only from here. The seven lie between 58 and 124 light years out, no two at the same distance.

    The governor

    The machine that stops the machine.

    The conductor promotes, demotes and refuses without asking anyone. Its decisions are published in the order they were made, and a cut is given the same room as a promotion — when it loses, the machine stops itself.

    Decision classes

    Promoted

    A graded record crossed the bar: positive expectancy over at least twenty graded trades. Real capital is granted.

    Demoted · cut off

    Expectancy fell to zero or below. Real-money rights are revoked; the engine keeps trading on paper and must earn its way back.

    Re-promoted

    An engine earned its way back from paper. It has happened inside a single week: cut on the Friday, trading again on the Monday.

    Veto · beta-tide

    An entry refused because the day's broader tape was against that class of trade. A measured losing bucket, closed at the signal layer.

    Veto · falling-knife

    An entry refused into a deteriorating tape. Also a measured losing bucket, also closed.

    Suspension · equity floor

    All live entries paused because equity touched a pre-committed floor. Banked progress is locked and the ratchet does not release.

    Blackout · macro

    No new entries around FOMC, CPI or NFP prints. The window closes before the print and opens after it.

    Connecting

      The decision log renders here, newest first, the moment the governor endpoint is live. Until then this section states nothing: no sample events, no illustrative numbers.

      The graveyard

      Most of what we build does not survive the tribunal. This is the part nobody publishes.

      A promotion funnel is only meaningful if things actually die in it. These did, with the measurement that killed them.

      22 synthetic TradFi markets
      — oil, indices, semiconductor names       Zero qualifying setups across 45 configurations. The order books are market-maker quoted; the event we trade never forms, not even on a 12% intraday range.
      Funding harvest $0.43 accrued against $12.50 of execution cost over ten cycles. The strategy is sound. Our capital base is not.
      Re-entry on a second touch
      of the same trigger       Positive on paper, negative across 600 trades once fees were real. One entry per trigger per day is the strategy, not a safety blanket.
      Two new markets,
      a full week of tape each       Neither produced a tradeable signal at any threshold. Volume is not liquidity, and liquidity is not structure.
      Four analysis engines,
      retired the same day       280 graded signals, −40R, no path to capital. Deleted, not demoted.
      A signal-quality discovery,
      retracted within a day       It held on the pooled data and evaporated once split by engine. Simpson's paradox, caught before it reached the sizing logic.
      CVD-divergence engine 2,734 divergence prints replayed from our own recorded tape: 51% directional accuracy, ≈0bps average forward return. The signal does not clear its own fees. No engine.
      Session-open range breakout 598 breakouts, three sessions, four coins, 60 days: negative before fees at every horizon. Crypto session opens mean-revert — and the fade doesn't clear fees either.
      Funding-flip as a
      directional signal       66 qualifying flips; the apparent edge lived entirely in the first half of the sample and vanished in the second. Not robust, not traded.
      BTC→alt lead-lag
      at the minute scale       The alts don't follow the leader — they fade it, hardest on the biggest moves. Neither direction survives costs.
      Three dip-buy sides, retired by their own re-exam AAVE longs: negative across 90 days and 0 wins in 11 live attempts — two independent measurements agreeing. BTC and ETH longs: flat, which after costs is not flat. With three concurrent seats, a slot that earns nothing is a slot denied to a side that does.
      BTC on the breakout engine The most efficient market in crypto: gross result indistinguishable from zero across a full month of turnover, with the largest fee bill in the book. The entries were fine and the exits were vindicated — there was simply no edge left to pay for them. Live entries stopped.
      Liquidation-cascade
      snapback       56 forced-flow bursts from our own tape: the median bounce is real, the mean is negative. The cascades that keep cascading eat the ones that snap back. Knife, not edge.

      Everything above was built, measured and thrown away. The list is longer than the list of what runs.

      Live record

      Read from exchange fills. Percentages only.

      Account size and dollar figures are not published. When the capital base moves, only a percentage stays honest.

      Return on capital — —
      Today — Since 00:00 UTC
      Closed trades — —
      Win rate — Closed positions

      The curve is one number made of two. What the strategies earn and what it costs to execute them are different things — and for a month ours disagreed. We publish both, because an account that loses while its book wins has a fixable problem, and one that hides the split does not know which it has.

      Strategies, before costs — What the closed trades produced
      Cost of execution — Exchange fees on the same trades
      What reached the account — The difference, and the whole problem
      Max drawdown — Deepest peak-to-trough on the published curve
      Days live — Continuous, since first fill
      Equity curve — return on capital
      Connecting
      ClosedMarketResult

      Waiting for the exchange feed.

      EngineMandateGradedExpectancyGradeStatus

      Per-engine grades appear here once the governor endpoint publishes them.

      Research

      Notes from the parts of the system that are hard to get right.

      Method

      How we nearly shipped a false discovery

      A signal-quality metric separated winners from losers cleanly across thirty live trades. Split by engine, the effect disappeared: it had been measuring which engine traded, not which signal was good. Simpson's paradox, one day from reaching the sizing logic.

      Draft
      Execution

      The wick our breakeven never saw

      The backtest armed protection the moment price touched its trigger. The live module sampled price on an interval, so a wick that crossed and bounced between two samples was forgotten entirely. The policy was sound; the implementation was less faithful than the test that authorised it.

      Draft
      Microstructure

      Two engines, one coin

      Positions net per market on-chain. When one strategy opened against another's live position, the exchange resolved the conflict silently and one engine's trade simply ceased to exist — while its stop order remained, sized for a position that was no longer there.

      Draft

      Live voice

      The system files its own reports.

      One post per closed position, percentages only, and a card every morning. Nothing below is written by a person, and nothing is held back for being unflattering.

      Waiting for the exchange feed.

      What we don't claim

      The record is short. It starts on — and a short record is not evidence, no matter how well it is presented. Read it as a sample, not as a result.

      A losing month is expected. When one arrives it will be published here, in the same place and at the same size as the gains, because a record that only shows the good periods is a brochure.

      Anyone promising steady returns is selling something. Perpetual futures do not produce steady returns. They produce a distribution with a real tail, and the tail arrives on its own schedule.

      We publish no account size and no dollars. The unit on this page is percent of capital, and every figure on it is read from exchange fills. If a number cannot be sourced there, it is shown as an em dash.

      Past performance measures the past. It is not a forecast, and nothing on this page should be read as one.

      Waitlist

      Leave an address and we'll write when the record gets longer.

      No newsletter, no launch sequence. One note when there is something measured worth reporting, and the same figures this page shows.

      Address only. Nothing else is stored, nothing is passed on.