Which macro chart crypto moves with, per half hour. Moving together is not a cause: this says what moves with crypto now, not what comes next.
CoinsHoursHalf hoursHorizon
Table as it stood
Moving together
Strength is the correlation of 30-minute changes over the same half hours, from −1 to +1. Clear = beyond the two-sided 5% bar with a Bonferroni correction over the eight main charts (an extra chart such as S&P 500: over 15), from at least 20 half hours. That bar treats the half hours as independent, so it is checked a second way: the same half hours, with whole New York trading days resampled 4,000 times at the same level, so that a few strong days cannot carry it alone. Bold and full colour = clear on both. Dotted underline = clear only on the first: grey when the window has fewer than 10 trading days to resample (3 and 7 days, 90 days and the earlier windows of the comparison are not in the half-hour data here), orange when the day blocks include zero. The tooltip of each value gives both intervals. Minus = opposite direction (VIX up, BTC down). Click a column header to sort on it (highest correlation first, click again for most negative first); click a row or a chart to enlarge it, hover it (tap on a phone) to read each half hour. Under the table you switch the extra charts on and off; S&P 500 is on by default. US hours: 08:00 to 16:00 New York, Monday to Friday, so the 8:30 data releases (CPI, jobs, PPI) are in; every chart over the same half hours. Outside US hours: Asia, Europe and evening, each chart over its own half hours. The Nasdaq, the Russell and VIX come from their futures (NQ, RTY, VX), which trade nearly around the clock. Lead: each chart is also measured one half hour ahead of crypto, and "Ran ahead" appears only when that is clear; a 30-minute measurement cannot show which market moved first within the same half hour. The long view on daily returns over months and years is in Create Backtest, Correlation Lab. Source: TradingView public feed, 30m bars; measured on the station, refreshed at most every 30 minutes when this screen is opened.
Breadth · top 60
How broad the move is across the top 60: how many coins are green, beat BTC, sit above their 50- and 200-day average or have EMA33 above SMA44, and how the equal-weighted basket did against the cap-weighted one.
Loading…
Engine Room
Scheduled tasks, what runs now, what starts at login, and the Claude sessions per folder. The switches are yours.
Engine Room is not running. Start it with the desktop button Achtergrond.
Alerts
One list for the whole station: the measures listed under New alert (prices, indicators, funding, open interest, volume, liquidations, breadth, TradingView symbols, early warnings and trending lists), Polymarket markets and the wallets you follow. The station checks them every minute and sends what fires to Telegram, also when no screen of this app is open, as long as the station pc is on, connected and the alert service runs; this pc and the phone see the same list.
Night rest 23:00–07:00: what fires then comes as one message after 07:00. This page plays no sound.
Armed
Went off or paused
These fired once, or you paused them, and are off now. Re-arm one to watch the same level again, or remove it.
New alert
Pick what to watch, then the condition and the level: type it, take a step, or drag on the chart. Every chart in the app also has Add alert… on its right-click menu.
Log
Every alert that fired, from every source, with what it said and whether it reached Telegram. For Polymarket the price an hour and a day later is on Polymarket → Alerts.
Polymarket
Live from Polymarket's public APIs, read straight from this page. Nothing is placed or signed here.
Funding-Arb Portfolio Backtester
Backtests portfolios of the crypto-&-tradfi funding trade (short a Hyperliquid perp to
collect funding, long the matching asset on a broker) and searches a rule space for the
rules that optimize each objective. Models leverage on equity, transaction fees and broker
financing. Liquidations are checked per leg (Liquidation check, below): the Hyperliquid
perp against its own margin and maintenance margin, hour by hour, and the broker leg against
half its margin, trading only while its market is open. Switch the check off for the upper
bound without it.
Asset classes ⓘ
Asset classes ⓘ
Universe & direction ⓘ
Ready.
Carry Lab
Computing this backtest…
Calculating…
This configuration has not been fetched before, so the history is being pulled from
scratch. Hyperliquid answers about one call a second, which is what the wait is. The series
are cached in this tab, so re-running with different rules is instant.
Your setup
How positions are chosen: at each rebalance the engine ranks every eligible asset by
its trailing 7-day (168-hour) average funding rate, annualized, minus the broker
financing cost, and holds the best ones (up to this rule's position cap) that clear your
hurdle until the next rebalance. The trailing-7d funding shown in the hover tooltip is
exactly that selection signal, as it stood at each point in time (it only uses data
available before that date, so there is no look-ahead).
Performance by day of week
Your setup's (the rule in the fields) average daily equity return per weekday (Monday–Sunday) over
the whole backtest, as a plain daily percentage. It is not annualized: compounding one weekday's
average over a year turns a few good days into a meaningless number. Funding accrues every day,
but the tradfi leg only trades on market days, so weekday returns can differ. The note under the
chart says how large a difference has to be before it is more than noise.
Impact of each knob
Your setup (the rule in the fields) held fixed, one axis varied at a time, on the same history.
It shows how sensitive the result is to each setting; if a column barely moves, that knob is not
doing what its name suggests. It is not a robustness test: every row is measured on the data it is
shown on, and settings that interact are never varied together.
Monthly returns
Holdings over time
Trades
Each row is one pair held continuously. “Spread” is the
expected carry at entry; “Corr” is the two legs’ trailing correlation, i.e. how good the
hedge was expected to be. Funding is what the carry earned, Price drift is what the two
legs’ divergence did, Fees are that pair’s trading costs (its share of each rebalance it
entered or left on, and the close at a stop): funding + drift − fees = return, each as a share of
the pair’s notional (both legs). A trade whose return is mostly drift was a directional bet, not carry.
Trade statistics
Per-trade averages across all trades of your setup (the rule in the fields, not the best row of the frequency table). "Per trade" = on the
notional of one position slot (one leg: the book's notional per leg divided by the position cap). Gross is before transaction fees (still net of
broker financing); net is after fees.
Impact of number of positions
Your setup's other settings held fixed, varying only how many positions it
holds at once: the diversification vs profit tradeoff. More positions usually lower the
volatility of returns (the diversification benefit) but dilute capital into lower-ranked,
lower-funding names (often lower return). Capital is split equally across the positions held.
Funding-Rate Arbitrage
Delta-neutral funding capture: buy the underlying on your broker (costs financing
to carry) and short a crypto perp (collect funding when positive).
Net carry = captured funding − financing, before transaction cost. The round-trip
cost is amortized over the holding period — a 90-day hold means paying the
round-trip fee about 4× a year, so shorter holds eat more of the funding you
capture — and it is shown separately in Tx drag and subtracted in
Net · turnover.
Each row is the best venue to short for that asset (highest funding among liquid venues).
Click the ▶ caret to compare all venues; click the asset name for its funding history.
Loading…
—
Funding rate (APY) by exchange
Smoothing
One APY funding line per exchange (right axis), smoothed
over the window you pick above (Hour = raw per-hour, then Day / Week / Month moving averages) so the
venues stay legible and the cross-venue spread is visible directly — the gap between lines is
the extra funding one venue pays over another for the same trade. The dashed breakeven hurdle
= financing + tx drag; a line above it = positive net carry there.
Click a label to show/hide a line. Scroll to zoom, drag to pan, hover for values.
Full cached history.
Best venue (Hyperliquid): each bar = that day's mean hourly funding, shown as
APY . Orange line = the Hyperliquid perp daily close.
Hyperliquid — each bar = that calendar week's mean hourly funding, as APY.
Green = longs pay shorts (you collect when short the perp); red = the reverse.
Hyperliquid — average funding by day of the week (Mon–Sun, UTC), as APY.
Each bar averages every hour that fell on that weekday. Watch for the carry fading or going
red at weekends, when the tradfi market is shut — a cue to trim the short leg.
Hyperliquid and Binance rows rank on the trailing mean over the chosen funding
window, fetched per market — a single hour annualizes to nonsense (ZHIPU's last hour reads
+146% APR against a seven-day mean of −3.5%). Bitget and MEXC have no equivalent public
endpoint here and stay on their current snapshot, so treat those two columns as indicative;
a leg still on a snapshot is never ranked on. Venues that block browser CORS fail gracefully
and show as unavailable.
Mode
Backtest
Single Indicator Optimization
Indicator Combination
Leverage Optimization
Seasonality
Regression Analysis
Return Autocorrelation
DCA Optimization
Rolling Window
Asset
BBTC▾
÷ Denominator asset
—Select asset▾
Sweep ranges
–
Mode settings
Indicators
A dead band around the crossing, in percent of
indicator 2. The signal has to run above it to go long and below it to go
flat or short, and inside it the previous position is held — so the level that turns the
trade on is not the level that turns it off. That gap is the hysteresis. It cannot
create a trade, only refuse one.
Threshold signals already have this by construction: an RSI buy at 30 and a sell
at 70 is a 40-point dead band. The band applies to the crossover path only.
Whether it is worth using is a question about your fee, not about the band.
Measured over four series, ten average lengths and four folds each — 156 comparisons
on data the setting was not chosen on — a band beats the unfiltered run 44% of the time
at 5 bps a leg and 68% at 40 bps. It buys fewer trades with worse timing, so it only
pays where trading is expensive. Break-even is near 20 bps a leg; the line under the
fields works out where your own numbers put you.
How many bars the new side has to hold before
it is acted on. A band filters by size, this filters by time — a move can be large and
last one bar, and the two catch different whipsaws. Applies to both crossover and
threshold signals.
The delay is real: entries land later and so do exits, which costs in a clean
trend what it saves in chop.
Same economics as the band, and the same break-even near 20 bps a leg. It is the
less reliable of the two: across folds a confirmation count swings much more than a
band does, because how many bars are enough is exactly the thing that changes between a
trending market and a choppy one. Treat it as an experiment rather than a setting.
Exposure & leverage
Optimal — Daily rebalance for long positions, set & forget for short positions. Best of both worlds. Daily Rebalance — Leverage is reset to target every day. Consistent exposure but higher fees in volatile markets. Set & Forget — Leverage is applied at entry and drifts naturally. Lower fees but exposure changes over time.
Compounding — Position size scales with equity. Gains compound but so do losses (volatility drag). Fixed — Always trade the initial capital amount. No volatility drag — reversing the signal gives exactly the opposite P&L. Useful for measuring pure signal quality.
Date range & capital
Configure parameters and press Run Backtest
Calculating…
Growth per leverage, reset every day
The sweep above holds the leverage fixed per trade; this curve resets the strategy to the same leverage every day and shows where growth stops.
How it works and settings
Up to the optimum L*, more leverage earns more; beyond it, less, and further right a loss, even when the strategy
itself wins. That is volatility drag: the return grows with L, but the damage of the swings grows with L² (after −50% you
need +100% to get back, and leverage makes those falls deeper). The line is the formula g(L) = L·a − L²·σ²/2 − costs, with a and
σ² from the strategy’s daily returns at 1×; the dots are what would have happened on the real days, with funding, fees and
liquidations. One history gives a rough L*: the lane under the chart shows how far the optimum moves from year to year.
—— strategy equity · —— buy & hold · ------ initial cash ·
● position change. Signals execute on the next bar (no lookahead); fees per position change; financing accrues daily on borrowed and short notional.
Optimizing combinations — 0%
Elapsed 0s
Sweep results
Best price prediction
Each row uses the regression that fits its mode
best. The model takes the most recent observed past lookback-day log return and projects a
forward log return using the regression's slope and intercept; the predicted price applies
that log return to today's close.
Raw values never leave the screen: every normalized
reading is printed beside the number it came from, and the method that produced it is named
with the property of the variable that chose it. A missing day is missing — never zero,
never carried forward — and a bar whose transformation cannot be formed is dropped and
counted. Below 30 observations no percentile is quoted at all; between 30 and 120 it is
quoted and marked limited, with its resolution stated. On a pair the same rule applies to the
intersection: a day on which either leg is missing is dropped rather than filled, and every
interval is computed on the observations that are genuinely separated rather than on the row
count. A conditional outcome is an association in the stored history — never a cause,
never a signal — and its sample is the separated draws, not the days.
Open interest & volume profile
—
Where positions were opened or closed, next to where market orders traded. Estimated from bars, not from individual trades.
Correlation Lab
—
/ of windows
ⓘ
Two prices are joined on the
stamp before anything is computed, and every return runs between the same two common days on
both legs — a weekend or a missing day is spanned by both or by neither, never filled and
never carried forward. Every rolling value comes from the window that ended on its own day.
Below 30 paired returns no coefficient is quoted; every interval is computed on the
observations that are genuinely separated, not on the row count. There is no lag, no search
over windows or lags, and no ranking of pairs: which pairs matter is a claim with a different
burden of proof, and a lead-lag found by scanning is found in noise about as often as not.
The one exception is Find diversifiers, which ranks on purpose and pays for it: a candidate
counts only when its result held in both halves of the window, and a candidate without enough
history is named with the reason, never dropped or filled.
Positioning Quadrants
—
The Daily flow report was retired on 9 Oct 2026 and no longer has a screen; its old address opens Positioning Quadrants.
De acht reeksen achter deze markt verouderen niet samen,
dus staan verse heid, bruikbare historie, ontbrekende dagen en de weggelaten vlakke aanloop
er per reeks bij — met de klasse van de bron. Niets in deze store is door deze app zelf
waargenomen: het zijn aggregaties van de provider, van een chart-studie gelezen.
Positioning Episodes & Historical Analogs
same coin · Liquidation Heatmaps = the CoinGlass model, kept per day when it is fetched, with Proprietary beside it (Daily archive: compare, or Proprietary for our maps only) · Positioning Heatmaps = our Net Positioning model on the CoinGlass series · Liquidation Ladder = CoinGlass’ estimate of the amount that would be liquidated at each price level, per exchange, for 1, 7 and 30 days
Across venues top 60 × Binance, Bybit, OKX, Bitget, Hyperliquid
Derivatives Scan
—
Narrow to ⓘ
This screen finds
states worth looking at. It does not say what to do about one, and nothing on it is a signal, a
direction or a trade. The catalogue is fixed: fourteen readings across eight evidence
families, each one a declared metric with a declared transformation, so no combination of metric,
transformation and horizon is ever searched for the one that looks most extreme.
Nothing is scored. There is no composite, no weighting and no vote — several readings off
one series are one angle, and the only count allowed to be read as “how many separate
angles” is the number of families: different data, which is not proof of statistical independence. Two extremities, never merged: how unusual a
reading is for this market comes back as a percentile against its own prior history, where this
market stands among the others comes back as a rank among today’s cross-section, and a
reading whose raw values are not comparable between markets gets no rank at all rather than one
that is really a table of market sizes. Breadth is the third question: how many other
markets are in the same tail today, which is what separates a discovery from the weather.
Missing is never zero — a gap is dropped and counted, a series behind the scan date is
refused rather than aged, a population under the floor produces no percentile, and a provider
restatement cuts the history rather than being ranked across.
A state is an episode, not a day: the same placement is computed for each of the 120 days
before the scan date, each against the population that existed on that day, and a state
is entered at the band and left only at the next looser rung — so a reading resting on the band
stays one state rather than being rediscovered every other morning, and a market that has been
extreme for six weeks says so instead of being announced again. The phases are not
directions and carry no ordering, no ranking and no score; a reading deep in its low tail
can be new and one deep in its high tail can be resolving. Novelty can refuse to answer:
an episode is only called a first where the trailing window reaches far enough back to show the
state was absent before it, and where it does not the answer is that it does not.
Horizons are a shape, never a vote: the 30-day change contains the 7-day one, so the
scale profile beside a change reading is one deviation seen at three sizes and only the horizon
you chose can put a reading beyond the band.
Liquidation Heatmaps
—
Orderflow Board
—
Board options
Bands on normalised panes
every normalised pane without a setting of its own, and the Correlation Lab's bands, read these three numbers (RC.FL.cfg); a pane with its own ⚙ setting is left alone
Arrangement
which panes, in what order, with what settings — named and kept; the layout chips on the board switch between the ones you saved
Mark this board
this market, this grain, this arrangement and this pin, with a note — something worth coming back to
Studies
everything stored for this market at this grain, by research family; what cannot be drawn stays on the list with its reason
Marked boards
every change is applied as you make it; the board keeps it
Downturn Accumulation
—
Increases payments after deeper falls, subject to the budget and supply stop. The separate tilt test did not establish a benefit beyond investing less.
Universe ⓘ
Event Explorer
—
A positive story is never a reason to ignore a valid stop.
Signals
Two kinds, as lists on this screen only: nothing is sent anywhere. Both describe what happened; neither is a trade signal.
Loading…
Cost and key
Catalyst list
Price changes by news kind
What the coins did around each kind of news, against the median of the other top-100 coins, against BTC or on their own, from the daily closes in this app's store. Observed association; not an estimate of the effect caused by the news. One case per coin per week; the rules were fixed before any result was seen. Under 30 coin-weeks a group is too few to judge.
Scan log
every scan, also the ones that found nothing
Rond de gebeurtenis: wat de koers deed
De toets: verschil met gewone dagen per termijn
Instellingen welke soorten meedoen, welke eisen, peildatum en catalogus
Doen mee ?
Meenemen als ?
Meer de toets per termijn, de tweede toets, wat bij wat lag, de catalogus en de methode
De toets per termijn, in detail
This screen
is a register of events and a catalogue over it. It says what happened, when, how
precisely that is known and where the knowledge came from — and it deliberately says
nothing about what happened next.
An event is not a state. A state is true of a stretch of time and a market that has
been crowded for six weeks is one thing on its forty-second day. What is an event is
entering the state, so every persistent thing here is one row at its onset carrying
the window it went on to occupy, never one row per day.
Three clocks, never one. Event time is when it happened, information time is the
earliest anyone could know, detection time is the earliest this terminal with this store
could have detected it. They are separate fields, separately uncertain, and where one is not
in the data it is empty with a reason rather than filled with the nearest other one. A daily
bar for a day is complete when that day is, so anything computed from one is detectable the
day after and not on the day.
Nothing is re-detected here. Positioning states, unusual readings and relative
dislocations are read from the contracts those modules publish. Their coverage is theirs and
it is printed; a module that has not run leaves its definition silent, which is not
the same as empty.
Dependence is a field, not a filter. Events of one definition landing within a day of
each other are one cluster with a count on every member. None of them is dropped — they
are all real — but forty markets falling on one Tuesday is one shock observed forty
times, and a study that counts forty independent observations there is wrong by about that
factor.
Who ran against bitcoin in a window you set
Hedged against bitcoin only, with β from the 180 closed 4-hour candles before the window, the same β as the burst alerts. The asset, the hedge model and Build further down belong to the daily view; this block needs only a window. A row opens its pair in that daily view, with bitcoin as the hedge.
Quick window, runs at once:?
What the columns mean
in $
the move in dollars, from the open at the start of the window to the close of its last 5 minutes
in BTC
the same move priced in bitcoin: above zero, the market did better than bitcoin
after β
what is left once the part that usually comes with bitcoin is taken off; β from the 180 closed 4-hour candles before the window
× normal
that rest divided by its own usual size for a window this long; orange from 2 either way
· 1 h
the second look: the same two with β from the 180 hourly candles before the window; a difference with the 4-hour columns comes from the hedge, not from the market
4 h after
in bitcoin, from the end of the window to four hours later
to the last closed hour
in bitcoin, from the end of the window to the last hour that has closed
order
click a column heading or use Order by; strongest on top, weakest apart below, the fixed rows with their place last
Against its sector your sectors, your rubrics and CoinGecko sectors · 7, 30 and 90 days
Your sectors Rotation Radar · 20 sectors, each coin on the exchange you chose · this week, 7, 30 and 90 days · 26 weeks
Compared in dollars, bitcoin, ether and against the other alts · spot and perpetual volume
Relative Valuations, by day
—
Discovery — where in the universe the residual has run
Worth a closer look when ?
This screen
describes one asset, one hedge model and the residual between them. It does not say what
to do about a residual, and nothing on it is a signal, a direction or a trade — the convergence
tester on Alt Strategies takes one of these objects and trades it; this is the research side of
the same object and it does not backtest.
The hedge is always visible. A spread chart is a picture of one asset minus another,
which is a coefficient of exactly one that nobody chose and nobody looked at. Where the real
relation is 1.3 that picture leaves 0.3 of the hedge inside the “spread”. Every
coefficient here is estimated, printed with an interval, and shown against the 1 the ratio
assumes.
Point-in-time is structural. The coefficient applied to a day is fitted on the days
strictly before it, so the residual is out of sample by construction; include the day and a
violent move partly explains itself and comes back smaller than it was.
Nothing is scored. How much of the move the hedge accounts for and whether the
coefficient stays put are two questions with two answers, printed in words with the rule that
produced them, and never added: a relation can account for most of a move and still have
reversed twice this year.
Missing is never zero — a day either side is missing is dropped and counted, a return is
never formed across a gap, a constant leading pad is dropped, and a pair whose relation cannot
be estimated on data that existed at the time produces no residual rather than a shortened one.
A convergence claim needs evidence for one. “The spread is two sigma wide” is
a statement about nothing unless the spread has somewhere to return to. The level reading
appears only beside an augmented Dickey-Fuller rung read against the test’s own critical
values and a variance ratio with a heteroskedasticity-robust interval, and where those say
nothing it says so.
Nothing here searches. The hedge models are a declared catalogue with a written reason
each; no part of this screen scans four hundred markets for the pair that looks most
dislocated. Comparing eight declared candidates is a comparison; the best of a hundred
thousand asset-hedge pairs is a different object that would look spectacular on pure noise.
An extreme residual is only worth as much as the hedge under it. The last block puts
today’s move under every declared candidate, because a move that is beyond the band
against one hedge and ordinary against the other seven is a statement about the hedge. The
diagnosis above it separates the four things that are otherwise one word: a hedge that
holds, one that is coming apart now, one whose level moved, and one that never held at all.
There is no best hedge and this screen will not print one. Neutralising the common
factor, keeping the residual quiet and keeping the coefficient still are different jobs that
often name different hedges; all three are shown, none is weighted against another.
And choosing is measured against how it would have had to be done. Picking the hedge
that turned out best over a history and then reporting how well it did on that same history
uses the answer; every candidate here is scored on the days ending before the one it is used
on, and the gap to the hindsight figure is printed as the bias it is.
A dislocation is two axes and they are never multiplied. How far the residual has come
loose from what the hedge explains is one measured thing; whether that hedge can be trusted is
another. Where the hedge is unreliable or came apart during the very days the residual ran,
the reading is less model reliability — not a stronger finding.
A percentile is context, never evidence that anything returns. Nothing here computes a
fair value, nothing assumes mean reversion, and the question of whether the spread comes back
has its own block, its own two diagnostics and its own refusal.
Successive extreme days are not successive findings. The residual’s own placement is
recomputed for each of the 120 days before the as-of date, each against the observations that
existed on that day, and a state is entered at the band and left only at the next
looser rung — the same episode machinery the scan cuts its states with.
Horizons are a shape, never a vote: the 90-day move contains the 30-day one.
And what the other modules published is read, never re-derived. The block about the two
legs’ derivatives and positioning states comes out of their records through
RC.anomaly and RC.positioning; it is context about the legs and is
counted toward nothing.
Volatility Screener
Crypto perps on Hyperliquid, ranked by recent realized volatility to surface long/short candidates.
Vol-expansion = 24h volatility ÷ 30-day baseline — the "volatile lately" signal: above 1× means the asset is more turbulent than its own monthly norm.
Volatility is annualized close-to-close: 1h vol from the last hour's 1-minute returns, 24h vol from 5-minute returns, 7d and 30d from hourly returns. Vol-expansion therefore sets a 5-minute figure against an hourly one, so it is not like for like: finer bars can read higher or lower than hourly ones. Vol-exp. 1h beside it takes the same 24 hours on hourly returns and divides by the same hourly 30-day baseline: like for like, but cut from only 24 returns, so coarser. Funding APR around +11% is Hyperliquid's fixed interest component (0.01% per 8 hours, about 10.95% a year): a real payment, not by itself a sign of crowding. Use the momentum (% change) columns to read direction, and the liquidity/funding columns to judge whether a name is actually tradeable. Click a row for the price chart and a per-timeframe breakdown.
Minimum 24h volume · turnover only ⓘ●—
—
Hyperliquid perpetual
—
Last 30 days of hourly close prices (Hyperliquid). Scroll to zoom, drag to pan, hover for values.
Volatility is annualized close-to-close realized vol over each window; the small figure under each one is the same measure without the yearly scaling, one standard deviation of an hourly or daily return (not a forecast of the move).
Vol-expansion above 1× means the asset is more volatile than its own 30-day baseline — the
“volatile lately” signal. Read the change rows for direction (long vs short), and funding for
the cost (or rebate) of holding the position. Per-window figures show n/a when there isn’t enough
candle history (newly listed markets).
Convergence Trades · legacy backtester
Mean reversion on the price ratio of two related commodities: when the ratio stretches beyond
its rolling average (entry at ±z), go long the cheap leg / short the rich leg and hold until
the ratio converges back. Dollar-neutral legs on Hyperliquid perps. Two data views:
Long history (daily futures) backtests decades of futures history with measured HL
funding means as constant carry; Hyperliquid only (real funding) uses only real
Hyperliquid perp prices and real hourly funding. The reference has a third,
Perp-like history + Hyperliquid perps, which is built on its server and cannot be
reproduced here — it is listed in the field below and greyed out.
Strategy configuration
Data and test
Signal
Execution
Result
—
Equity (3.5x gross, compounded, log scale)
━━ in the market ·
━━ flat (ratio inside its band) ·
━━ S&P 500 buy & hold
Ratio z-score · entry ±1σ, exit 0σ
Yearly returns vs S&P 500
Robustness sweep
Every cell is a full backtest over the same window with only lookback and
entry threshold changed. A strategy that only works in one cell is an artifact; the
research's WTI/Brent result was profitable in all cells out-of-sample. Here every cell is scored on
the same history it is shown for, so this grid is an in-sample sensitivity check: it shows how much
the result leans on these two settings, not that it holds on data it has not seen. The same goes for
Set optimal parameters, which copies the best cell of this grid into the form. The
Follow-up test below is the out-of-sample half: it picks on the first part of the window and
measures only on the rest.
Follow-up test
Leverage as an input
The same configuration at every gross leverage from 1x to 6x. Return and
worst drawdown both scale with leverage; pick the point whose drawdown you can actually sit
through. The single-run form above accepts any value in this range.
Recent trades
My Backtests
Stored on the machine that serves this page, so this is the
same list on the desktop and on the phone. Opening one restores every field and runs it.
My Collections (0)
Drag to pan · pinch to zoom · double-tap to fit
!
Are you sure?
Edit Backtest
Trade
Results & costs
Your real Bybit account first, in dollars from the account's own records.
Reading the account figures…
Chart: example profit added up, $10,000 per trade▾
Each closed example counts at a flat stake, nothing reinvested, so the line starts at $0. Hypothetical, before costs; not your account.
Execution & costs in detail fees per week and per coin, execution quality, the path inside each trade, every trade
Reading the figures…
Open positions
Live trades, unrealized P&L at current market prices.
Unrealized figures are estimated from current market prices and
refresh every minute while this screen is open.
Your trades after the exit▾
Your own closed trades, and what the price did after you closed them.
Everything except the ticker is optional. Leave Closed empty for a
live position; fill in the entry price and its unrealized P&L is marked to market.
What, and which way
?Pick an asset▾
–none▾
When
● goes in as a live position
What it did — fill in a result, or the prices and let it be computed
In your own words
The book
Trades live in this browser (localStorage). Export keeps a copy; import
replaces the book. On Lovable this is the one place that becomes a table.
Global Stocks
SHARES OF THE WORLD
RATIO VS PARITY AND ITS OWN MEDIAN
How to buy it
Every market against its own history
Biggest economies first. Click a column heading to
re-sort, or any row to load that market in the charts above. Sorting by Rel puts the markets
furthest below their own long-run norm at the top. Rel and Pctile are 2024 readings, the last
full year of each record: Ratio 2025E is a newer estimate, so Rel is not that ratio over the median.
Where this measure breaks down, in seven places
This is a blunt instrument. It is a decent way to spot where a market has drifted a
long way from its own historical relationship with its economy, and a poor way to decide
whether that drift should be corrected. None of this is investment advice.
Where a company lists distorts everything. Market cap is booked to the
exchange, not to the economy. Many large Chinese companies list in Hong Kong or New
York and never show up in China's figure, which understates China badly. In the other
direction, the UK figure counts FTSE 100 multinationals whose revenues are
overwhelmingly not British.
Private ownership varies enormously. Germany's Mittelstand and Italy's family
firms are privately held. That keeps their ratios permanently low for reasons that will
never mean-revert.
State ownership. In China and Brazil, large parts of the economy sit inside
state-owned entities with little or no free float.
Sector mix drives valuation, not geography. The US ratio is high in large part
because America hosts the world's software and semiconductor design companies, which
carry structurally higher multiples. That is not a mispricing.
It says nothing about earnings. A market can be small relative to its economy
because its companies are unprofitable, not because they are cheap.
GDP is measured in current US dollars. Currency swings move both the top and
the bottom of the fraction in ways that are not economically meaningful from one year
to the next.
Data quality. The World Bank market-cap series has gaps and outright errors
that had to be patched from exchange, central-bank and CEIC data. Patched years carry
roughly 5 to 10 percent uncertainty.
Sources and method. World Bank World Development Indicators: NY.GDP.MKTP.CD for
GDP in current US dollars and CM.MKT.LCAP.CD for the market capitalization of listed
domestic companies, full cap rather than free float. The raw market-cap series has gaps and
outright errors, so it was patched from exchange, central-bank, CEIC and IFC data, and the
world denominator was rebuilt to stay consistent with those patches. It tracks SIFMA's
global equity market-cap figures within about 3 percent. Europe here means the EU-27 plus
the United Kingdom, Switzerland and Norway, summed from member-level data. Dataset compiled
August 2026.
Copy Wallets
Scores on-chain wallets on Robinhood Chain the way a copy trader would have to: it
rebuilds every swap a wallet made from the chain itself, then asks what a follower who put
the same dollar amount into every one of its buys would have ended up with. A swap
there is a token leg against the Uniswap v4 PoolManager and a money leg the wallet paid or
received — native ETH on most of them, a stablecoin on some, and on the wallets that
trade through a pre-funded router, nothing at all until deep pricing goes and looks at
what the router paid the pool. Every trade is then valued at the ETH/USD price of the hour it
settled, so nothing is taken on trust from a leaderboard. Whatever still cannot be
priced is counted and shown as such rather than quietly dropped. The wallets themselves come
from a list you paste, or from asking the chain who has been trading a given token.
⚠ Three of the source’s columns are not here, because
they need a price history per token that the sources used here do not provide:
entry vs peak, early winners and the Early Bird style that is built on
them. Open positions are marked at the last price this wallet itself traded at, which
is the newest honest mark available and is stale for a token it left weeks ago. Wallets that
received tokens by transfer rather than by buying them are flagged: their cost basis is not
on the chain, so their own PnL is unreliable (cost basis unknown).
Settings who to score, and how the follower is simulated
Each wallet is read from the chain's own node: two event queries
for everything it ever moved, then its transactions in batches of twenty-five. A wallet with a few
hundred trades takes ten to twenty seconds the first time; the explorer is only used when the
node will not answer.
Rebuilding wallets from the chain…
Starting…
Change criteria
Copy size and delay are under Settings (they recompute all scores).
Funding Momentum
When a Hyperliquid perp's hourly funding prints deeply negative, shorts are paying
longs a premium to hold a coin that is being sold hard — and the selling usually
continues. This strategy shorts every coin whose settled funding crosses the entry
threshold (deepest first, up to the position limit), then holds each short through
and beyond funding normalization (“linger”) to harvest the continuation
drift. Position sizes are capped per coin by real order-book capacity (execution-cost
budget + share of hourly volume). Everything runs on this site's own hourly settled funding
and candle history; BTC/ETH/SOL excluded, listings younger than 14 days skipped, entries
gated on 6h momentum. 1× short margin only — no leverage. A long ETH leg
is carried alongside the basket, sized each hour from the held coins’ own volatility,
which takes the book’s market direction out: daily beta to ETH runs near zero hedged
against about −0.34 unhedged. The leg shares the same 1× rule — shorts plus
hedge never exceed capital — so it needs no extra margin and adds no leverage. Set
Hedge to 0 to see the unhedged book.
⚠ Execution costs are an approximation (flat fees + spread
plus a volume-scaled impact model with a selloff stress factor), and the backtest window is
a falling-alt-market regime — the basket has not yet lived through a sustained alt
rally, where entries would naturally dry up but lingering shorts would get stopped.
Drawdowns are measured on hourly closes.
Strategy configuration
Signal and exit
Size
Execution
Hedge
The first run of the day fills the server's hourly cache for
every market, which takes minutes; after that it is a file read.
Running the funding-momentum backtest…
Starting…
Building the hourly funding and price grids for every perp in the universe, then simulating
the basket hour by hour over the whole window. The first run after the server's daily cache
has gone stale walks Hyperliquid market by market and takes a minute or so; after that it is
a file read and lands in a few seconds. The result appears here by itself.
Result
Equity ($, fixed capital, net of modeled costs)
Capital deployed (% of capital in open shorts)
Current basket
Monthly returns vs shorting the alt market
“Short alts” = the inverse of an equal-weight index of every
coin in the strategy's universe (hourly rebalanced, no costs) — what a passive bet that
altcoins fall would have returned. “Excess” is the strategy's return minus that
benchmark scaled to the strategy's average deployment for the month: the part of performance
the funding-gated selection and timing added beyond alts simply doing badly.
Biggest coin contributions
Recent trades
Capital Rotation
Crypto's total capitalisation, split into Bitcoin, Ethereum, large alts
(the rest of the top ten), the long tail (everything outside it) and stablecoins
— five buckets that nest exactly, with no gap and no overlap. Pick any past window and the
screen splits each bucket's dollar change into the part the whole market explains and the
part left over, which is what it calls rotation. That column sums to zero, so
“out of one and into another” is an accounting statement rather than a figure of
speech.
Individual coins can be lifted out of the alt pool to stand beside Bitcoin and Ethereum in
the same balance — the partition still closes on the whole market, with
Rest of alts holding what is left. Per-coin capitalisations only reach back a year,
so ticking one narrows the window to that.
It is a performance attribution, not a flow. A bucket's capitalisation falls when its
price falls and nothing has to move anywhere for that to happen; nothing on this page can see
a transfer. Read it as “who won and who lost relative to the market, measured in
dollars”.
Loading the rotation dataset…
Window
Split out of the alt pool — last year only
Market share over time
Every day of the window as 100%, stacked bottom-up in the same order the
tables use. Hover for the composition on any single day; drag a box anywhere on the
chart to zoom into it — sideways narrows the window and re-runs every panel below, upward
also zooms the share axis so a two-point band fills the figure. Double-click steps back
out one level.
Relative capitalisation change
Rotation clock
How each bucket behaves
Rotation sequence
Whether the market moves first and the composition follows, or the other
way round — measured rather than assumed.
Market phases
The presets above are round numbers; the market's own turning points are
not. Total capitalisation is cut at every swing of the threshold from its running extreme,
which gives alternating expansions and contractions with real dates on them. Click a row to
load that phase as the window.