How it works

How the odds work, and how we test them.

yoru gives every trade two numbers. Target first is where you stand on a line, and needs no forecast. Liquidation in 24 h is a clock: it needs how fast the coin moves today and how real prices fall. Try both on one trade below, then see how each held up on data the model never saw.

BTC

Side
Leverage
Target
1

The line

Draw a line from your liquidation price to your target. Your chance of reaching the target first is your cushion divided by the whole line. That holds whenever the price is a fair game, whatever the volatility.

entry
liquidationtarget
cushionto target
P(target first)=cushioncushion+distance to target

Waxing crescent. The odds are against you.

2

Why volatility can’t move it

Each line is one way the price could move from here, with no trend up or down. It stops when it reaches your target, a white dot along the top, or your liquidation, an orange one along the bottom. Count the white dots and you have your chance. It lands on the line’s number.

Now turn the volatility up. The same futures replay on a faster clock: the whole picture takes less time, the shaded first day covers more of it, and more futures end inside it. They still end in the same places. Volatility changes the clock, not the odds.

Volatility

Playing the futures forward…

target

liquidation

When the tide drifts, it drifts away from the nearer end: the futures that headed there finished first and left it. Each future on its own has no trend.

3

The clock

The line says which end comes first, never when. For the 24-hour number yoru needs how fast this coin moves today and how real prices fall. Real prices are not the smooth walk above: they sit still, then jump.

Chance this trade touches liquidation in the next 24 hours

Textbook walk
–
yoru, calibrated
–

How yoru’s number was graded
Read the moon

Every chance yoru shows is drawn as a moon lit by exactly that share of its face, and the light is always the good outcome. Our logo is a 20x BTC long aiming for +10%: lit just over a quarter.

Accuracy

Graded on 2.8 million forecasts it never saw.

Like a rain forecast, yoru is graded on what happened next. Every quarter from January 2021 to September 2026, each model was refitted on earlier prices only, then scored on the next three months: eleven coins, every hour, LUNA’s collapse included. The rules were written down before any data was downloaded.

98liquidation touches happened for every 100 yoru expected. On twelve coins it never trained on: 99.
20 of 20leverage and target settings where the target-first odds held within 5%
11xas often as the textbook formula said, when it called a move under 1% likely
In the backtest, when yoru said 10%, it happened about 10% of the time.Every 24-hour forecast, grouped by what was predicted. Dots on the dashed line are perfectly calibrated. Where the textbook formula said 0.1%, 1.4% happened. On twelve coins it never trained on, 99 happened per 100 expected. It runs cautious only at the biggest moves, 15 to 20% away: 76 per 100.
0.1%1%1%10%10%100%100%it happenedyoru saidyorutextbook formula0.1%1%1%10%10%100%100%it happenedyoru saidyorutextbook formula
The method in full: five models raced, every formulaHide the method

Five models raced. One shipped.

Sharper means its odds scored better than the textbook’s on the same 2.76M forecasts. Longer bar, better.

Honest on big moves: of every 100 moves of 15 to 20% a model promised, how many came. 100 is honest; the dot should sit in the middle.

  1. 1
    HAR-RVBlends the last day, week and month of swings.
    +1.5%
    79
    ShippedLive: per coin for the ten majors, pooled for every other coin
  2. 2
    yoru v1Real price history, scaled by the last two days of swings.
    +0.8%
    69
    BackupCoins under a month old
  3. 3
    GARCHThe Wall Street classic for forecasting volatility.
    +0.6%
    66
    OutWorse than yoru v1
  4. 4
    High-lowReads each hour’s whole range, not only the close.
    +0.5%
    69
    OutWorse than yoru v1
  5. 5
    The textbookThe Black-Scholes probability of touch from options trading, given the same volatility as yoru.
    baseline
    144
    BaselineWhat we measure against

yoru v1 was replaced for the majors because it overstated the biggest moves, 15 to 20% away; on overall sharpness the two are close. A pooled version then replaced it for every other coin with a month of prices, after a pre-registered test on twelve fresh coins. The fanciest model was not the best one. Sharpness is the Brier score, the grade weather forecasters use. Target-first odds need no model: they follow from the liquidation rule, and the chart above shows they hold.

These run in the bot and on this page as the same code, and an automated test checks both give the same answer to the ninth decimal. S is the mark price, L your leverage.

Hyperliquid ruleLiquidation distance
d=L−1−m1∓m, m=12Lmax
Maintenance margin is half the initial margin at the market's maximum leverage. Upper sign for longs, lower for shorts. The liquidation price is S(1 − d) for a long and S(1 + d) for a short. Exact below the first margin tier, which starts between $20M and $150M of position by market.
Fair-market oddsTarget before liquidation
p=dd+t
t is the move to your target. When the mark price is a fair game, optional stopping makes this exact, whatever the volatility.
Calibrated modelLiquidation within 24 hours
p24h=F(ln(1−d)σ√1/365)
F is the distribution of 143,312 real 24-hour falls in twelve Binance perps from 2019 to 2026, LUNA and FTT included, each divided by the volatility forecast at its start. Shorts use the rises the same way.
HAR-RV, Corsi 2009Volatility
ln RV+1=b0+bd ln RVd+bw ln RVw+bm ln RVm
RVd sums the squared hourly returns of the last day; RVw and RVm the last week and month, per day. Fitted per coin for ADA, AVAX, BNB, BTC, DOGE, ETH, LINK, LTC, SOL, XRP. It beat the 48-hour EWMA in 6 of 6 years and was the only volatility model whose big-move calibration passed. Every other coin with a month of prices uses one pooled fit: on twelve fresh coins it never trained on, it was slightly sharper than the EWMA on all of them and calibrated about as well.
Generalised ParetoBeyond the data
F(z)=0.025(1+ξz0−zβ)−1/ξ
For moves past the outer 2.5% of the sample. Fitted on the same data: ξ = 0.31, β = 1.37 for falls and ξ = 0.36, β = 0.78 for rises. The tail is heavier than a normal curve’s, so crashes are rare, never impossible.
48-hour EWMAVolatility
σ2t=λσ2t−1+(1−λ)r2t, λ=2−1/48
r is the hourly log return, from 30 days of Hyperliquid candles, annualised by √8760. Used for coins with less than a month of prices.
Session clockVolatility, US stocks
σ224h=L ∑h w(h), L=EWMA(r2/w)
w(h) is how much a stock usually moves in each hour of the week on the New York clock: opens, closes, nights and weekends. The next 24 hours are priced by the hours they actually contain, with a fat-tailed table of real stock moves. Used for US single stocks with a month of prices; other stock-like markets keep the 48-hour EWMA and the lognormal formula.
New listingsMarkets under 30 days old
p=Fnew(z)
A coin in its first 30 days is read through tables built from other coins’ first 30 days (107,120 24-hour windows), one for the first week and one for the rest. The majors’ table warned too often about a new listing. Tested on 462 coins never used before (1,705,396 forecasts): touches against warnings went from 0.93 to 0.97, in a coin’s first week from 0.87 to 0.95, and the number was sharper on 364 of 462 coins.
Hourly rangesVolatility, coins outside the ten majors
ln σ2=b·(day, week, month, last 6 h range, last hour range)
For a coin outside the ten majors with a month of prices, the forecast also reads how far each of the last six hours travelled between its high and its low. A market that has just woken up, or just gone quiet, shows in those hours before it shows in the day. Learned from 150 coins and tested on 50 others never used before (5,076,060 forecasts): the number was sharper on 50 of 50, by 0.85% of the Brier score, and by 0.66% when the prices it reads are an hour old. A second set of 50 unseen coins gave the same: sharper on 49 of 50, by 0.84%. On Hyperliquid’s own prices (69 coins, 141 days) the gain holds. The coefficients in use were then refitted on 200 coins.
Jump memoryThe coin’s own biggest days
logit p=logit p1+γ(largest day, distance)
For a level 8% or more away, the chance is moved by the coin’s largest daily rise in the last 90 days (a level above the price) or its largest daily fall (a level below it). A coin that has jumped 40% in a day reaches far levels above the price more often than the tables say; far levels below the price are reached less often than they say. On the test coins, touches against warnings for far levels above the price went from 1.06 to 1.04, and below the price from 0.86 to 0.91.
Far levelsMore than 20% from the price
δ→w(distance)·δ, when δ<0
Most large positions keep their liquidation price more than 20% away, further out than the levels the layers here were learned on. Out there a layer may still raise a chance. A shift that would lower one fades out between 20% and 40%. Carried unchanged, those shifts made the model warn too seldom about far levels above the price. Checked on 100 coins never used before, in two sets of fifty (3,685,080 forecasts at 25% to 50% away): touches against warnings above the price went from 1.29 to 1.07 on one set and from 1.41 to 1.14 on the other, so above the price the model still warns a little too seldom. Below the price they stand at 0.61 and 0.70: there it warns too often. Beyond 50% nothing has been tested.
Weekly clockVolatility, by day of the week
σ24h=σ·s(h)
s(h) is how much the next 24 hours usually move when they start in hour h of the week, against the forecast: about 0.79 on a Saturday, 1.17 on a Sunday. Crypto is quieter at weekends, and a forecast built from the last days does not know what day it is. Tested on 12,776,148 forecasts on 100 coins it had never seen: on Saturdays, touches against warnings went from 0.83 to 1.01.
Level memoryHow long since the price was there
logit p=logit p0+δ(age, distance)
The age is how long ago the price last traded at your liquidation level: within a day, three days, a week, two weeks, one, two or four months, or not at all in four months. A level the market has not visited for weeks is hit more often than the tables alone say once the price is close. A level the price left days ago and is now far from is hit less often. The shifts δ are learned from the model’s own misses over the last two years and were tested on 50 coins it had never seen (6,293,826 forecasts): for levels untouched for two weeks, touches against warnings went from 1.12 to 0.99. Used for coins with a month of prices.
Earnings daysUS stocks
σ→mσ, m=2.09
When a company reports inside the next 24 hours its stock moves about twice as much as on other days, so the volatility is multiplied by m for those 24 hours. Fitted on reports before 15 June 2026 and tested on those after: touches against warnings went from 1.66 to 0.96.
Brier scoreHow forecasts are graded
B=1N∑(pi−yi)2
p is the chance yoru gave, y is 1 if the level was touched. Lower is better. Weather services grade rain forecasts the same way.
Diebold-MarianoBetter, not lucky
DM=d̄seHAC(d̄)
d is the daily difference in Brier scores; the standard error allows for 20 days of correlation (Newey-West). HAR-RV against the previous model: t = -5.3 over 2,094 days.
Share of marginCosts
fees=2fL, funding a day=24ρL
f = 0.045%, the taker fee per side at the base tier. ρ is the current hourly funding rate; longs pay when it is positive.
Share of marginOutcomes
win=tL−2fL, loss≈dL+2fL
An order-book liquidation returns the leftover margin, so you lose the move times your leverage. A backstop liquidation keeps the rest.
Lognormal modelStock perps
p24h=N(a+v2/2v)+e−aN(a−v2/2v)
a = ln(Sliq/S) and v = σ√(1/365): first passage of a zero-drift price. Used for stock perps until the calibration covers them.

The bot is live.

Free in Telegram, in a private chat or in your group. Type a trade like btc 20x 10% and get the chance and the card.