Methodology
A deeper look at how the scores and predictions are built — and how we test them so the numbers you see are honest, not curve-fit. For the plain-English version of what each metric means, see How it works.
A layered read of every perp
For every USDT-perpetual coin on Binance Futures we build three things daily: a Stress score (how fragile the setup is), a Pulsescore (how fast it's changing), and a set of model predictions(which way it leans, and whether a big move is brewing). Each layer answers a different question, and they're designed to be read together.
Stress & Pulse — composite reads
Both scores are 0–1 composites blended from several normalized market signals, so no single input can dominate and coins are comparable to one another.
Stress — structural fragility
Captures how much leveraged exposure a coin carries relative to its size, how violently positioning is swinging, and how much forced liquidation is already under way. High stress = the conditions for a cascade are stacked.
Pulse — rate of change
Ranks how unusual a coin's current positioning, funding and volatility are versus its own recent history, and how fast they're accelerating. Because it measures change, it tends to move ahead of the static stress read.
Both are published in full below. Every input is public exchange data, so a score you can't recompute is a score you shouldn't trust — you should be able to check our arithmetic.
Stress score
Stress = 0.50·LISn + 0.25·DBSn + 0.25·LCRn → 0…1 LIS = MarketCap / OpenInterest LISn = max(0, 1 − LIS / 5) DBS = |ΔOI 24h / OI| × |funding %| × 1000 DBSn = min(DBS, 1) LCR = (Liquidations 24h / OI) × 100 LCRn = min(LCR, 1)
LISis leverage relative to size — a coin whose market cap is only a few times its open interest is structurally fragile, which is why it's inverted and carries half the weight. DBS catches positioning swinging violently while funding is expensive. LCR is how much forced liquidation is already happening relative to the position base.
Validity rule: all three parts need a real, non-zero input. If any is missing the score is void— we show “—”, never a 0, because a missing input is not the same as a calm market. That's why some coins carry no score.
Pulse score
Pulse = Σ wᵢ·vᵢ / Σ wᵢ (over the parts that exist) → 0…1 0.20 |funding| percentile within its own 90d history 0.20 min( |ΔOI 24h| / 20% , 1 ) 0.20 min( |OI acceleration| / 10% , 1 ) 0.20 volatility compression (1 − ATR percentile; 1 = tightest range) 0.10 volume spike (≥ 3× the 30d median = 1) 0.10 min( |perp basis| / 0.5% , 1 )
Every part is a rarity measure — how unusual today is for thiscoin against its own history — not an absolute level. That keeps a $50M alt and BTC on the same 0–1 scale, and it's why Pulse tends to light up before Stress does: it reacts to change, not to state.
What stays unpublished is the machine-learning side: the feature list, model hyper-parameters and training recipe behind the direction and big-move predictions. The scores above are deterministic arithmetic on public data; the models are not.
How the predictions are made
The direction and big-move predictions come from gradient-boosted decision-tree models — an approach that consistently beats heavier deep-learning models on this kind of tabular market data at our scale. Each model learns from years of historical coin-days: for every past day it sees the market conditions and what price actually did next, and it learns the patterns that separate the moves from the noise.
Behavior-group routing
Coins don't all move alike. We group them by behavior (e.g. BTC-followers vs erratic low-caps vs funding-driven names) and let each group get predictions tuned to how it actually trades.
Calibrated probabilities
A raw model score isn't a probability. We calibrate the outputs so that when the model says “60%,” it's been right about 60% of the time on past data — the number means what it says.
Two questions, two models
One model estimates direction (up vs down); a separate one estimates big-move odds (a large move either way). Keeping them separate keeps each honest.
How we keep it honest
It's easy to make a model look brilliant on the data it was trained on. The hard part — and what most numbers on the internet skip — is proving it works on data it has never seen. Everything we publish is measured that way.
Walk-forward testing
We never score the model on days it learned from. We train on the past, then test on the followingdays it hasn't seen, and roll that window forward through history. Every accuracy figure on the site is out-of-sample by construction.
Edge over the base rate
A 65% hit rate sounds great — but if a big move happens 65% of the time anyway in that condition, the model added nothing. So we always compare against the “just guess the common outcome” baseline and report the edge over it. Only real edge counts as skill.
We only flag validated conditions
The Today's setups page and the big-move flags appear only for conditions where the edge is statistically significant— not just positive on a small sample. When the data doesn't support a claim, we show the model's lean as context and say so, rather than dressing it up.
It concentrates on high-stress coins
Be skeptical of any blanket “our model is X% accurate” claim — including ours. Across the whole market, the 1-day direction call is only ~55% right: real, but thin.
The edge concentrates where the service is pointed: high-stress coins. In those crowded, over-leveraged setups the hit rate climbs toward 60%+, and the big-move read is sharper still. That's the core thesis — the value isn't a market-wide oracle, it's identifying high-stress coins and reading them well, because that's where the probabilities pay off. Filter to high stress and high pulse, and act there.
What it runs on
Everything is built from public exchange data — open interest across multiple venues, funding, liquidations, volume and price — refreshed on a daily cycle.
The measurements behind all of this
Our research write-ups go through what we actually tested — including the parts that failed, the traps that produced fake results, and why we treat any accuracy figure without a baseline as meaningless.