Methodology · v1.0-EXT
How the Coil Index works
A five-axis composite of crypto market stress, refined by a small machine-learning residual, calibrated on 11,154 hours of true out-of-sample data. This page documents what the score does, how it’s built, what it’s been tested against, and what it cannot do.
What the Coil Index predicts
The Coil Index estimates the probability of a sharp BTC or ETH move within the next 24 hours. The score itself is bi-directional — a high reading means a sharp move is more likely, not which way it’ll go.
The model is trained on a single binary label: did either BTC or ETH move ≥ 4% in absolute terms over the next 24 hours? We then evaluate the same score against three thresholds, each useful for a different kind of decision:
- 3% — Common move threshold. Captures the broader range of meaningful 24-hour moves. Useful for swing positioning and intraday risk awareness.
- 4% — Daily risk gauge. The training target. Leveraged positions begin facing real stress at this magnitude. This is the threshold the model is optimized for.
- 5% — Severe cascade view. Out-of-sample threshold for tail-risk decisions. Leveraged positions face significant stress; liquidation cascades become possible.
Calibration tables for all three thresholds, broken out per-asset (BTC, ETH, Combined), are on the homepage.
How to read the score
The 0–100 score maps to six tiers. The tiers aren’t model outputs — they’re a reading aid. The number is what matters; the tier label just makes it interpretable at a glance.
| Range | Tier | What the market looks like | Frequency |
|---|---|---|---|
| 0–20 | QUIET | Realized vol compressed, OI stable, funding flat, options pricing in low stress. Sharp moves are unusual but not impossible. | Common |
| 20–45 | NORMAL | Typical two-way flow. No single stress axis registering meaningfully above baseline. | The default state |
| 45–60 | ELEVATED | One or two stress axes printing above baseline — usually realized vol or OI velocity. Worth watching but not actionable on its own. | Periodic |
| 60–70 | TENSE | Multiple axes co-elevated. Funding, OI, or vol structure showing stress. Probability of a sharp 24h move is meaningfully above baseline. | Uncommon |
| 70–85 | DANGEROUS | Several stress axes co-elevated. Conditions resemble historical pre-event setups (sustained FOMC week, leverage build before unwind). | Rare |
| 85–100 | EXTREME | Composite stress at the upper tail. Conditions resemble historical cascade pre-states (LUNA, FTX, October 2025 deleveraging). | Very rare |
The tier you see in the gauge always reflects the latest score. Hourly tier changes are normal during volatile periods.
Architecture
The Coil Index is a two-stage model. The two stages do different things and are kept separate on purpose.
Stage 1 — Transparent composite. A weighted sum of five stress axes that any quantitative crypto trader could compute themselves: realized volatility, open interest velocity, liquidation pulse, funding-price divergence, and options market stress. Stage 1 is the backbone of the score — interpretable, auditable, and resilient to the kinds of regime changes that break opaque models.
Stage 2 — Residual refinement. A small XGBoost model that corrects Stage 1’s known blind spots — primarily around weekend liquidity, calendar effects, and tradfi macro events (FOMC and CPI release windows). Stage 2 modifies Stage 1’s output by ±15 points maximum; the cap exists on purpose so a single ML failure can’t override the transparent composite.
final_score = clip(stage1 + stage2_residual, 0, 100)Stage 1 components
Realized Volatility (RV)
What it measuresThe annualized standard deviation of BTC and ETH log-returns over the trailing 24 hours.
Why it mattersRealized volatility is the cleanest backward-looking measure of how much the market is actually moving. Empirically, it’s the single best leading indicator of sharp moves over the next 24 hours — markets that are already moving tend to keep moving. It’s also the most resilient to manipulation: unlike funding or OI, you can’t fake realized vol because it’s computed directly from observed prices.
Open Interest Velocity
What it measuresThe rate of change of total perpetual-futures open interest across major venues (Binance, Bybit, OKX, Deribit). Both rapid build-ups and rapid unwinds count.
Why it mattersOpen interest is the stockpile of leveraged crypto positions. When it’s expanding rapidly, leverage is being added — usually faster than spot-market liquidity can absorb if those positions reverse. When it’s contracting rapidly, leverage is being forcibly closed (often via liquidations). Both directions raise the probability of a sharp move because they signal that the market is leveraged, fragile, or unwinding.
Liquidation Pulse
What it measuresThe total volume of forced liquidations across major perpetual-futures venues over the most recent complete day, ranked against historical liquidation totals.
Why it mattersLiquidations are the cleanest evidence that leverage has already broken. A high liquidation pulse means traders just got forcibly closed out — and historically, large liquidation days cluster: one day of cascading liquidations often precedes another, because the surviving leverage is now sitting on shakier ground. The signal isn’t predicting the cascade itself; it’s measuring the residue of recent stress, which is empirically a leading indicator of more stress.
Funding-Price Divergence
What it measuresWhether leveraged traders are positioned in the opposite direction of where price is actually moving. Specifically: positive funding (longs paying shorts) while price is falling, or negative funding (shorts paying longs) while price is rising.
Why it mattersFunding rates show what crowded leverage looks like; price movement shows what the market is actually doing. When the two disagree, leveraged positions are getting hurt — longs who paid to be long are watching price fall, or shorts who paid to be short are watching price rise. Either case is a setup for forced unwinds: the longer the divergence persists, the more positions are sitting on losses, and the higher the probability of a liquidation cascade in either direction. This is the cleanest forward-looking stress signal in the composite.
Options Market Stress
What it measuresWhat the BTC and ETH options market is pricing in for the next two weeks — specifically through two signals: how expensive short-dated implied volatility is relative to the past year, and how much more traders are paying for downside protection (puts) versus upside speculation (calls).
Why it mattersThe options market is where sophisticated traders express forward-looking conviction with real money. When implied vol gets expensive or put protection gets bid up, it means the most informed segment of the market is buying insurance — they see something. Options pricing tends to move ahead of spot during periods of building stress, which makes it a useful early-warning component.
Calibration
Calibration measures whether the score’s promises are honest. A score of 70 should mean something specific, repeatedly, on data the model has never seen.
The Coil Index is calibrated on a strict 16-month HOLDOUT — the period from January 2025 through April 2026 — covering 11,154 hourly out-of-sample predictions. The XGBoost residual layer was trained on data ending December 31, 2024. Everything in the holdout is data the model has never been exposed to during training, validation, or threshold tuning. No retroactive adjustment, no peeking, no overfitting.
The calibration tables on the homepage are broken out three ways:
- By move threshold — 3% (common move), 4% (daily risk gauge, the training target), 5% (severe cascade view).
- By asset— BTC, ETH, and Combined (whichever moved more).
- By score band — 40+, 50+, 60+, 70+, 80+, 90+, 95+.
For each combination, four numbers are reported:
- Hit rate— the percentage of hours at that score band where the threshold was met within the next 24 hours.
- Base rate— how often the threshold was met across the full holdout, regardless of score. The number to beat.
- Lift— hit rate divided by base rate. A lift of 2× means the score doubles your odds vs. coin-flipping the holdout.
- 95% confidence interval — Wilson score interval. Wider CIs at the top score bands (95+) reflect smaller sample sizes — the highest-conviction signals are also the rarest.
Across the full holdout, the score crossed 70 in 13.6% of hours, 90 in 1.8%, and 95 in 1.2% — a selective alert distribution by design. The score is meant to spend most of its time below 60.
What we tested and rejected
The path to v1.0-EXT involved discarding several candidate approaches. A short, honest list of the ones worth knowing about:
| Approach | Why it was rejected |
|---|---|
| Pure deep-learning model (transformer-based predecessor) | Predicted comparably on average but failed on the events that mattered most. Sophisticated stress regimes need interpretable axes more than they need parameter count. |
| Direction-specific scoring (separate “crash” and “rally” scores) | Crypto’s biggest moves are bi-directional (LUNA was a downside cascade; Q4 2024 was an upside cascade with similar fragility signatures). Splitting the score doubled the noise and halved the signal. |
| Multi-horizon scoring (12h, 24h, 48h, 72h) | The 24h horizon dominated. Shorter horizons were too noisy; longer horizons absorbed too many unrelated catalysts. We ship one horizon and keep it honest. |
| Put-Call Ratio (PCR) | Not rejected — deferred. PCR is being captured by the live options collector and will be evaluated as a candidate feature in v1.1. It was excluded from v1.0-EXT to avoid expanding the feature set during the calibration window. |
The point of disclosing these isn’t completeness. It’s that any score that claims to predict crypto moves should be able to name what it gave up on the way to its current form. Signals that survive ablation are stronger than signals that were never tested against alternatives.
Methodology versioning
The Coil Index is versioned. Every meaningful change to the score’s computation gets a new label, and the version that produced any given historical reading is preserved. Today’s live score is v1.0-EXT— a five-axis composite with an XGBoost residual layer, calibrated on a 16-month true holdout. Earlier versions exist in the production database for comparison but are not the live signal.
Historical readings are not retroactively rewritten when the methodology changes. The chart on the homepage shows v1.0-EXT scores backfilled across the full historical window for consistency, but the production system retains every methodology’s readings independently.
Direction-neutrality
The Coil Index does not predict direction. A reading of 78 means a sharp move is more likely than baseline — not whether it’ll be up or down. We made this design choice deliberately:
The most violent crypto events of the past five years split roughly evenly between upside and downside cascades. Forcing a directional prediction onto a fundamentally bi-directional signal would have meant either training two separate models (which we tested and rejected, see above) or biasing the score toward the regime that happened to dominate the training window. Both paths weaken the signal.
If you want a directional view, combine the score with your own thesis: spot positioning, funding skew, options flow, or whatever you anchor on. The Coil Index tells you when the market is fragile. What you do with that is yours.
Honest limitations
A score that promises to predict every sharp crypto move would be lying. A few things the Coil Index does not do:
- It does not catch every move. Across the holdout, 43% of 4%+ moves were preceded by a 70+ reading. The remaining 57% — moves driven by single-event catalysts (regulatory news, exchange outages, individual-asset shocks) — are largely outside what a market-fragility composite can see. Recall is a tradeoff, not a goal in isolation.
- It cannot predict direction. Bi-directional by design. See above.
- It is calibrated for BTC and ETH. Other assets often correlate, but the score is not validated against ALT-specific fragility. Use it as a market-wide gauge, not a single-asset signal.
- It is a probability, not a guarantee. Even at score 95+ — where the holdout shows a 100% hit rate at 4% combined — the sample is 133 hours over 16 months. Real-world deployment will surface false positives. Plan for that.