COIL Index Methodology
The Coil Index is a 0–100 score that measures the probability of a major crypto market move in the next 24 hours. Refreshed every hour. Backed by six years of historical data. Out-of-sample validated. Look-ahead audited.
Methodology version v0.8-CF
What it predicts
The score answers one question: how likely is BTC or ETH to move significantly in the next 24 hours?
The model is trained against a 4% move threshold — the level at which leveraged positions start facing pressure, market makers widen spreads, and DeFi protocols reassess collateral buffers. The Calibration section presents results at both 4% and 5% thresholds. Same model, two operational yardsticks for different risk appetites.
4% — daily risk gauge. Routine risk windows. Higher signal frequency. Used for active risk sizing.
5% — cascade view. Severe events. Rarer, more catastrophic. Used for tail-risk hedging.
The Coil Index does NOT predict direction. It predicts magnitude — “is something big about to happen, regardless of which way?” This direction-neutrality is the most important methodological commitment in the index. A high reading means a sharp move is more likely. It does not say which way. Use Coil Index to size risk, not to decide long vs short.
How to read the score
| Score band | Tier | What it means |
|---|---|---|
| 0–15 | Calm | Markets are quiet. Hit rate well below average. |
| 15–30 | Quiet | Below-average risk of a 4%+ move. |
| 30–45 | Tense | Slightly elevated. Watch for catalysts. |
| 45–60 | Elevated | Above-average risk. Tighten stops. |
| 60–80 | Dangerous | High probability. Reduce leverage. |
| 80+ | Cascading | Highest risk band. Defensive positioning. |
Exact hit rates per threshold are in the Calibration section below.
Architecture
Coil Index v0.8 is a two-stage model. Each stage runs every hour automatically.
Stage 1 — Base score.A composite of three direct measurements of crypto-market fragility, blended into a single 0–100 score.
Stage 2 — Refinement.A second layer trained on six years of historical patterns refines the base reading using sixteen contextual features. The corrected score is the final v0.8 output, clipped to the 0–100 range.
The architecture is intentionally interpretable. Each stage has a clear theoretical basis, and the refinement layer's behavior can be inspected one feature at a time.
Stage 1 — Base components
Realized Volatility
Hourly realized volatility from major spot exchanges (BTC, ETH). The current 24-hour volatility is mapped to a 0–100 score via a fixed function calibrated against six years of empirical distribution. This component anchors the score in observable market behavior — when prices have been moving, the probability of further movement is mathematically higher. Volatility clustering is the most documented stylized fact in financial econometrics, formalized by Engle's 1982 ARCH framework.
Open Interest Velocity
Rate of change in open interest across major exchanges. Rapid OI buildup precedes liquidation cascades; rapid OI unwind precedes range-bound periods. This component captures the “leverage stacking” dynamic that turns ordinary corrections into cascading liquidations.
Liquidation Pulse
Recent liquidation activity, normalized as a 90-day percentile. A high percentile means recent forced unwinds are unusually large — a signal that the market is actively clearing leverage stress. This component captures cascade fuel directly, rather than inferring it from upstream signals.
The three components blend into a single 0–100 fragility score. Each component is independently meaningful. The composite covers volatility clustering, leverage build-up, and active cascade dynamics simultaneously.
Stage 2 — Refinement
The base score from Stage 1 isn't read in isolation. A second layer learns from six years of historical patterns to recognize when context shifts the meaning of a reading.
Why context matters
A Coil Index reading of 70 behaves differently before a CPI release than on a quiet weekend. Cross-venue funding divergence amplifies fragility. BTC-ETH correlation breaks signal regime shifts. The base composite is direction-neutral and treats all 70 readings the same — the refinement layer learns when 70 means “one thing” and when it means “something different.”
Sixteen contextual features
The refinement layer takes the base score plus sixteen contextual features at category level:
- Time-of-day patterns (US session, Asia session, weekend dynamics)
- Macro event windows (proximity to FOMC, CPI, major data releases)
- Cross-venue divergence (funding rate spreads across exchanges)
- BTC-ETH correlation breaks (regime shift indicators)
- Recent volatility regime context
- Other structural features capturing market microstructure shifts
The refinement layer is trained on out-of-sample data using strict walk-forward validation — train on years through N, test on year N+1, never the reverse. The model is interpretable and heavily regularized to prevent overfitting to spurious patterns in the training data.
Calibration
Coil Index is calibrated against 29,134 hourly out-of-sample predictions across four years (2023–2026). The table below presents per-threshold hit rates at every alert level, with Wilson 95% confidence intervals and lift over the no-skill baseline. Toggle between 4% and 5% target thresholds to see both the daily-risk and cascade-view calibration.
COIL Index — Calibration
29,134 hourly out-of-sample predictions · 2023–2026 walk-forward backtest
Daily risk gauge — leveraged positions begin facing stress
| Score | Hours flagged | % of all hours | Hit rate | 95% CI | Lift vs base |
|---|---|---|---|---|---|
| 70+ | 4,137 | 14.20% | 77.7% | 76.4–78.9 | 1.99× |
| 75+ | 2,781 | 9.55% | 83.4% | 81.9–84.7 | 2.14× |
| 80+ | 1,787 | 6.13% | 86.1% | 84.4–87.6 | 2.21× |
| 85+ | 1,021 | 3.50% | 89.0% | 87.0–90.8 | 2.28× |
| 90+ | 519 | 1.78% | 92.3% | 89.7–94.3 | 2.37× |
| 95+ | 263 | 0.90% | 98.1% | 95.6–99.2 | 2.51× |
Wilson 95% confidence intervals · Walk-forward backtest, no look-ahead · AUC reported as per-fold mean across 4 walk-forward folds (2023, 2024, 2025, 2026) · One model trained on 4% labels, evaluated against both 4% and 5% thresholds — same score, two operational yardsticks
Validation discipline
This is the most important section of the methodology. Most crypto data products claim accuracy without a validation framework. Coil Index publishes its validation framework explicitly.
Pre-commitment
Before every backtest, predicted AUC ranges and validation gates are locked in writing. After the backtest, realized values are published alongside the predictions. If we predicted +0.025 AUC lift and got +0.027, that's reported. If we predicted +0.020 and got +0.005, that's also reported. We do not retroactively adjust predictions to match results.
Walk-forward validation
The model is trained on years through N and tested on year N+1, never the reverse. No future information feeds into training. Each test year is unseen during model selection.
Look-ahead audit
Every feature is verified to use strictly past-only data at the moment of prediction. Rolling windows are backward-only. Calendar lookups are deterministic. Daily aggregates of partial in-progress days are explicitly forbidden. No backfill anywhere in the pipeline.
Anti-overfit rules
After seeing results, the following are explicitly prohibited:
- Adjusting weights to improve specific event scores
- Lowering thresholds to reclassify failures as passes
- Cherry-picking which events count
- Re-running with different windows until one passes
- Adjusting predictions to match realized values
These rules are enforced through methodology versioning. Every methodology change gets a new version label, and historical comparisons stay clean.
What we tested and rejected
Honesty about misses is part of the methodology. After locking the v0.8 architecture, we tested five additional feature classes to see if any could push the index further. All under pre-committed AUC ranges with walk-forward validation.
| Feature class tested | Realized lift | Verdict |
|---|---|---|
| Cross-asset (DXY, VIX, US10Y) | −0.0043 | Rejected |
| Per-feature cross-asset variants | −0.0017 to −0.0019 | Rejected |
| Liquidation cluster density | −0.0049 | Rejected |
| L/S global cohort divergence | −0.0006 | Rejected |
| Coinbase Premium 12-feature bundle | −0.0021 | Rejected |
Each test pre-committed AUC ranges in writing, walk-forward held out test years, audited for look-ahead, and reported the realized number honestly. Five separate orthogonal feature classes failed to clear the threshold. Coil Index v0.8 captures most recoverable signal from free data sources at this point in time.
Methodology versioning
The Coil Index methodology is versioned aggressively. Every meaningful change gets a new label.
- v0.6, v0.7, v0.8 — major architectural shifts
- v0.X-X — incremental refinements (e.g., v0.8-CF for the current Config F)
Historical backtest data under previous methodologies stays in our database, never silently mixed with current data. Each public methodology page reflects the live production version.
Methodology change without versioning is the most common form of validation theater in financial data products. By versioning explicitly and publishing the progression, every claim about Coil Index can be traced to the exact methodology version that produced it.
Direction-neutrality (the locked rule)
Coil Index measures fragility, not direction.
A high reading means a sharp move is more likely — but not which way. A market with crowded leverage on both sides can break in either direction depending on the trigger. The index reads the structural fragility; the trigger and direction are exogenous.
Use Coil Index to
- Size risk and adjust position sizing
- Tighten or loosen stops
- Decide whether to hold cash or stay deployed
- Time when to reduce leverage
Do not use Coil Index to
- Decide long vs short
- Predict the direction of the next move
- Time individual trades
Honest limitations
Three things Coil Index does not do well.
Predict direction
Already covered above. The index is structurally direction-neutral by design.
Predict timing within the 24-hour window
A reading of 80 on Monday morning means the next 24 hours have elevated risk — but not whether it manifests at noon or at 11:55 PM. Sub-hourly precision is not currently measured.
Distinguish between trigger types
A liquidation cascade triggered by macro news, a DeFi exploit, and a low-liquidity weekend wick all look the same to the index ex ante. The index reads structural fragility, not the specific catalyst.
We are working on signals that address all three. Updates ship as new methodology versions, with full transparency on what changed.