Volatility Pulse
Will the next few minutes be calm or violent? A live forecast of short-horizon volatility for BTC and ETH — deliberately never direction, which the same research measured as untradable at these horizons. Use it to time entries into quiet windows and to get early warning while short premium.
- P(move > 0.21%)
- 1%(0.0× base 10%)
- Forecast bar sd
- ±0.05%
- Last 5m bar moved
- ±0.01%
Bottom-quintile forecast — big moves are historically rare here.
- P(move > 0.36%)
- 3%(0.3× base 11%)
- Forecast bar sd
- ±0.12%
- Last 15m bar moved
- ±0.13%
Mid-range forecast — nothing unusual priced into the next bar.
- P(move > 0.30%)
- 1%(0.1× base 8%)
- Forecast bar sd
- ±0.07%
- Last 5m bar moved
- ±0.01%
Bottom-quintile forecast — big moves are historically rare here.
- P(move > 0.51%)
- 2%(0.2× base 9%)
- Forecast bar sd
- ±0.15%
- Last 15m bar moved
- ±0.04%
Bottom-quintile forecast — big moves are historically rare here.
How traders use this
- Entering premium sales: a CALM reading (bottom-quintile forecast) has historically meant a <1% chance of a big bar — the market is unlikely to run through your strikes while you leg in. Pair with the settlement-hour window on the Movement Odds page.
- Defending short options: a jump to STORM is a measured early warning — big-bar odds in the top bucket ran 33–38% versus a ~9–11% base rate, out-of-sample. Not a prediction of direction; a prompt to check your risk.
- What it is not: a trade signal. There are no 5-minute options, and chasing this with taker-fee futures scalps feeds the fee schedule, not you.
About the Supertrend row
Each card shows the classic Supertrend (10, 3) state for its timeframe — because many traders use it, and because it pairs naturally with the pulse: Supertrend is an ATR trailing stop, a volatility-scaled level that answers “where would I be wrong?”, and the pulse forecasts the very volatility that sets its width.
What our own backtest says: as a directional signal on this venue, Supertrend failed — 48 configurations across 5m–1h timeframes, and zero beat taker fees over 577 sessions (the least-bad variants only approached breakeven with maker fills and flip-only entries). We show the trend state and stop level as risk context: use the line to place and trail stops, not to predict the next move.
Does it tell the truth? (out-of-sample scorecard)
Each model was validated on a held-out chronological test set (~23k–72k bars per combination, 2.3 years of tape). The table shows the observed frequency of a “big bar” in each bucket of the model’s forecast — the same mapping the live gauges use, frozen at training time.
| Model | OOS R² | Rank IC | Big bar = | Base rate | P(big | bottom decile) | P(big | top decile) |
|---|---|---|---|---|---|---|
| BTC · 5m | 0.52 | 0.71 | >0.21% | 10.4% | 0.5% | 37.8% |
| BTC · 15m | 0.63 | 0.79 | >0.36% | 10.6% | 0.8% | 37.4% |
| ETH · 5m | 0.49 | 0.69 | >0.30% | 8.5% | 0.5% | 33.0% |
| ETH · 15m | 0.60 | 0.77 | >0.51% | 8.7% | 0.7% | 32.4% |
Method: per-bar realized variance from 1-minute returns; HAR-style model (last bar, ~30-minute mean, ~4-hour mean, hour-of-day seasonality); chronological 70/30 split. Volatility clustering is the most replicated effect in market data — the skill here is real but ordinary; the honesty is showing you its exact calibration. Models trained through 2026-06-23 (BTC) / 2026-07-26 (ETH); see methodology.
Educational analytics, not investment advice. The pulse forecasts movement magnitude only — never direction — and can be wrong on any given bar. Probabilities are historical calibrations, not guarantees.