Research protocol
Scope and protocol specification
Working paper · prequential probability forecasting · status: preliminary
Abstract
This study evaluates, in real time, whether parametric two-sided Weibull tail models yield calibrated probabilities for multi-day price corridors of digital assets. Each forecast is fixed and hashed before its outcome is observable and enters an append-only ledger. Evaluation follows the prequential principle (Dawid, 1984): the model is judged only on forecasts issued before the data that score them.
1Research questions
RQ1. Do emitted corridor probabilities match observed frequencies, as measured by the Brier score and reliability diagram?
RQ2. Does calibration hold across volatility regimes without ex-post refitting of the emission rule?
2Specification
| Estimand | Probability that the close at horizon h remains within a stated price corridor, h ∈ {1, 3, 7} days |
| Model | Two-sided Weibull fitted separately to upside and downside daily log-return tails |
| Emission rule | C+cap: target probability 85%, total width capped at 2·W_max·√h; observation-only below 65% |
| Universe | Ten liquid digital assets, Binance spot, daily candles (00:00 UTC) |
| Evaluation | Brier score (all and 10–90% subset), reliability diagram, regime-conditional split |
| Out of scope | Directional signals, portfolio advice, intraday horizons, ex-post recalibration |
3Integrity protocol
Pre-registration. Parameters and corridors are written before the target candle opens; each record carries a SHA-256 hash of its original payload.
Non-selectivity. All forecasts are retained and scored. No record is edited, filtered or removed.
Reproducibility. Outcomes reference public daily closes and can be checked independently against the stored hash.
4Sample and statistical power
- First issue date
- —
- Issue days
- —
- Recorded forecasts
- —
Forecasts issued on the same day are cross-sectionally correlated and are not independent trials. Calibration statistics are treated as preliminary until at least 30 issue days have been evaluated.
5Limitations
The estimation window retains memory of past volatility, so estimates may be conservative in calm regimes. Results apply to the stated universe and horizons only and carry no implication of investment performance.
Method details: Methodology · References: Literature