Research specification
Methodology
Daily close-to-close estimates. Explicit assumptions. Persisted evidence.
Study status
Study in progress · — issue days · Updated daily
The ledger has run for a short period. All calibration, Brier and coverage figures are preliminary: predictions issued on the same day share the same market moves, so the effective sample size is the number of issue days, not the number of rows.
Low volatility and conservatism: the study so far has been observed almost entirely in a below-median volatility regime. In that regime the model's 365-day window retains memory of past volatility, so its estimates are more conservative than the 180-day naive base rate, and observed frequencies can sit below the stated probabilities. This is expected and is not evidence of a calibration error or a reason to recalibrate.
No recalibration is applied. Model parameters stay fixed until calibration can be evaluated across multiple regimes.
What this distribution is chosen for
- Asymmetric tailsUpside and downside returns are fitted separately, each with its own shape and scale.
- Direct tail probabilityP(close beyond x) is computed from the fitted distribution, not from a normal approximation.
- Regime diagnosticCalibration is split by 30-day realized volatility, so a calm period is not mistaken for a general result.
The distribution choice is supported by published research (see Literature). Whether it delivers on the live ledger is a separate question, answered only by the ledger itself.
Two-sided Weibull estimation
Daily Binance prices provide a rolling 365-day return window for 1, 3 and 7-day horizons. Positive and negative returns are fitted separately, allowing different tail shapes and scales.
P(move > x) = p · exp(−(x / λ)k)
These are model-implied probabilities, not known future frequencies. Regime changes, limited tail observations and overlapping horizons can affect estimates.
Price corridors · C+cap
The target corridor probability is 85% before the width constraint. When total target width exceeds 2 × W_max, one shared scale factor reduces both sides, preserving their asymmetry. Final implied coverage is recomputed as one minus both weighted tail probabilities.
W_max(h) = max(1.5 × σ_90d × √h, 3 × round-trip cost)
Final bounds, implied probability, cap status and policy audit fields are persisted with each new record. Existing bounds are never refitted in the ledger. The reference stop represents tail risk at the daily close, not an execution price or intraday guarantee.
Event selection & baseline comparison
The daily candidate set comprises 10 assets × 3 horizons × 6 thresholds × 4 event types. Every candidate is recorded; selection only controls the overview, never the ledger.
- Selected event probabilities range from 65% to 90%.
- Minimum divergence vs baseline (in either direction): |model − base rate| ≥ 3 pp, with positive historical Brier skill and a minimum backtest size. A negative divergence means the model is more cautious than recent history. Divergence is not a quality score; calibration is.
- In calm regimes, the model's 365-day window retains memory of past volatility, producing estimates more conservative than the 180-day naive base rate. This is by design. Calibration should be evaluated across regimes, not within a single calm period.
- Confidence in estimate: share of historical backtest windows in which the model's estimate was well calibrated. Fewer than 30 backtest windows is flagged as a small sample.
- 90% vs 85%: the base policy corridor is calibrated to 90% coverage in the walk-forward historical test; C+cap corridors target 85% before the width cap, and each record stores its final implied probability.
- Legacy: records created before canonical payload storage. Their hash can only be checked by reconstruction and may be inconclusive.
- At most 5 selected events and 2 per asset.
Price-corridor selection is separate: implied probability from 65% to 90%, subject to width constraints. Low-probability corridors remain as observation-only records.
Calibration scope & evidence integrity
Walk-forward width calibration uses past data only. The historical out-of-sample test targets 90% under the base policy; it is not a C+cap backtest. Recorded coverage can include multiple policies and should not be interpreted as a single-policy accuracy estimate.
Original predictions and price levels are append-only. Complete outcomes may be filled once after the horizon closes. Browser SHA-256 verification compares original payload bytes with the stored digest; legacy serialization may be inconclusive.
Daily checkpoint manifests start at activation, with no independent external time anchor. Payload integrity does not establish forecasting accuracy. Correlated rows are not independent samples.