H-031: Live: an improved daily model (coin-specific levels, recalibrated) gives each coin's chance of rising above fees better than its base rate

— Family: forecast · Status: registered

Does an improved model predict crypto direction live? Data is being collected. The pre-registered test runs on 2027-11-01, once a year of data has accumulated.

Forward data collected: 0 of 365 days (0.0%). First test on 2027-11-01.

Result

Sharpe: return per unit of risk (higher is better). CAGR: average yearly growth. Max drawdown: worst fall from a peak. p-value: chance the result is luck (lower is better). “Hold” columns are buy-and-hold for the same asset and days. Glossary

No engine results yet: this study is kept for the record and is not re-run.

Why it might work

Asked for by F1DevBin after H-027 failed, and designed after seeing H-027's history result, so it is judged only on forecasts published from 2026-11-01 on, data nobody has seen. H-027's probabilities were biased and overconfident: one shared intercept ignored that Gate.io coins need a larger move to beat their costs (Brier skill -0.10 on EVER, -0.03 on GRAM), and where it said 53 % only 46 % happened. Two changes, nothing else: (1) a separate intercept per coin, ridge-penalised like the other weights; (2) Platt recalibration: a model trained on data up to a year before the month's cutoff scores that last year, a one-variable logistic regression maps those scores to outcomes, and the map is applied to the model trained on all data. Same inputs, coins, costs, monthly refits (frozen in datasets/forecast_models/YYYY-MM_v2.json), paper portfolio and scoring as H-029, published in the same daily log before 12:00 UTC (columns ending in _v2) and scored noon to noon. Expectation written before any result: the bias disappears and the skill moves to about zero; beating the base rate clearly is unlikely, because H-027's forecasts carried only a trace of information. Alpha: H-029 is locked at 0.025, so this direction test takes half of it, 0.0125.

Verdict history

Pre-registered specification

kind: forecast_live
target: direction
model:
  type: pooled_logit
  features:
  - r1
  - r3
  - r7
  - r30
  - d50
  - d200
  - v7
  - v30
  - btc_r1
  - btc_r7
  - funding
  - fear_greed
  - stable30
  - flow7
  zscore_window: 365
  zscore_min_periods: 180
  clip: 5
  l2: 1.0
  refit: monthly
  external_lag: 1
  ffill_limit: 3
  coin_intercepts: true
  calibration: platt_365
log_suffix: _v2
universe:
- BTCUSDT
- ETHUSDT
- SOLUSDT
- BNBUSDT
- XRPUSDT
- ADAUSDT
- LTCUSDT
- LINKUSDT
- DOGEUSDT
- XLMUSDT
- TRXUSDT
- EVERUSDT
- GRAMUSDT
costs:
  default: 0.001
  EVERUSDT: 0.004
  GRAMUSDT: 0.004
portfolio:
  weight: probability
  band: 0.1
execution:
  time_utc: '12:00'
periods:
- name: forward
  start: '2026-11-01'
significance:
  method: circular_shift
  n: 2000
  alpha: 0.0125
criterion:
  metric: brier_skill
  pooled: true