{
  "_comment": "Empirical band-widening factors applied at chart render time. Archived round artifacts under forecasts/rounds/ are never modified; the correction exists only in the published fan charts.",
  "source": {
    "repo": "PolicyEngine/boe-var-model",
    "file": "results/coverage_evaluation.json",
    "revision": "c13a9c912e106ad6de02809f99897d8ef79ea69d",
    "code_version": "47e8ab54d92cee11bcfeed15a1e9e0d584b85d9b088fe2b9dc324643eff7bc92",
    "origins": 50,
    "draws_per_origin": 1000,
    "horizons_evaluated": 8,
    "block": "band_scale_factors_yoy.production_spec",
    "spec": "production_spec: the coverage evaluation re-run with the six Covid dummies (2020Q1-2021Q2) that every published forecast is estimated with, scored on four-quarter differences (coverage_yoy)"
  },
  "derivation": "Gaussian half-width approximation: k(h, level) = z(nominal) / z(empirical), where z(p) is the standard-normal quantile at (1+p)/2 and empirical is the measured coverage of the nominal band at horizon h. Each band half-width is multiplied by k around the archived median. k below 1 (an over-covering band) narrows it: this is a calibration, not a one-way widening.",
  "extrapolation": "The evaluation covers horizons h1-h8; archived rounds run to 13 quarters. For h > 8 the h8 factor is held constant.",
  "series_mapping": {
    "gdp": "uk_gdp",
    "cpi": "cpisa"
  },
  "caveat": "Coverage is measured on four-quarter differences of the model's log series (pseudo-out-of-sample, overlapping origins), the same year-on-year basis as the published fans. The shock factor is the Cholesky factor (boe-var-model #26/#30), so the measured coverage is the same on every machine. If Bank Rate is ever fanned, use the level factors (band_scale_factors.production_spec), since it is published as a level.",
  "unemployment_satellite": "Left raw. The svar-unemployment (Okun) satellite is not among the series in the coverage evaluation, so no measured factor applies to it; borrowing the boe-svar factors would assert a calibration that was never tested. Its fan is labelled as raw model quantiles.",
  "k": {
    "gdp": {
      "68": {
        "h1": 0.8464,
        "h2": 0.9205,
        "h3": 1.1328,
        "h4": 1.0864,
        "h5": 1.0864,
        "h6": 1.1328,
        "h7": 1.1816,
        "h8": 1.1328
      },
      "90": {
        "h1": 1.0579,
        "h2": 1.1146,
        "h3": 1.2268,
        "h4": 1.3411,
        "h5": 1.3999,
        "h6": 1.3411,
        "h7": 1.3999,
        "h8": 1.4603
      }
    },
    "cpi": {
      "68": {
        "h1": 1.1816,
        "h2": 1.346,
        "h3": 1.408,
        "h4": 1.7062,
        "h5": 1.797,
        "h6": 1.797,
        "h7": 1.797,
        "h8": 1.6228
      },
      "90": {
        "h1": 1.3411,
        "h2": 1.587,
        "h3": 1.7239,
        "h4": 1.9544,
        "h5": 2.0397,
        "h6": 2.0397,
        "h7": 2.0397,
        "h8": 2.2262
      }
    }
  },
  "correction": "Superseded factors measured without Covid dummies. The coverage run had always estimated with no dummies while every published fan estimates with them, so the factors were derived from a different model than the one they were applied to — and in the same direction at every horizon for UK GDP, which was being narrowed by 36% at h1 where the model actually shown wants 14%. Both sets are published in the source artifact.",
  "standard_error": "Binomial standard error sqrt(p(1-p)/n) at the NOMINAL level with n = 50 origins: 0.066 for the 68% band and 0.0424 for the 90% band. Origins overlap, so the number of independent observations is well below n and this understates the true sampling error. A one-standard-error move in measured coverage shifts k by far more than the fourth decimal these factors are quoted to; the digits are the arithmetic of the run, not the precision of the estimate."
}
