{
  "as_of": "2026-10-19",
  "question": "Do removing factor families and choosing smaller fundamentals scores improve Senate forecasts?",
  "groups": {
    "lean": [
      "pvi_scaled"
    ],
    "incumbency": [
      "incumbency_advantage",
      "appointed_advantage",
      "multiple_elections_advantage"
    ],
    "prior_performance": [
      "incumbent_prior_scaled",
      "incumbent_prior_missing",
      "nonincumbent_prior_scaled",
      "nonincumbent_prior_missing"
    ],
    "experience": [
      "nonincumbent_experience",
      "career_missing_advantage"
    ],
    "scandals": [
      "news_scaled"
    ],
    "generic_ballot": [
      "generic_scaled"
    ],
    "fundraising": [
      "individual_share_difference"
    ]
  },
  "cohorts": {
    "core": [
      "lean",
      "incumbency",
      "prior_performance",
      "experience",
      "scandals"
    ],
    "generic": [
      "lean",
      "incumbency",
      "prior_performance",
      "experience",
      "scandals",
      "generic_ballot"
    ],
    "financial": [
      "lean",
      "incumbency",
      "prior_performance",
      "experience",
      "scandals",
      "generic_ballot",
      "fundraising"
    ]
  },
  "subset_counts": {
    "core": 32,
    "generic": 64,
    "financial": 128
  },
  "primary_test_cycles": [
    2016,
    2018,
    2020,
    2022,
    2024
  ],
  "fixed_test": "Every subset uses lambda 1 and a fixed 25% blend, with the existing +/-3-point cap. This is a retrospective diagnostic, not an adopted winner.",
  "removal": "Remove one whole family and its associated source/performance missing indicators. Other correlated features can retain related information; these are predictive ablations, not isolated causal effects.",
  "selection": "Primary algorithm selects the subset, lambda in [.25,1,4], and blend in [0,.10,.25,.50] on earlier chronological validation elections only, with at least two such elections. It then refits the chosen pure outcome index on earlier elections and scores the new election. No current race polls enter the index; polling is used only in the separate blend.",
  "eligibility": "All subsets within a cohort train and test on exactly the same race membership, even if fundraising or generic ballot is removed. Core uses 218 historical outcome rows; generic requires actual generic readings; financial also requires matched individual-dollar reports. Missing reports never become zero.",
  "regularization": "Retain the earlier fundamentals protocol, priors, scales, chronological fits and inherited per-cycle uncertainty. The empty subset is an intercept-only index and can choose zero blending.",
  "inference": "Use whole-election paired MAE differences for the selected algorithm and diagnostic winners. Holm adjust fixed-subset p-values within each cohort. Retrospective winners remain exploratory even after correction; overlapping training, development-cycle reuse and few national cycles limit inference.",
  "production_changed": false
}