forecasting/bootstrap/README.md

Forecasting bootstrap reference

This subdir holds the bootstrap reference for the Track-3 (forecasting) leaderboard recompute — the long-format per-draw frame that the skill / rank / fairness CIs reduce from.

Layout

forecasting/bootstrap/
├── draws.parquet      # per-(reduction, model, scope, metric, draw) value
└── draws.meta.json    # provenance: seed, n_boot, methods, baseline, git commit

What's it for

Each row of draws.parquet is one bootstrap draw of one headline scope for one method, for one of three reductions (reduction column):

• skill — paired skill score S = 1 − exp(mean_task log R) vs seasonal_naive, per (model, scope) (scope = channel_<i>_score, sleep_score, workout_score, activity_score, physiology_score, overall_score).
• rank — cross-method average rank, per (model, scope, metric).
• fairness — mean-absolute-pairwise-difference (MAPD) disparity-ratio fair skill score, per (model, scope) (scope = age_group, sex, overall, the 4 sensor categories, and per-channel rows).

Reducing each draw group to mean / SE / percentile-CI (and BCa for the headline scopes) reproduces the published forecasting_*_bootstrap.csv tables.

Note — scope-level, not task-level. Unlike the imputation track (whose draws.parquet is per-task E/R/rank, reducible to any scope), forecasting aggregates tasks → scopes within each draw, so its draws are already at the headline-scope level. They reproduce the headline CIs but cannot be re-aggregated to other scopes. (Re-aggregation is unnecessary for the leaderboard, which shows exactly these scopes.)

Provenance

Generated by scripts/paper_results/forecasting/produce_forecasting_bootstrap_draws.py in the code repo. The current snapshot was built with:

• seed = 42, n_boot = 1000, ci_level = 0.95
• split = test (canonical sharable_users_seed42_2026)
• 10 methods (see draws.meta.json:methods)
• Baseline: seasonal_naive; scored metrics: mae (continuous), auroc (binary)
• Within-user aggregation: micro; age_bins = [18, 30, 40, 50, 60]
• Fairness disparity: MAPD (mean absolute pairwise difference)

See draws.meta.json for the exact git commit and runtime metadata.

Note: no pooled per_user_errors.parquet

The per-user substrate (the BCa LOO + point input) is not stored here — it is exactly the concatenation of the per-method substrate files one level up:

import glob, pandas as pd
pooled = pd.concat(
    [pd.read_parquet(p) for p in glob.glob("forecasting/*.parquet")],
    ignore_index=True,
)

Loading

from huggingface_hub import hf_hub_download
import pandas as pd, json

draws = pd.read_parquet(hf_hub_download(
    "MyHeartCounts/OpenMHC-leaderboard-data",
    "forecasting/bootstrap/draws.parquet", repo_type="dataset"))
meta = json.loads(open(hf_hub_download(
    "MyHeartCounts/OpenMHC-leaderboard-data",
    "forecasting/bootstrap/draws.meta.json", repo_type="dataset")).read())
print(meta["seed"], meta["n_boot"], len(meta["methods"]))

See SCHEMA.md for the full column spec.

Uploaded with

tools/upload_leaderboard_bootstrap.py --track forecasting in the code repo.