Skip to content

m6_forward_tape_2weekends_variants.md

Latest commit

 

History

History
64 lines (52 loc) · 3.46 KB

m6_forward_tape_2weekends_variants.md

File metadata and controls

64 lines (52 loc) · 3.46 KB
Raw

M6 σ̂ variant comparison — forward tape, 2 weekends since freeze

Generated: 2026-05-12 13:30 UTC. Variant bundle: lwc_variant_bundle_v1_frozen_20260504.json (SHA-256 7cef6132d970…, freeze date 2026-05-04). Forward window: 2026-05-01 → 2026-05-08.

Role of this report. §13.6 of reports/m6_sigma_ewma.md describes the selection-procedure transparency layer — the canonical M6 σ̂ rule (EWMA HL=8) was selected from a 5-variant ladder under a multi-test-exposed criterion (80 split-date Christoffersen cells). To re-validate the selection on data it never saw, this report scores all five variants on the same forward weekends. The intent is re-validation, not re-selection: a different variant looking cleaner here is a flag to revisit, not to re-deploy.

Preliminary — fewer than 4 forward weekends accumulated; treat the comparison as anecdotal until ≥ 4 weekends land. Per-variant Christoffersen and Kupiec are uninformative at this n.

1. Pooled OOS — all variants at every served τ

variant τ n realised half-width (bps) Kupiec p Christoffersen p
baseline_k26 0.68 20 0.9000 109.4 0.0197 nan
baseline_k26 0.85 20 1.0000 173.6 0.0108 nan
baseline_k26 0.95 20 1.0000 291.4 0.1520 nan
baseline_k26 0.99 20 1.0000 495.0 0.5261 nan
ewma_hl6 0.68 20 0.9000 107.3 0.0197 nan
ewma_hl6 0.85 20 0.9500 170.0 0.1543 nan
ewma_hl6 0.95 20 1.0000 291.2 0.1520 nan
ewma_hl6 0.99 20 1.0000 479.9 0.5261 nan
ewma_hl8 (canonical) 0.68 20 0.9000 106.3 0.0197 nan
ewma_hl8 (canonical) 0.85 20 0.9500 166.9 0.1543 nan
ewma_hl8 (canonical) 0.95 20 1.0000 290.7 0.1520 nan
ewma_hl8 (canonical) 0.99 20 1.0000 456.9 0.5261 nan
ewma_hl12 0.68 20 0.9000 103.7 0.0197 nan
ewma_hl12 0.85 20 0.9500 160.6 0.1543 nan
ewma_hl12 0.95 20 1.0000 269.2 0.1520 nan
ewma_hl12 0.99 20 1.0000 444.1 0.5261 nan
blend_a50_hl8 0.68 20 0.9000 107.3 0.0197 nan
blend_a50_hl8 0.85 20 0.9500 169.6 0.1543 nan
blend_a50_hl8 0.95 20 1.0000 291.9 0.1520 nan
blend_a50_hl8 0.99 20 1.0000 457.6 0.5261 nan

2. Headline comparison — variant × τ pooled half-width (bps)

variant τ=0.68 τ=0.85 τ=0.95 τ=0.99
baseline_k26 109.4 173.6 291.4 495.0
ewma_hl6 107.3 170.0 291.2 479.9
ewma_hl8 (canonical) 106.3 166.9 290.7 456.9
ewma_hl12 103.7 160.6 269.2 444.1
blend_a50_hl8 107.3 169.6 291.9 457.6

3. Headline comparison — realised coverage

variant τ=0.68 τ=0.85 τ=0.95 τ=0.99
baseline_k26 0.9000 1.0000 1.0000 1.0000
ewma_hl6 0.9000 0.9500 1.0000 1.0000
ewma_hl8 (canonical) 0.9000 0.9500 1.0000 1.0000
ewma_hl12 0.9000 0.9500 1.0000 1.0000
blend_a50_hl8 0.9000 0.9500 1.0000 1.0000

4. Reproducibility

uv run python scripts/freeze_sigma_ewma_variant_bundle.py
uv run python scripts/collect_forward_tape.py
uv run python scripts/run_forward_tape_variant_comparison.py

The variant bundle is read-only. To advance the freeze date, re-run scripts/freeze_sigma_ewma_variant_bundle.py with a new --date and re-run this evaluator.