PHASE 8 — Paper (2026-11-03)
Manuscript
paper/main.tex → paper/main.pdf (latexmk, pdflatex + bibtex, style aer, UQO template: newtxtext/newtxmath,
title page macros). Sections: title page + abstract, introduction, hypotheses (5 formal statements), related work,
data, model, experimental design, results (pilot, development test 2024, calibration, subgroups, economics, zero-shot
and untouched 2025), ablations and falsification (incl. Stage-B paths), limitations, reproducibility and release,
conclusion, bibliography, appendices (baseline specifications, additional tables, universe). 36 pages.
Provenance of every number
paper/results_macros.tex: 3 468 macros generated byscripts/evaluate.pyfrom the saved leaderboards, comparisons and pairwise files (\crps<exp><model>H<h>,\rel…,\pw<exp><model>vs<ref>H<h>,\tpw…,\ppw…, coverage macros); digits in macro names are spelt out (Hsixzero= H = 60).paper/tables/*.tex(53 files): leaderboards (scaled CRPS, WIS, QLIKE, calibration) and relative-change tables for pilot / main / main held-out / ablations / final / final held-out, universe strata, groups and ticker lists, baseline specifications (fromconfigs/baselines.yaml), measured benchmark (fromresults/metrics/benchmark_M3U96b.json).paper/figures/*.pdf(27): scores by horizon, coverage, PIT histograms, forest plots, training curves, drawn only from result files bysrc/chronos_x/eval/figures.py(Times/STIX, monochrome + one accent, panel labels).- Hand-typed numbers in the prose (subgroups, ablation ranges, copula scores, economic results) are copied from
results/metrics/*/pairwise_*.csv,subgroups_*.csv,stress_calibration_*.csv,path_scores_*.json,economic_test.csvand listed indocs/PHASE_5..7_REPORT.md. - Bibliography: 54 entries, every DOI/arXiv id fetched and title-matched (
scripts/build_references.py,results/metrics/references_verification.json); four entries without DOI are typed by hand (LightGBM, Holm, Sklar, López de Prado).
Claims → evidence (central ones)
| Claim | Evidence |
|---|---|
| All models beat the random walk by 3.5–4.9 % | main/compare_crps_vs_naive_rw_test.csv, final/…_final.csv |
| H1 effect ≤ 0.1 % in sample, absent in 2025 | main/pairwise_test.csv, final/pairwise_final.csv, ablations/pairwise_test.csv |
| HAR-RV better from 15 min, neural better at 5 min | same files, tables/main_crps_scaled.tex, final_crps_scaled.tex |
| Transfer without penalty | main/pairwise_test_heldout.csv, final/pairwise_final_heldout.csv |
| H3, H4, H5 nulls | ablations/pairwise_test.csv, stress_calibration_test.csv, path_scores_mrpt_small_copula_test.json, main/economic_test.csv |
| Time-only ≈ 90 % of the skill | ablations/pairwise_test.csv (mrpt_small_time_only vs mrpt_small_b4k; b4k vs naive_rw) |
Checks
latexmkexit 0; 0 undefined references/citations after the final pass; 1 remaining overfull box (appendix table, cosmetic).- Visual check of the results pages (figures and tables render; legends below panels).
- No proprietary bars, grids or per-origin returns appear in the manuscript, tables or figures (scores and tickers only).