feat(analysis): N time-series stability comparison across A-matrix methods by MoLi7 · Pull Request #381 · cornerstone-data/bedrock

MoLi7 · 2026-05-06T05:12:08Z

cc:
Closes:

What changed? Why?

Adds bedrock/analysis/a_matrix_time_series/compare_method_stability.py, a new analysis script that ranks the 3 A-matrix methods (commodity-PI, industry-PI, summary-tables) by year-over-year stability of N across 2019–2023. useeio is excluded (no temporal scaling on A — not comparable in this framing).

The script reads the per-pair tabs in ef_comparison.xlsx, pivots to a long panel of N_new_ref (deflated to 2023$), and emits:

output/results/n_yoy_ranking.csv — per-approach mean_abs_yoy_pct, max_abs_yoy_pct, abs_total_drift_pct, each rolled up as median, p95, and emissions-weighted (by |mean_N|).
output/results/n_yoy_per_sector.csv — per-sector per-year N, the 4 transition YoY %s, and the aggregates.
output/plots/n_indexed_lines.png — head sectors covering 30% of |mean_N| (cap 8), N rebased to 2019=100, faceted by method.
output/plots/n_yoy_distribution.png — per-method boxplot of mean |YoY %| + per-transition |YoY %| boxplots grouped by method.

Also fixes a year-coercion bug in compile_ef_diagnostics.py: the year column from ef_run_index.csv reads as float-string ("2019.0"), so int(year) raised; switched to int(float(year)).

Testing

Ran compile + the new script end-to-end against the 20 dispatched bundle_v0_2 cells; all 4 artifacts written, ranking matches expectations (industry-PI ≈ commodity-PI ≪ summary-tables on weighted YoY).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

MoLi7 · 2026-05-06T05:12:22Z

This stack of pull requests is managed by Graphite. Learn more about stacking.

Adds rebuild_run_index_from_drive.py — lists Google Sheets in a diagnostics Drive folder and parses each title into (approach, baseline, sheet_id, year?, scenario?) rows for ef_run_index.csv. Closes the manual- flow gap: users who triggered diagnostics via the GH Actions UI can now auto-build the run index instead of hand-typing sheet IDs. Title regex handles both formats currently in use: Manual: [DATE, BASELINE based, A matrix with APPROACH] EFs diagnostics Dispatch: [DATE, YEAR, BASELINE based, A matrix with APPROACH, SCENARIO] EFs diagnostics Also rewords the FileNotFoundError in compile_ef_diagnostics.py to point at the new script as the auto-rebuild path, with hand-write as a fallback. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…patcher Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…-series dispatch Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…race Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…ta-overwrite Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Each diagnostics cell runs at model_base_year=Y so D_new/N_new are denominated in year-Y dollars. Cross-year comparison requires a single dollar reference; add D_new_ref/N_new_ref columns that apply inflation_adjust_ef_denom_to_new_base_year to land every cell on REFERENCE_DOLLAR_YEAR (2023). Step 6 single-year rows (empty year column) skip the step and behave unchanged. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Dispatcher previously crashed mid-run when `gh run list` returned non-zero exit (transient API hiccup mid-batch). Now retries 3× per status with a 5s backoff and falls back to a "still busy" sentinel so the poll loop keeps spinning instead of unwinding the dispatch and losing the queue. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…thods

mypy flags ta.cast as redundant since pd.to_numeric/concat/__getitem__ return inferable types. The casts were added to placate Pyright, which isn't a CI check. Switching back to no-cast satisfies mypy + black; ruff unchanged. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

- Removed 6 redundant casts that mypy flagged. The casts placated Pyright but mypy infers the same types without them. - pd.ExcelFile.sheet_names is typed as list[int|str]; explicitly str() the tab name before passing to _parse_tab and pd.read_excel. - Apply black formatting. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

WesIngwersen · 2026-05-07T13:48:41Z

Useful line charts and indexing but the n values need to be converted to a common dollar year to see stability

MoLi7 · 2026-05-07T15:55:18Z

-                joined, source_year=int(year), ref_year=REFERENCE_DOLLAR_YEAR
+                joined,
+                source_year=int(float(year)),
+                ref_year=REFERENCE_DOLLAR_YEAR,


@WesIngwersen this step here deflates the new N values in each year-based diagnostics from the varying model_base_year to a shared REFERENCE_DOLLAR_YEAR, so that the N comparison in a time series is comparing N values based on the same dollar year.