SPB Git

Repository Public Research preview

Commit

L1/L4 post-pull run 5 (P4 learning curve 12k/24k/36k + budget verdict, ablations_h3 complete): budget tables, learning-curve and H3-control records in inline_numbers.json, compute ledger rows (WP41-EDIT-08, WP41-AI_C-02, WP41-THEO-09, WP41-THEO-05, WP41-AI_C-04)

budget_verdict.json (P4): criterion (i) fails at H = 5 and 15 (the 36k H1 gap +0.05 % / -0.01 % falls outside the five-seed band of the
12k contrast) and holds at H = 30 and 60 (-0.09 / -0.25 %); criterion (ii) holds (MRPT-S vs HAR-RV keeps its sign at 30/60 min, +0.49 /
+0.62 % at 36k vs +0.53 / +0.65 % at 12k) -> the budget condition stays in the abstract. Validation-NLL slope per 1,000 steps at 36k:
-0.00001 (MRPT-S, converged) vs +0.00006 (1m-only). latexmk 42 p., 0 errors, 0 overfull.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Simon-Pierre Boucher committed · Nov 6, 2026 parent bc13620
9 changed files +643 −1 Browse files
modified paper/main.pdf +0 −0

Binary file not shown.

modified paper/results_macros.tex +80 −0
@@ -11674,4 +11674,84 @@
11674 11674 \newcommand{\pwunseedsmrptsmallsfivevsmrptsmallonemonlysfiveHsixzero}{-0.14}
11675 11675 \newcommand{\tpwunseedsmrptsmallsfivevsmrptsmallonemonlysfiveHsixzero}{-4.8}
11676 11676 \newcommand{\ppwunseedsmrptsmallsfivevsmrptsmallonemonlysfiveHsixzero}{1.58e-06}
11677 +\newcommand{\crpsbudgetmrptsmallonemonlyctwofourkHfive}{0.4532}
11678 +\newcommand{\crpsrawbudgetmrptsmallonemonlyctwofourkHfive}{7.936e-04}
11679 +\newcommand{\covfiftybudgetmrptsmallonemonlyctwofourkHfive}{55}
11680 +\newcommand{\covninetyfivebudgetmrptsmallonemonlyctwofourkHfive}{95.7}
11681 +\newcommand{\diraccbudgetmrptsmallonemonlyctwofourkHfive}{0.501}
11682 +\newcommand{\crpsbudgetmrptsmallonemonlyctwofourkHonefive}{0.4418}
11683 +\newcommand{\crpsrawbudgetmrptsmallonemonlyctwofourkHonefive}{1.338e-03}
11684 +\newcommand{\covfiftybudgetmrptsmallonemonlyctwofourkHonefive}{57}
11685 +\newcommand{\covninetyfivebudgetmrptsmallonemonlyctwofourkHonefive}{96.0}
11686 +\newcommand{\diraccbudgetmrptsmallonemonlyctwofourkHonefive}{0.500}
11687 +\newcommand{\crpsbudgetmrptsmallonemonlyctwofourkHthreezero}{0.4319}
11688 +\newcommand{\crpsrawbudgetmrptsmallonemonlyctwofourkHthreezero}{1.847e-03}
11689 +\newcommand{\covfiftybudgetmrptsmallonemonlyctwofourkHthreezero}{57}
11690 +\newcommand{\covninetyfivebudgetmrptsmallonemonlyctwofourkHthreezero}{96.2}
11691 +\newcommand{\diraccbudgetmrptsmallonemonlyctwofourkHthreezero}{0.500}
11692 +\newcommand{\crpsbudgetmrptsmallonemonlyctwofourkHsixzero}{0.4199}
11693 +\newcommand{\crpsrawbudgetmrptsmallonemonlyctwofourkHsixzero}{2.533e-03}
11694 +\newcommand{\covfiftybudgetmrptsmallonemonlyctwofourkHsixzero}{58}
11695 +\newcommand{\covninetyfivebudgetmrptsmallonemonlyctwofourkHsixzero}{96.3}
11696 +\newcommand{\diraccbudgetmrptsmallonemonlyctwofourkHsixzero}{0.499}
11697 +\newcommand{\crpsbudgetmrptsmallonemonlycthreesixkHfive}{0.4526}
11698 +\newcommand{\crpsrawbudgetmrptsmallonemonlycthreesixkHfive}{7.923e-04}
11699 +\newcommand{\covfiftybudgetmrptsmallonemonlycthreesixkHfive}{54}
11700 +\newcommand{\covninetyfivebudgetmrptsmallonemonlycthreesixkHfive}{95.2}
11701 +\newcommand{\diraccbudgetmrptsmallonemonlycthreesixkHfive}{0.500}
11702 +\newcommand{\crpsbudgetmrptsmallonemonlycthreesixkHonefive}{0.4411}
11703 +\newcommand{\crpsrawbudgetmrptsmallonemonlycthreesixkHonefive}{1.335e-03}
11704 +\newcommand{\covfiftybudgetmrptsmallonemonlycthreesixkHonefive}{56}
11705 +\newcommand{\covninetyfivebudgetmrptsmallonemonlycthreesixkHonefive}{95.6}
11706 +\newcommand{\diraccbudgetmrptsmallonemonlycthreesixkHonefive}{0.501}
11707 +\newcommand{\crpsbudgetmrptsmallonemonlycthreesixkHthreezero}{0.4313}
11708 +\newcommand{\crpsrawbudgetmrptsmallonemonlycthreesixkHthreezero}{1.843e-03}
11709 +\newcommand{\covfiftybudgetmrptsmallonemonlycthreesixkHthreezero}{57}
11710 +\newcommand{\covninetyfivebudgetmrptsmallonemonlycthreesixkHthreezero}{95.8}
11711 +\newcommand{\diraccbudgetmrptsmallonemonlycthreesixkHthreezero}{0.501}
11712 +\newcommand{\crpsbudgetmrptsmallonemonlycthreesixkHsixzero}{0.4195}
11713 +\newcommand{\crpsrawbudgetmrptsmallonemonlycthreesixkHsixzero}{2.529e-03}
11714 +\newcommand{\covfiftybudgetmrptsmallonemonlycthreesixkHsixzero}{58}
11715 +\newcommand{\covninetyfivebudgetmrptsmallonemonlycthreesixkHsixzero}{96.1}
11716 +\newcommand{\diraccbudgetmrptsmallonemonlycthreesixkHsixzero}{0.502}
11717 +\newcommand{\crpsbudgetmrptsmallctwofourkHfive}{0.4529}
11718 +\newcommand{\crpsrawbudgetmrptsmallctwofourkHfive}{7.929e-04}
11719 +\newcommand{\covfiftybudgetmrptsmallctwofourkHfive}{55}
11720 +\newcommand{\covninetyfivebudgetmrptsmallctwofourkHfive}{95.6}
11721 +\newcommand{\diraccbudgetmrptsmallctwofourkHfive}{0.501}
11722 +\newcommand{\crpsbudgetmrptsmallctwofourkHonefive}{0.4413}
11723 +\newcommand{\crpsrawbudgetmrptsmallctwofourkHonefive}{1.336e-03}
11724 +\newcommand{\covfiftybudgetmrptsmallctwofourkHonefive}{56}
11725 +\newcommand{\covninetyfivebudgetmrptsmallctwofourkHonefive}{95.9}
11726 +\newcommand{\diraccbudgetmrptsmallctwofourkHonefive}{0.500}
11727 +\newcommand{\crpsbudgetmrptsmallctwofourkHthreezero}{0.4312}
11728 +\newcommand{\crpsrawbudgetmrptsmallctwofourkHthreezero}{1.843e-03}
11729 +\newcommand{\covfiftybudgetmrptsmallctwofourkHthreezero}{57}
11730 +\newcommand{\covninetyfivebudgetmrptsmallctwofourkHthreezero}{96.0}
11731 +\newcommand{\diraccbudgetmrptsmallctwofourkHthreezero}{0.501}
11732 +\newcommand{\crpsbudgetmrptsmallctwofourkHsixzero}{0.4188}
11733 +\newcommand{\crpsrawbudgetmrptsmallctwofourkHsixzero}{2.525e-03}
11734 +\newcommand{\covfiftybudgetmrptsmallctwofourkHsixzero}{58}
11735 +\newcommand{\covninetyfivebudgetmrptsmallctwofourkHsixzero}{96.1}
11736 +\newcommand{\diraccbudgetmrptsmallctwofourkHsixzero}{0.502}
11737 +\newcommand{\crpsbudgetmrptsmallcthreesixkHfive}{0.4528}
11738 +\newcommand{\crpsrawbudgetmrptsmallcthreesixkHfive}{7.926e-04}
11739 +\newcommand{\covfiftybudgetmrptsmallcthreesixkHfive}{54}
11740 +\newcommand{\covninetyfivebudgetmrptsmallcthreesixkHfive}{95.5}
11741 +\newcommand{\diraccbudgetmrptsmallcthreesixkHfive}{0.500}
11742 +\newcommand{\crpsbudgetmrptsmallcthreesixkHonefive}{0.4411}
11743 +\newcommand{\crpsrawbudgetmrptsmallcthreesixkHonefive}{1.335e-03}
11744 +\newcommand{\covfiftybudgetmrptsmallcthreesixkHonefive}{56}
11745 +\newcommand{\covninetyfivebudgetmrptsmallcthreesixkHonefive}{95.8}
11746 +\newcommand{\diraccbudgetmrptsmallcthreesixkHonefive}{0.501}
11747 +\newcommand{\crpsbudgetmrptsmallcthreesixkHthreezero}{0.4310}
11748 +\newcommand{\crpsrawbudgetmrptsmallcthreesixkHthreezero}{1.842e-03}
11749 +\newcommand{\covfiftybudgetmrptsmallcthreesixkHthreezero}{57}
11750 +\newcommand{\covninetyfivebudgetmrptsmallcthreesixkHthreezero}{96.0}
11751 +\newcommand{\diraccbudgetmrptsmallcthreesixkHthreezero}{0.502}
11752 +\newcommand{\crpsbudgetmrptsmallcthreesixkHsixzero}{0.4185}
11753 +\newcommand{\crpsrawbudgetmrptsmallcthreesixkHsixzero}{2.523e-03}
11754 +\newcommand{\covfiftybudgetmrptsmallcthreesixkHsixzero}{57}
11755 +\newcommand{\covninetyfivebudgetmrptsmallcthreesixkHsixzero}{96.1}
11756 +\newcommand{\diraccbudgetmrptsmallcthreesixkHsixzero}{0.503}
11677 11757 \newcommand{\pairwiseConvention}{scaled CRPS}
added paper/tables/budget_calibration.tex +18 −0
@@ -0,0 +1,18 @@
1 +\begin{table}[htbp]\centering\small
2 +\caption{Learning curve (continued training): empirical 50/80/95\% coverage and PIT KS statistic}
3 +\label{tab:budget_calib}
4 +\begin{adjustbox}{max width=\textwidth}
5 +\begin{tabular}{lcccccccccccccccc}
6 +\toprule
7 +Model & \multicolumn{4}{c}{$H=5$} & \multicolumn{4}{c}{$H=15$} & \multicolumn{4}{c}{$H=30$} & \multicolumn{4}{c}{$H=60$} \\
8 + & 50 & 80 & 95 & KS & 50 & 80 & 95 & KS & 50 & 80 & 95 & KS & 50 & 80 & 95 & KS \\
9 +\midrule
10 +MRPT-S continued to 24k steps & 0.55 & 0.83 & 0.96 & 0.029 & 0.56 & 0.84 & 0.96 & 0.033 & 0.57 & 0.85 & 0.96 & 0.037 & 0.58 & 0.85 & 0.96 & 0.040 \\
11 +MRPT-S continued to 36k steps & 0.54 & 0.83 & 0.96 & 0.029 & 0.56 & 0.84 & 0.96 & 0.034 & 0.57 & 0.84 & 0.96 & 0.039 & 0.57 & 0.85 & 0.96 & 0.043 \\
12 +MRPT-S 1m only, 24k steps & 0.55 & 0.83 & 0.96 & 0.026 & 0.57 & 0.84 & 0.96 & 0.037 & 0.57 & 0.85 & 0.96 & 0.043 & 0.58 & 0.85 & 0.96 & 0.050 \\
13 +MRPT-S 1m only, 36k steps & 0.54 & 0.82 & 0.95 & 0.024 & 0.56 & 0.83 & 0.96 & 0.030 & 0.57 & 0.84 & 0.96 & 0.036 & 0.58 & 0.85 & 0.96 & 0.042 \\
14 +\bottomrule
15 +\end{tabular}
16 +\end{adjustbox}
17 +\begin{minipage}{0.95\textwidth}\scriptsize \emph{Notes.} development test 2024, \num{986468} evaluation origins over 251 sessions. Empirical coverage of the central 50/80/95\,\% prediction intervals of the cumulative return (nominal 0.50/0.80/0.95) and the Kolmogorov--Smirnov distance of the PIT values from uniformity; no scaling is involved.\end{minipage}
18 +\end{table}
\ No newline at end of file
added paper/tables/budget_crps_scaled.tex +17 −0
@@ -0,0 +1,17 @@
1 +\begin{table}[htbp]\centering\small
2 +\caption{Learning curve (continued training): mean scaled CRPS of the cumulative-return forecast by horizon (minutes)}
3 +\label{tab:budget_crps_scaled}
4 +\begin{adjustbox}{max width=\textwidth}
5 +\begin{tabular}{lrcccc}
6 +\toprule
7 +Model & Parameters & $H=5$ & $H=15$ & $H=30$ & $H=60$ \\
8 +\midrule
9 +MRPT-S continued to 24k steps & 3{,}455{,}619 & 0.4529 {\scriptsize[0.4474, 0.4588]} & 0.4413 {\scriptsize[0.4350, 0.4478]} & 0.4312 {\scriptsize[0.4249, 0.4381]} & 0.4188 {\scriptsize[0.4106, 0.4284]} \\
10 +MRPT-S continued to 36k steps & 3{,}455{,}619 & 0.4528 {\scriptsize[0.4473, 0.4587]} & 0.4411 {\scriptsize[0.4348, 0.4476]} & 0.4310 {\scriptsize[0.4247, 0.4380]} & 0.4185 {\scriptsize[0.4104, 0.4281]} \\
11 +MRPT-S 1m only, 24k steps & 3{,}200{,}259 & 0.4532 {\scriptsize[0.4478, 0.4591]} & 0.4418 {\scriptsize[0.4355, 0.4483]} & 0.4319 {\scriptsize[0.4257, 0.4387]} & 0.4199 {\scriptsize[0.4119, 0.4297]} \\
12 +MRPT-S 1m only, 36k steps & 3{,}200{,}259 & 0.4526 {\scriptsize[0.4471, 0.4585]} & 0.4411 {\scriptsize[0.4348, 0.4477]} & 0.4313 {\scriptsize[0.4250, 0.4381]} & 0.4195 {\scriptsize[0.4113, 0.4291]} \\
13 +\bottomrule
14 +\end{tabular}
15 +\end{adjustbox}
16 +\begin{minipage}{0.95\textwidth}\scriptsize \emph{Notes.} development test 2024, \num{986468} evaluation origins over 251 sessions. Brackets: 95\,\% trading-day cluster-bootstrap intervals (all instruments and origins of one session form one cluster). Convention: scaled CRPS, i.e.\ CRPS divided by $\sigma_t\sqrt{H}$. Parameters: trainable parameters of the trained models (\texttt{train\_summary.json}); baselines have none in this sense.\end{minipage}
17 +\end{table}
\ No newline at end of file
added paper/tables/budget_qlike.tex +17 −0
@@ -0,0 +1,17 @@
1 +\begin{table}[htbp]\centering\small
2 +\caption{Learning curve (continued training): mean QLIKE of the realised-variance forecast by horizon (minutes)}
3 +\label{tab:budget_qlike}
4 +\begin{adjustbox}{max width=\textwidth}
5 +\begin{tabular}{lcccc}
6 +\toprule
7 +Model & $H=5$ & $H=15$ & $H=30$ & $H=60$ \\
8 +\midrule
9 +MRPT-S continued to 24k steps & -12.475 {\scriptsize[-12.507, -12.439]} & -11.408 {\scriptsize[-11.443, -11.369]} & -10.737 {\scriptsize[-10.772, -10.697]} & -10.080 {\scriptsize[-10.117, -10.038]} \\
10 +MRPT-S continued to 36k steps & -12.476 {\scriptsize[-12.508, -12.439]} & -11.408 {\scriptsize[-11.443, -11.370]} & -10.736 {\scriptsize[-10.772, -10.697]} & -10.079 {\scriptsize[-10.117, -10.038]} \\
11 +MRPT-S 1m only, 24k steps & -12.469 {\scriptsize[-12.501, -12.431]} & -11.399 {\scriptsize[-11.437, -11.356]} & -10.727 {\scriptsize[-10.766, -10.682]} & -10.069 {\scriptsize[-10.108, -10.019]} \\
12 +MRPT-S 1m only, 36k steps & -12.469 {\scriptsize[-12.501, -12.432]} & -11.400 {\scriptsize[-11.438, -11.357]} & -10.729 {\scriptsize[-10.767, -10.685]} & -10.070 {\scriptsize[-10.108, -10.023]} \\
13 +\bottomrule
14 +\end{tabular}
15 +\end{adjustbox}
16 +\begin{minipage}{0.95\textwidth}\scriptsize \emph{Notes.} development test 2024, \num{986468} evaluation origins over 251 sessions. Brackets: 95\,\% trading-day cluster-bootstrap intervals (all instruments and origins of one session form one cluster). Convention: QLIKE of the realised variance, $\log v + \mathrm{RV}/v$ without the data-only constant (levels can be negative). The LightGBM $H=5$ QLIKE is dominated by a few near-zero variance forecasts (floored at $10^{-14}$ in \texttt{eval/scores.py}); the floor share is tabulated once \texttt{recal/lightgbm\_floor.csv} is in.\end{minipage}
17 +\end{table}
\ No newline at end of file
added paper/tables/budget_wis.tex +17 −0
@@ -0,0 +1,17 @@
1 +\begin{table}[htbp]\centering\small
2 +\caption{Learning curve (continued training): mean WIS ($\times 10^{4}$) of the cumulative-return forecast by horizon (minutes)}
3 +\label{tab:budget_wis}
4 +\begin{adjustbox}{max width=\textwidth}
5 +\begin{tabular}{lcccc}
6 +\toprule
7 +Model & $H=5$ & $H=15$ & $H=30$ & $H=60$ \\
8 +\midrule
9 +MRPT-S continued to 24k steps & 7.138 {\scriptsize[7.003, 7.279]} & 12.022 {\scriptsize[11.780, 12.281]} & 16.589 {\scriptsize[16.218, 16.970]} & 22.733 {\scriptsize[22.165, 23.355]} \\
10 +MRPT-S continued to 36k steps & 7.135 {\scriptsize[7.000, 7.276]} & 12.015 {\scriptsize[11.773, 12.275]} & 16.578 {\scriptsize[16.209, 16.958]} & 22.711 {\scriptsize[22.144, 23.333]} \\
11 +MRPT-S 1m only, 24k steps & 7.144 {\scriptsize[7.011, 7.286]} & 12.043 {\scriptsize[11.801, 12.303]} & 16.629 {\scriptsize[16.259, 17.006]} & 22.814 {\scriptsize[22.242, 23.431]} \\
12 +MRPT-S 1m only, 36k steps & 7.133 {\scriptsize[6.999, 7.274]} & 12.019 {\scriptsize[11.777, 12.281]} & 16.596 {\scriptsize[16.224, 16.976]} & 22.776 {\scriptsize[22.209, 23.397]} \\
13 +\bottomrule
14 +\end{tabular}
15 +\end{adjustbox}
16 +\begin{minipage}{0.95\textwidth}\scriptsize \emph{Notes.} development test 2024, \num{986468} evaluation origins over 251 sessions. Brackets: 95\,\% trading-day cluster-bootstrap intervals (all instruments and origins of one session form one cluster). Convention: weighted interval score of the cumulative return (unscaled). Values are multiplied by $10^{4}$ (raw WIS of order $10^{-3}$).\end{minipage}
17 +\end{table}
\ No newline at end of file
modified paper/tables/compute.tex +5 −1
@@ -36,11 +36,15 @@ seeds & MRPT-S seed 4 & MacStudnpierre4 & Apple M4 Max & 12{,}000 & 3{,}455{,}61
36 36 seeds & MRPT-S seed 5 & MacStudnpierre4 & Apple M4 Max & 12{,}000 & 3{,}455{,}619 & 2{,}024 & 379 \\
37 37 budget & MRPT-S 1m only, 24k steps & M5M36 & Apple M5 Max & 20{,}000 & 3{,}200{,}259 & 743 & 689 \\
38 38 budget & MRPT-S 1m only, 36k steps & M5M36 & Apple M5 Max & 30{,}000 & 3{,}200{,}259 & 912 & 702 \\
39 +budget & MRPT-S continued to 24k steps & MacStudnpierre2 & Apple M3 Ultra & 24{,}000 & 3{,}455{,}619 & 1{,}928 & 398 \\
40 +budget & MRPT-S continued to 36k steps & MacStudnpierre2 & Apple M3 Ultra & 34{,}000 & 3{,}455{,}619 & 1{,}564 & 409 \\
39 41 transfer300 & MRPT-S trained on all 300 instruments & m3u96b & Apple M3 Ultra & 12{,}000 & 3{,}455{,}619 & 1{,}902 & 404 \\
42 +ablations\_h3 & MRPT-S no regime, reduced global token & MacStudnpierre2 & Apple M3 Ultra & 4{,}000 & 3{,}416{,}643 & 640 & 400 \\
43 +ablations\_h3 & MRPT-S regime stats, reduced global token & MacStudnpierre2 & Apple M3 Ultra & 4{,}000 & 3{,}455{,}619 & 640 & 400 \\
40 44 ablations\_h3 & MRPT-S shuffled price/volume columns & M5M36 & Apple M5 Max & 4{,}000 & 3{,}455{,}619 & 933 & 274 \\
41 45 ablations\_h3 & MRPT-S time only, reduced global token & M5M36 & Apple M5 Max & 4{,}000 & 3{,}455{,}619 & 936 & 273 \\
42 46 \midrule
43 −Total & 33 runs & & & & & 38{,}868 (10.8 GPU-h; baseline study 6.0 GPU-h) & \\
47 +Total & 37 runs & & & & & 43{,}639 (12.1 GPU-h; baseline study 6.0 GPU-h) & \\
44 48 \bottomrule
45 49 \end{tabular}
46 50 \end{adjustbox}
modified results/inline_numbers.json +472 −0
@@ -6944,5 +6944,477 @@
6944 6944 }
6945 6945 }
6946 6946 }
6947 + },
6948 + "budget": {
6949 + "source": "results/metrics/budget/learning_curve.csv + budget_verdict.json (P4, scripts/review/learning_curve.py)",
6950 + "learning_curve": [
6951 + {
6952 + "budget": 12000,
6953 + "model": "mrpt_small",
6954 + "name": "mrpt_small",
6955 + "host": null,
6956 + "steps": 12000.0,
6957 + "wall_s": 1759.3876690865,
6958 + "ex_per_s": 436.5155067835,
6959 + "val_subset": null,
6960 + "val_nll_best": -6.4534611303,
6961 + "val_nll_last": -6.4534611303,
6962 + "val_nll_slope_per_1k": -0.0001691746,
6963 + "n_val_evals": 12.0,
6964 + "lr_at_resume": null,
6965 + "crps_scaled_5": 0.4530927718,
6966 + "crps_5": 0.0007932828,
6967 + "rel_har_5_pct": -1.9056344036,
6968 + "t_cl_har_5": -51.6870464186,
6969 + "rel_rw_5_pct": -4.0537189763,
6970 + "t_cl_rw_5": -41.4739939006,
6971 + "crps_scaled_15": 0.4413554209,
6972 + "crps_15": 0.0013358037,
6973 + "rel_har_15_pct": 0.3419293208,
6974 + "t_cl_har_15": 8.813362543,
6975 + "rel_rw_15_pct": -4.0316649432,
6976 + "t_cl_rw_15": -39.7797357585,
6977 + "crps_scaled_30": 0.4311301909,
6978 + "crps_30": 0.0018423706,
6979 + "rel_har_30_pct": 0.5292370008,
6980 + "t_cl_har_30": 9.2709593131,
6981 + "rel_rw_30_pct": -4.0527658699,
6982 + "t_cl_rw_30": -35.6420338188,
6983 + "crps_scaled_60": 0.4187017316,
6984 + "crps_60": 0.002523578,
6985 + "rel_har_60_pct": 0.6525650454,
6986 + "t_cl_har_60": 8.9204972459,
6987 + "rel_rw_60_pct": -4.2698639812,
6988 + "t_cl_rw_60": -32.1737574342,
6989 + "gap_rel_pct_5": null,
6990 + "gap_t_cl_5": null,
6991 + "gap_rel_pct_15": null,
6992 + "gap_t_cl_15": null,
6993 + "gap_rel_pct_30": null,
6994 + "gap_t_cl_30": null,
6995 + "gap_rel_pct_60": null,
6996 + "gap_t_cl_60": null
6997 + },
6998 + {
6999 + "budget": 12000,
7000 + "model": "mrpt_small_1m_only",
7001 + "name": "mrpt_small_1m_only",
7002 + "host": null,
7003 + "steps": 12000.0,
7004 + "wall_s": 1087.6674129963,
7005 + "ex_per_s": 706.0981976874,
7006 + "val_subset": null,
7007 + "val_nll_best": -6.4520200365,
7008 + "val_nll_last": -6.4520200365,
7009 + "val_nll_slope_per_1k": -0.0001421914,
7010 + "n_val_evals": 12.0,
7011 + "lr_at_resume": null,
7012 + "crps_scaled_5": 0.4532714872,
7013 + "crps_5": 0.0007935916,
7014 + "rel_har_5_pct": -1.8674466033,
7015 + "t_cl_har_5": -51.8908135155,
7016 + "rel_rw_5_pct": -4.016367418,
7017 + "t_cl_rw_5": -41.8341982228,
7018 + "crps_scaled_15": 0.4415638994,
7019 + "crps_15": 0.0013363016,
7020 + "rel_har_15_pct": 0.3793291671,
7021 + "t_cl_har_15": 9.6283302875,
7022 + "rel_rw_15_pct": -3.9958952405,
7023 + "t_cl_rw_15": -40.6351208841,
7024 + "crps_scaled_30": 0.4314122831,
7025 + "crps_30": 0.0018432955,
7026 + "rel_har_30_pct": 0.5797083842,
7027 + "t_cl_har_30": 10.2181128163,
7028 + "rel_rw_30_pct": -4.0045949121,
7029 + "t_cl_rw_30": -36.3956022031,
7030 + "crps_scaled_60": 0.4192359584,
7031 + "crps_60": 0.0025261483,
7032 + "rel_har_60_pct": 0.7550823259,
7033 + "t_cl_har_60": 10.9749079299,
7034 + "rel_rw_60_pct": -4.172360324,
7035 + "t_cl_rw_60": -32.5223804654,
7036 + "gap_rel_pct_5": null,
7037 + "gap_t_cl_5": null,
7038 + "gap_rel_pct_15": null,
7039 + "gap_t_cl_15": null,
7040 + "gap_rel_pct_30": null,
7041 + "gap_t_cl_30": null,
7042 + "gap_rel_pct_60": null,
7043 + "gap_t_cl_60": null
7044 + },
7045 + {
7046 + "budget": 24000,
7047 + "model": "mrpt_small",
7048 + "name": "mrpt_small_c24k",
7049 + "host": "MacStudnpierre2",
7050 + "steps": 24000.0,
7051 + "wall_s": 1927.6599280834,
7052 + "ex_per_s": 398.4105229409,
7053 + "val_subset": 50000.0,
7054 + "val_nll_best": -6.1220689368,
7055 + "val_nll_last": -6.1218395602,
7056 + "val_nll_slope_per_1k": -0.0001630566,
7057 + "n_val_evals": 12.0,
7058 + "lr_at_resume": 0.0001685938,
7059 + "crps_scaled_5": 0.452924514,
7060 + "crps_5": 0.0007929011,
7061 + "rel_har_5_pct": -1.9528320004,
7062 + "t_cl_har_5": -52.3988358765,
7063 + "rel_rw_5_pct": -4.0998830333,
7064 + "t_cl_rw_5": -41.3339362172,
7065 + "crps_scaled_15": 0.4413040312,
7066 + "crps_15": 0.0013357356,
7067 + "rel_har_15_pct": 0.3368106517,
7068 + "t_cl_har_15": 8.3123049256,
7069 + "rel_rw_15_pct": -4.0365605053,
7070 + "t_cl_rw_15": -39.3052422262,
7071 + "crps_scaled_30": 0.4311654587,
7072 + "crps_30": 0.0018429486,
7073 + "rel_har_30_pct": 0.560775491,
7074 + "t_cl_har_30": 9.5829338584,
7075 + "rel_rw_30_pct": -4.0226648666,
7076 + "t_cl_rw_30": -35.1506776245,
7077 + "crps_scaled_60": 0.4187689452,
7078 + "crps_60": 0.002525162,
7079 + "rel_har_60_pct": 0.715742186,
7080 + "t_cl_har_60": 9.2621874546,
7081 + "rel_rw_60_pct": -4.2097765283,
7082 + "t_cl_rw_60": -31.1260754817,
7083 + "gap_rel_pct_5": null,
7084 + "gap_t_cl_5": null,
7085 + "gap_rel_pct_15": null,
7086 + "gap_t_cl_15": null,
7087 + "gap_rel_pct_30": null,
7088 + "gap_t_cl_30": null,
7089 + "gap_rel_pct_60": null,
7090 + "gap_t_cl_60": null
7091 + },
7092 + {
7093 + "budget": 24000,
7094 + "model": "mrpt_small_1m_only",
7095 + "name": "mrpt_small_1m_only_c24k",
7096 + "host": "M5M36",
7097 + "steps": 20000.0,
7098 + "wall_s": 743.2170901299,
7099 + "ex_per_s": 688.8969680589,
7100 + "val_subset": 50000.0,
7101 + "val_nll_best": -6.1198707079,
7102 + "val_nll_last": -6.1194296813,
7103 + "val_nll_slope_per_1k": 5.22885e-05,
7104 + "n_val_evals": 8.0,
7105 + "lr_at_resume": 0.0001685938,
7106 + "crps_scaled_5": 0.4532151677,
7107 + "crps_5": 0.0007935599,
7108 + "rel_har_5_pct": -1.8713624941,
7109 + "t_cl_har_5": -54.2244520434,
7110 + "rel_rw_5_pct": -4.020197558,
7111 + "t_cl_rw_5": -41.3971915353,
7112 + "crps_scaled_15": 0.4418304208,
7113 + "crps_15": 0.0013378214,
7114 + "rel_har_15_pct": 0.4934911362,
7115 + "t_cl_har_15": 12.4285828702,
7116 + "rel_rw_15_pct": -3.8867092385,
7117 + "t_cl_rw_15": -39.8689997176,
7118 + "crps_scaled_30": 0.4319130602,
7119 + "crps_30": 0.0018469954,
7120 + "rel_har_30_pct": 0.7815930931,
7121 + "t_cl_har_30": 12.5976496918,
7122 + "rel_rw_30_pct": -3.8119118677,
7123 + "t_cl_rw_30": -35.964718943,
7124 + "crps_scaled_60": 0.4199447727,
7125 + "crps_60": 0.0025334924,
7126 + "rel_har_60_pct": 1.047999968,
7127 + "t_cl_har_60": 12.5309568992,
7128 + "rel_rw_60_pct": -3.8937678638,
7129 + "t_cl_rw_60": -31.5502836674,
7130 + "gap_rel_pct_5": null,
7131 + "gap_t_cl_5": null,
7132 + "gap_rel_pct_15": null,
7133 + "gap_t_cl_15": null,
7134 + "gap_rel_pct_30": null,
7135 + "gap_t_cl_30": null,
7136 + "gap_rel_pct_60": null,
7137 + "gap_t_cl_60": null
7138 + },
7139 + {
7140 + "budget": 36000,
7141 + "model": "mrpt_small",
7142 + "name": "mrpt_small_c36k",
7143 + "host": "MacStudnpierre2",
7144 + "steps": 34000.0,
7145 + "wall_s": 1563.6283888817,
7146 + "ex_per_s": 409.3044130887,
7147 + "val_subset": 50000.0,
7148 + "val_nll_best": -6.1223114292,
7149 + "val_nll_last": -6.1218243954,
7150 + "val_nll_slope_per_1k": -9.3733e-06,
7151 + "n_val_evals": 10.0,
7152 + "lr_at_resume": 9.88803e-05,
7153 + "crps_scaled_5": 0.4527753305,
7154 + "crps_5": 0.0007926279,
7155 + "rel_har_5_pct": -1.9866196764,
7156 + "t_cl_har_5": -49.5143744088,
7157 + "rel_rw_5_pct": -4.1329308219,
7158 + "t_cl_rw_5": -41.4659723695,
7159 + "crps_scaled_15": 0.4411368498,
7160 + "crps_15": 0.0013349884,
7161 + "rel_har_15_pct": 0.2806854891,
7162 + "t_cl_har_15": 6.722016244,
7163 + "rel_rw_15_pct": -4.0902393458,
7164 + "t_cl_rw_15": -38.9222677398,
7165 + "crps_scaled_30": 0.4309795153,
7166 + "crps_30": 0.0018417289,
7167 + "rel_har_30_pct": 0.4942246442,
7168 + "t_cl_har_30": 8.0308731555,
7169 + "rel_rw_30_pct": -4.086182405,
7170 + "t_cl_rw_30": -34.8617265841,
7171 + "crps_scaled_60": 0.4185283397,
7172 + "crps_60": 0.0025228662,
7173 + "rel_har_60_pct": 0.6241739458,
7174 + "t_cl_har_60": 7.3565662921,
7175 + "rel_rw_60_pct": -4.2968666098,
7176 + "t_cl_rw_60": -30.6639632503,
7177 + "gap_rel_pct_5": null,
7178 + "gap_t_cl_5": null,
7179 + "gap_rel_pct_15": null,
7180 + "gap_t_cl_15": null,
7181 + "gap_rel_pct_30": null,
7182 + "gap_t_cl_30": null,
7183 + "gap_rel_pct_60": null,
7184 + "gap_t_cl_60": null
7185 + },
7186 + {
7187 + "budget": 36000,
7188 + "model": "mrpt_small_1m_only",
7189 + "name": "mrpt_small_1m_only_c36k",
7190 + "host": "M5M36",
7191 + "steps": 30000.0,
7192 + "wall_s": 912.2271790504,
7193 + "ex_per_s": 701.579622596,
7194 + "val_subset": 50000.0,
7195 + "val_nll_best": -6.120694511,
7196 + "val_nll_last": -6.1205803238,
7197 + "val_nll_slope_per_1k": 5.78628e-05,
7198 + "n_val_evals": 10.0,
7199 + "lr_at_resume": 0.000143646,
7200 + "crps_scaled_5": 0.4526041256,
7201 + "crps_5": 0.0007922578,
7202 + "rel_har_5_pct": -2.0323796092,
7203 + "t_cl_har_5": -51.7236941458,
7204 + "rel_rw_5_pct": -4.1776886971,
7205 + "t_cl_rw_5": -41.7867386688,
7206 + "crps_scaled_15": 0.4411428699,
7207 + "crps_15": 0.001335139,
7208 + "rel_har_15_pct": 0.291998869,
7209 + "t_cl_har_15": 7.9942337119,
7210 + "rel_rw_15_pct": -4.0794190811,
7211 + "t_cl_rw_15": -39.3623285293,
7212 + "crps_scaled_30": 0.4312809373,
7213 + "crps_30": 0.0018433191,
7214 + "rel_har_30_pct": 0.5809963109,
7215 + "t_cl_har_30": 10.6489541594,
7216 + "rel_rw_30_pct": -4.0033656875,
7217 + "t_cl_rw_30": -35.0803124675,
7218 + "crps_scaled_60": 0.4194674603,
7219 + "crps_60": 0.0025292926,
7220 + "rel_har_60_pct": 0.8804899075,
7221 + "t_cl_har_60": 12.4698191559,
7222 + "rel_rw_60_pct": -4.0530858192,
7223 + "t_cl_rw_60": -30.3938767976,
7224 + "gap_rel_pct_5": null,
7225 + "gap_t_cl_5": null,
7226 + "gap_rel_pct_15": null,
7227 + "gap_t_cl_15": null,
7228 + "gap_rel_pct_30": null,
7229 + "gap_t_cl_30": null,
7230 + "gap_rel_pct_60": null,
7231 + "gap_t_cl_60": null
7232 + },
7233 + {
7234 + "budget": 12000,
7235 + "model": "H1_gap",
7236 + "name": "full_minus_1m_12000",
7237 + "host": null,
7238 + "steps": null,
7239 + "wall_s": null,
7240 + "ex_per_s": null,
7241 + "val_subset": null,
7242 + "val_nll_best": null,
7243 + "val_nll_last": null,
7244 + "val_nll_slope_per_1k": null,
7245 + "n_val_evals": null,
7246 + "lr_at_resume": null,
7247 + "crps_scaled_5": null,
7248 + "crps_5": -3.088e-07,
7249 + "rel_har_5_pct": null,
7250 + "t_cl_har_5": null,
7251 + "rel_rw_5_pct": -0.0389145079,
7252 + "t_cl_rw_5": -4.3973787059,
7253 + "crps_scaled_15": null,
7254 + "crps_15": -4.979e-07,
7255 + "rel_har_15_pct": null,
7256 + "t_cl_har_15": null,
7257 + "rel_rw_15_pct": -0.0372585138,
7258 + "t_cl_rw_15": -3.6705851182,
7259 + "crps_scaled_30": null,
7260 + "crps_30": -9.25e-07,
7261 + "rel_har_30_pct": null,
7262 + "t_cl_har_30": null,
7263 + "rel_rw_30_pct": -0.0501804829,
7264 + "t_cl_rw_30": -3.3917796089,
7265 + "crps_scaled_60": null,
7266 + "crps_60": -2.5703e-06,
7267 + "rel_har_60_pct": null,
7268 + "t_cl_har_60": null,
7269 + "rel_rw_60_pct": -0.1017489918,
7270 + "t_cl_rw_60": -4.1229218089,
7271 + "gap_rel_pct_5": -0.0389145079,
7272 + "gap_t_cl_5": -4.3973787059,
7273 + "gap_rel_pct_15": -0.0372585138,
7274 + "gap_t_cl_15": -3.6705851182,
7275 + "gap_rel_pct_30": -0.0501804829,
7276 + "gap_t_cl_30": -3.3917796089,
7277 + "gap_rel_pct_60": -0.1017489918,
7278 + "gap_t_cl_60": -4.1229218089
7279 + },
7280 + {
7281 + "budget": 24000,
7282 + "model": "H1_gap",
7283 + "name": "full_minus_1m_24000",
7284 + "host": null,
7285 + "steps": null,
7286 + "wall_s": null,
7287 + "ex_per_s": null,
7288 + "val_subset": null,
7289 + "val_nll_best": null,
7290 + "val_nll_last": null,
7291 + "val_nll_slope_per_1k": null,
7292 + "n_val_evals": null,
7293 + "lr_at_resume": null,
7294 + "crps_scaled_5": null,
7295 + "crps_5": -6.588e-07,
7296 + "rel_har_5_pct": null,
7297 + "t_cl_har_5": null,
7298 + "rel_rw_5_pct": -0.0830231708,
7299 + "t_cl_rw_5": -5.7955317545,
7300 + "crps_scaled_15": null,
7301 + "crps_15": -2.0858e-06,
7302 + "rel_har_15_pct": null,
7303 + "t_cl_har_15": null,
7304 + "rel_rw_15_pct": -0.1559110771,
7305 + "t_cl_rw_15": -8.4853718992,
7306 + "crps_scaled_30": null,
7307 + "crps_30": -4.0469e-06,
7308 + "rel_har_30_pct": null,
7309 + "t_cl_har_30": null,
7310 + "rel_rw_30_pct": -0.2191050918,
7311 + "t_cl_rw_30": -7.0374665828,
7312 + "crps_scaled_60": null,
7313 + "crps_60": -8.3304e-06,
7314 + "rel_har_60_pct": null,
7315 + "t_cl_har_60": null,
7316 + "rel_rw_60_pct": -0.3288118341,
7317 + "t_cl_rw_60": -6.0666476581,
7318 + "gap_rel_pct_5": -0.0830231708,
7319 + "gap_t_cl_5": -5.7955317545,
7320 + "gap_rel_pct_15": -0.1559110771,
7321 + "gap_t_cl_15": -8.4853718992,
7322 + "gap_rel_pct_30": -0.2191050918,
7323 + "gap_t_cl_30": -7.0374665828,
7324 + "gap_rel_pct_60": -0.3288118341,
7325 + "gap_t_cl_60": -6.0666476581
7326 + },
7327 + {
7328 + "budget": 36000,
7329 + "model": "H1_gap",
7330 + "name": "full_minus_1m_36000",
7331 + "host": null,
7332 + "steps": null,
7333 + "wall_s": null,
7334 + "ex_per_s": null,
7335 + "val_subset": null,
7336 + "val_nll_best": null,
7337 + "val_nll_last": null,
7338 + "val_nll_slope_per_1k": null,
7339 + "n_val_evals": null,
7340 + "lr_at_resume": null,
7341 + "crps_scaled_5": null,
7342 + "crps_5": 3.701e-07,
7343 + "rel_har_5_pct": null,
7344 + "t_cl_har_5": null,
7345 + "rel_rw_5_pct": 0.0467092419,
7346 + "t_cl_rw_5": 3.8427215052,
7347 + "crps_scaled_15": null,
7348 + "crps_15": -1.506e-07,
7349 + "rel_har_15_pct": null,
7350 + "t_cl_har_15": null,
7351 + "rel_rw_15_pct": -0.0112804412,
7352 + "t_cl_rw_15": -0.909494338,
7353 + "crps_scaled_30": null,
7354 + "crps_30": -1.5902e-06,
7355 + "rel_har_30_pct": null,
7356 + "t_cl_har_30": null,
7357 + "rel_rw_30_pct": -0.0862704386,
7358 + "t_cl_rw_30": -4.4859784999,
7359 + "crps_scaled_60": null,
7360 + "crps_60": -6.4264e-06,
7361 + "rel_har_60_pct": null,
7362 + "t_cl_har_60": null,
7363 + "rel_rw_60_pct": -0.2540788233,
7364 + "t_cl_rw_60": -7.4893913719,
7365 + "gap_rel_pct_5": 0.0467092419,
7366 + "gap_t_cl_5": 3.8427215052,
7367 + "gap_rel_pct_15": -0.0112804412,
7368 + "gap_t_cl_15": -0.909494338,
7369 + "gap_rel_pct_30": -0.0862704386,
7370 + "gap_t_cl_30": -4.4859784999,
7371 + "gap_rel_pct_60": -0.2540788233,
7372 + "gap_t_cl_60": -7.4893913719
7373 + }
7374 + ],
7375 + "verdict": {
7376 + "criteria": "(i) H1 gap at 36k within the five-seed band of the 12k contrast; (ii) sign of MRPT-S vs HAR-RV at H=30 and 60 unchanged between 12k and 36k",
7377 + "i_gap_inside_band": {
7378 + "5": {
7379 + "gap36k_rel_pct": 0.046709241853904265,
7380 + "band_min_pct": -0.18256088703242176,
7381 + "band_max_pct": -0.038915392667973274,
7382 + "inside": false
7383 + },
7384 + "15": {
7385 + "gap36k_rel_pct": -0.011280441152003365,
7386 + "band_min_pct": -0.34177789642559,
7387 + "band_max_pct": -0.03723313038141836,
7388 + "inside": false
7389 + },
7390 + "30": {
7391 + "gap36k_rel_pct": -0.08627043863992985,
7392 + "band_min_pct": -0.4501766913992362,
7393 + "band_max_pct": -0.050118279111809695,
7394 + "inside": true
7395 + },
7396 + "60": {
7397 + "gap36k_rel_pct": -0.2540788233422957,
7398 + "band_min_pct": -0.5488716027403494,
7399 + "band_max_pct": -0.10156903627846983,
7400 + "inside": true
7401 + }
7402 + },
7403 + "i_pass": false,
7404 + "ii_sign_vs_har": {
7405 + "30": {
7406 + "rel_har_12k_pct": 0.5292370008204252,
7407 + "rel_har_36k_pct": 0.49422464418555406,
7408 + "same_sign": true
7409 + },
7410 + "60": {
7411 + "rel_har_12k_pct": 0.6525650454198514,
7412 + "rel_har_36k_pct": 0.6241739457910326,
7413 + "same_sign": true
7414 + }
7415 + },
7416 + "ii_pass": true,
7417 + "budget_condition_leaves_abstract": false
7418 + }
6947 7419 }
6948 7420 }
\ No newline at end of file
modified scripts/review/headline_numbers.py +17 −0
@@ -219,6 +219,23 @@ def main() -> None:
219 219 for r in q[(q.model == m) & (q.reference == ref)].itertuples()} for m, ref in q[["model", "reference"]].drop_duplicates().itertuples(index=False)}
220 220 e["pairwise"] = pc
221 221 out["lookahead66"][tag] = e
222 + # learning curve (P4): records of budget/learning_curve.csv and the budget verdict, copied verbatim
223 + bd = M / "budget"
224 + if (bd / "learning_curve.csv").exists():
225 + lc = pd.read_csv(bd / "learning_curve.csv")
226 + out["budget"] = {"source": "results/metrics/budget/learning_curve.csv + budget_verdict.json (P4, scripts/review/learning_curve.py)",
227 + "learning_curve": json.loads(lc.drop(columns=[c for c in lc.columns if c == "val_curve"]).to_json(orient="records"))}
228 + if (bd / "budget_verdict.json").exists():
229 + out["budget"]["verdict"] = json.loads((bd / "budget_verdict.json").read_text())
230 + # H3 control (P10): stress-day contrasts of the reduced-global-token twins, copied verbatim
231 + h3 = M / "ablations_h3"
232 + if (h3 / "stress_contrast_test.csv").exists():
233 + sc = pd.read_csv(h3 / "stress_contrast_test.csv")
234 + out["ablations_h3"] = {"source": "results/metrics/ablations_h3/{stress_contrast_test,pairwise_test}.csv (P10)",
235 + "stress_contrast": json.loads(sc.to_json(orient="records"))}
236 + if (h3 / "pairwise_test.csv").exists():
237 + pw3 = pd.read_csv(h3 / "pairwise_test.csv")
238 + out["ablations_h3"]["pairwise_crps_scaled"] = json.loads(pw3[pw3.metric == "crps_scaled"].to_json(orient="records"))
222 239 (ROOT / "results").mkdir(exist_ok=True)
223 240 (ROOT / "results" / "inline_numbers.json").write_text(json.dumps(out, indent=1, default=lambda x: x.item() if hasattr(x, "item") else str(x)))
224 241 # macros for the writing editor (letters only; percentages with two decimals; ranges as "a--b")
225 242
Menu

Theme