Results And Discussion

Research-candidate full-run artifact. This page is generated from tailrisk_20160719_20260522_20260527T083659Z_commit_7f628ff4. It summarizes the durable gold modeling sample and run outputs, not the older bounded access-check snapshot. It is still a research-candidate artifact: final manuscript claims require a clean committed run and author review of the tables and notes. It is organized as the paper's Results and Discussion section: sample and timing results, model setup, forecasting outcomes, diagnostics, tables, figures, and claim boundaries.

1. Overview And Link To Paper Plan

This page is the generated Results and Discussion companion to Paper Plan. It carries forward the planned manuscript sections: data/timing evidence, model/evaluation setup, benchmark results, ML nested information-set results, post-24-check comparisons, supporting diagnostics, and claim boundaries. Full data-source detail lives in Data.

Evidence Map

flowchart LR
  A["Vendor and calendar inputs"] --> B["Bronze / silver caches"]
  B --> C["Gold panel and timing map"]
  C --> D["Leakage and sample gates"]
  D --> E["Baseline benchmarks and advanced econometric benchmarks"]
  D --> F["Primary ML nested information sets"]
  E --> G["Metrics, DM, Murphy diagnostics"]
  F --> G
  G --> I["Tables and figures"]
  I --> J["Generated results snapshot"]

• The left branch binds vendor and calendar inputs into a timestamp-audited gold panel.
• The middle branch compares baseline benchmarks, advanced econometric benchmarks, and ML-tail forecasts on registered loss units.
• The right branch separates primary ML nested information sets, diagnostic model-family comparisons, unconditional DM inference, and supporting figures.

2. Data, Target, And Timing Results

Run Metadata

Field	Value
Run ID	tailrisk_20160719_20260522_20260527T083659Z_commit_7f628ff4
Artifact root	reports/runs/tailrisk_20160719_20260522_20260527T083659Z_commit_7f628ff4
Claim level	research_candidate
Requested window	['2016-07-19', '2026-05-22']
Combined clean start	2018-06-20
Gold panel dates	2016-07-19 to 2026-05-22
Forecast sample dates	2018-06-20 to 2026-05-22 (1722 rows)
Git commit	7f628ff4f66258a36314f492b652cdf7ef594b7e
Git dirty	False
FRED vintage safe	False

• combined_clean_start is the modeling lower bound; dates before it remain audit history rather than forecast evidence.
• git_dirty is recorded so dirty runs can be rejected before manuscript tables are frozen.
• fred_vintage_safe=False is an explicit limitation: FRED data are current historical values with conservative release lag, not real-time vintage observations.

Target Distribution And Tail Diagnostics

• These diagnostics are computed from the raw clean settlement-to-open target gap_t; left loss is -gap_t, and right loss is gap_t.
• The purpose is to show why the dependent variable is a tail-risk object before comparing VaR/ES forecasts.
• Positive tail-shape estimates, heavy empirical tails, and upward mean-excess patterns are empirical support for using heavy-tail approximations such as POT-GPD; they are not a finite-sample proof of Frechet max-domain attraction.
• Raw target diagnostics motivate VaR/ES and EVT modeling. They do not validate LightGBM+EVT forecasts; forecast validity must be read from out-of-sample VaR/ES backtests and loss comparisons.

Target Summary

Measure	Value
Clean forecast observations	1722
Date range	2018-06-20 to 2026-05-22
Mean gap	0.000599 log (+0.06%)
Standard deviation	0.011039 log (+1.11%)
Skewness	-0.066817
Excess kurtosis	11.159
1% quantile	-0.031062 log (-3.06%)
5% quantile	-0.015606 log (-1.55%)
Median	0.001031 log (+0.10%)
95% quantile	0.015357 log (+1.55%)
99% quantile	0.027480 log (+2.79%)
Max drawdown gap	-0.087513 log (-8.38%) on 2020-03-13
Max upside gap	0.096937 log (+10.18%) on 2025-04-10
Jarque-Bera p-value	0
Jarque-Bera statistic	8962.16

Raw-Tail EVT Diagnostics

Tail	Threshold probability	Threshold	Exceedances	Mean excess	GPD xi	GPD scale	Hill xi
left_tail_loss	0.900	0.0160237	78	0.0104227	0.148364	0.00886263	0.432871
left_tail_loss	0.925	0.0195554	59	0.00979449	0.318986	0.00680111	0.342056
left_tail_loss	0.950	0.0223228	39	0.0114201	0.232349	0.0088172	0.354783
left_tail_loss	0.975	0.0293044	20	0.0127979	0.257683	0.00960064	0.31884
left_tail_loss	0.990	0.0373166	8	0.0175619	0.204438	0.0142233	0.342351
right_tail_loss	0.900	0.0150066	92	0.00903798	0.403374	0.00560621	0.381845
right_tail_loss	0.925	0.0169408	69	0.00984642	0.47548	0.00563718	0.370713
right_tail_loss	0.950	0.0189956	46	0.0122032	0.29399	0.00878548	0.41336
right_tail_loss	0.975	0.0259629	23	0.0146916	0.225126	0.0115191	0.383413
right_tail_loss	0.990	0.0369444	10	0.0171855	0.218772	0.0136836	0.352692
absolute_gap	0.900	0.0155233	170	0.00965025	0.287887	0.00694789	0.401867
absolute_gap	0.925	0.0175227	127	0.0106185	0.249472	0.00801625	0.397719
absolute_gap	0.950	0.0208133	85	0.011767	0.256151	0.00884033	0.383021
absolute_gap	0.975	0.0269986	43	0.0144273	0.164437	0.0120976	0.372398
absolute_gap	0.990	0.0371795	17	0.0183049	0.0605131	0.0172189	0.353301

• The GPD threshold table is computed on raw left loss, raw right loss, and the absolute gap; it should not be read as a forecast-model diagnostic.
• The Hill and GPD shape estimates are deliberately reported over multiple thresholds because tail-index estimates are sensitive in samples of this length.

Target Distribution Figures

Figure	Tail side	Source	Claim scope	Docs file
target_tail_motivation	left_right_target_distribution	panel/modeling_panel.parquet	target_distribution_motivation_not_forecast_validation	figures/tailrisk_20160719_20260522_20260527T083659Z_commit_7f628ff4/target_tail_motivation.png

Gold Panel Construction

Measure	Value
Gold modeling rows	2403
Gold columns	1428
Target-audit rows	2403
Clean target rows	2206
Forecast-sample rows	1722
Rows before combined clean start	420
Target-not-clean rows	197
Mapping excluded rows	64

Target audit reason	Rows
None	2206
roll_sq_excluded	195
missing_previous_jpx_session	1
missing_reference_price	1

• The cache lower bound is 2016-07-19, but XLC/core predictor coverage pushes the actual forecast sample to the combined clean start.
• Target exclusion is explicit: roll/SQ windows and the single missing reference price are carried as audit evidence, not silently dropped.
• The forecast-sample reason column makes the sample boundary reproducible row by row.

Calendar And Timing Map

Measure	Value
Normal trading mappings	2333
U.S./Japan desync mappings	1
NYSE early-close mappings	32
EDT rows	1563
EST rows	840

• The map covers EST/EDT, early closes, U.S./Japan holiday desynchronization, and normal trading alignments.
• Desync rows are not treated as normal forecast rows.
• The timing map is part of the leakage-bound gold artifact, not ad hoc evaluation logic.

Feature Coverage

Source family	Block	Features	Mean missing	Max missing
Asia proxy	Asia proxy	10	0.000%	0.000%
cboe_volatility	fred_core	2	0.000%	0.000%
cross_market_derived	Asia proxy	1	0.000%	0.000%
cross_market_derived	fred_core	2	0.000%	0.000%
cross_market_derived	JP proxy	2	0.000%	0.000%
cross_market_derived	US core	2	0.000%	0.000%
event_calendar	calendar_controls	7	0.000%	0.000%
fred_core	fred_core	9	0.000%	0.000%
FRED credit enriched	FRED credit enriched	4	62.398%	62.427%
fx_core	fx_core	4	0.000%	0.000%
JP history	JP only	37	0.005%	0.058%
JP proxy	JP proxy	8	0.000%	0.000%
J-Quants N225 options	JP only	30	1.605%	14.634%
massive_daily	US core	40	0.001%	0.058%
massive_minute	Asia proxy	60	0.000%	0.000%
massive_minute	JP proxy	24	0.348%	4.181%
massive_minute	US late session	84	0.000%	0.000%
massive_optional	massive_optional	2	0.000%	0.000%

• U.S. core, proxy ETFs, minute late-session features, CBOE VIX, FRED rates, FRED H.10 FX, and any audit-gated options-risk fields are separated by source family and block.
• Credit-spread FRED features are enriched/optional and visibly late-starting, so they do not move the core clean start.
• Feature coverage should be read together with the leakage summary; high coverage alone is not enough without timestamp validity.

Leakage Audit

Field	Value
Status	pass_with_warnings
Rows audited	783378
Failures	0
Warnings	611790
Panel row count	2403
Panel signature seed	42
Panel signature	8094755ffc96b01af6fb904876e0abdd3920370fa1b07e44c2c95681cd3e5431

• Zero failures means no audited row violated the hard timestamp invariant.
• Warnings are retained because they identify conservative-lag or missing-feature situations that may matter for interpretation.
• The panel signature is deterministic and binds the leakage check to the current gold panel/config.

3. Model And Evaluation Setup

Pipeline Structure

Step	Layer	Purpose
1	Vendor and calendar sources	Pull or read J-Quants, Massive, FRED, CBOE, and exchange-calendar inputs.
2	Bronze and silver cache	Preserve typed vendor/cache rows, then normalize point-in-time research features.
3	Gold modeling panel	Join targets, calendar map, feature coverage, and leakage-bound signatures.
4	Leakage and coverage gates	Enforce timestamp ordering and sample eligibility before evaluation.
5	Baseline benchmarks and ML-tail registry	Run target-history/econometric baseline benchmarks and LightGBM tail-model families.
6	Metrics, inference, diagnostics	Build loss matrices, DM/Murphy diagnostics, stress windows, and result matrix artifacts.
7	Results snapshot	Summarize run-specific evidence and claim boundaries for reader review.

• Data-access and cache artifacts live under data/bronze and data/silver.
• Durable modeling evidence lives under data/gold; forecast/evaluation/reporting read from gold and reports.
• Run-specific forecasts, metrics, diagnostics, and LaTeX tables live under reports/runs/<run_id>.

Model And Evaluation Protocol

• The registered risk level is tail_level = 0.95; the nominal VaR exception rate is 5%.
• A VaR exception is counted when realized_loss > var_forecast; this follows the standard exception-counting logic of VaR backtesting, but the snapshot does not apply Basel green/yellow/red traffic-light capital zones.
• Forecast evaluation is based on coverage diagnostics, Kupiec/Christoffersen tests where available, quantile loss, Fissler-Ziegel joint VaR-ES loss, and DM inference.
• Benchmarks use target-history information only. ML-tail models add predictors through fixed nested information sets.
• Most specifications use expanding pre-forecast training histories. The rolling-quantile benchmark is the designed exception and uses the most recent 1,000 clean observations.
• LightGBM hyperparameters are held fixed across information sets and refit dates; the snapshot reports model-family evidence rather than tuning-search evidence.
• DM inference is read on average across the unconditional evaluation sample.

4. Forecasting Results And Discussion

4.1 Benchmark results

Status: completed; forecast rows: 15173; metric rows: 14; failures: 0.

Benchmark layer	Status	Forecast rows	Diagnostic rows	Failures	How to read it
baseline	completed	8664	12	0	Implemented evidence for target-history and econometric baseline benchmark models.
advanced econometric	completed_nonblocking	6509	2526	0	Implemented nonblocking advanced econometric benchmark forecasts; review with common-sample gates.

Model	Information set	Tail side	Rows	VaR breach rate	Exceptions	Mean quantile loss	Mean FZ loss
ewma_vol_scaled	Target history	left_tail	722	5.263%	38	0.00140906	-3.64746
ewma_vol_scaled	Target history	right_tail	722	4.571%	33	0.00134173	-3.65662
garch_t	Target history	left_tail	722	6.094%	44	0.00136831	-3.70108
garch_t	Target history	right_tail	722	4.155%	30	0.00127794	-3.70337
gas_t_location_scale	Target history	left_tail	722	6.371%	46	0.00135225	-3.69351
gas_t_location_scale	Target history	right_tail	722	4.986%	36	0.00130686	-3.66834
gjr_garch_evt	Target history	left_tail	722	6.094%	44	0.00133021	-3.74623
gjr_garch_evt	Target history	right_tail	722	5.817%	42	0.00123302	-3.70857
gjr_garch_t	Target history	left_tail	722	7.064%	51	0.00133844	-3.72346
gjr_garch_t	Target history	right_tail	722	4.294%	31	0.00122154	-3.73971
historical_quantile	Target history	left_tail	722	5.540%	40	0.00147622	-3.54066
historical_quantile	Target history	right_tail	722	6.925%	50	0.00150215	-3.40678
rolling_quantile	Target history	left_tail	722	5.817%	42	0.00148114	-3.52468
rolling_quantile	Target history	right_tail	722	7.202%	52	0.00149558	-3.42732

• Baseline benchmark rows set the target-history/econometric reference that ML models should be interpreted against.
• Advanced econometric benchmark families are nonblocking; rows with valid forecasts are empirical evidence subject to the same sample and inference gates, while unavailable rows remain diagnostics.
• The table is not a leaderboard by itself; coverage, exception counts, quantile loss, and FZ loss must be read together.
• Common-sample rows are reported directly so readers can see the effective evidence size.

4.2 Primary ML specifications across nested information sets

Status: completed LGBM ML-tail models; implemented models: LGBM direct quantile, LGBM location-scale empirical, LGBM POT-GPD plain MLE, LGBM POT-GPD UniBM block-maxima shape, LGBM median/MAD POT-GPD plain MLE, LGBM median/MAD POT-GPD UniBM block-maxima shape, LGBM median/IQR POT-GPD plain MLE, LGBM median/IQR POT-GPD UniBM block-maxima shape; forecast rows: 43088; failures: 0.

Model	Information set	Tail side	Rows	VaR breach rate	Exceptions	Mean quantile loss	Mean FZ loss
LGBM direct quantile	JP only	left_tail	527	8.159%	43	0.00141174	-3.48935
LGBM direct quantile	JP + US close core	left_tail	527	11.195%	59	0.00115754	-3.6684
LGBM direct quantile	JP + US close core + JP proxy	left_tail	527	11.765%	62	0.00111393	-3.86888
LGBM direct quantile	JP + US close core + JP proxy + Asia proxy	left_tail	527	11.765%	62	0.00111901	-3.8076
LGBM direct quantile	JP only	right_tail	527	9.677%	51	0.00130995	-3.48743
LGBM direct quantile	JP + US close core	right_tail	527	11.954%	63	0.00125168	-3.48624
LGBM direct quantile	JP + US close core + JP proxy	right_tail	527	11.575%	61	0.00121124	-3.55746
LGBM direct quantile	JP + US close core + JP proxy + Asia proxy	right_tail	527	12.903%	68	0.00122247	-3.55915

• This primary ML table remains strict and reports only ML-tail rows that pass the registered common-sample and forecast-validity gates; coverage is reviewed separately.
• Location-scale empirical and plain POT-GPD are primary candidates only after their valid OOS coverage, standardized-loss, exceedance, and ES-validity gates pass.
• Differences across information blocks are candidate forecast evidence only after the common-sample, coverage, and inference diagnostics are reviewed.
• Coverage review: 8/8 primary ML rows differ from the expected breach rate by more than 2.5 percentage points, so quantile/FZ loss differences alone must not be read as forecast improvement.

4.3 Side-specific ML-tail promotion gate

Role	Model	Information set	Tail side	Rows	Breach	Q loss	FZ loss	DM q	DM FZ	Gate
left promoted	LGBM median/IQR POT-GPD plain MLE	JP + US close core + JP proxy + Asia proxy	left_tail	527	5.882%	0.000917119	-4.22247	-0.000235864; p=0.028; reject10	-0.488666; p=0.049; reject10	pass
right promoted	LGBM location-scale empirical	JP + US close core + JP proxy	right_tail	493	6.085%	0.00102336	-4.02729	-0.000238584; p=0.026; reject10	-0.529731; p=0.003; reject10	pass

• This paper-facing bridge promotes side-specific ML-tail candidates only after the N/coverage gate and restricted common-sample inference are visible.
• The current run's promoted rows are exactly the rows shown above; read them as side-specific paper candidates, not as a universal family ranking.
• This is not a universal model-family ranking and does not replace the strict primary nested-information-set table above.

4.4 ML-tail artifact relationship

Artifact	Rows	Role	Claim boundary
ml_tail_metrics.parquet	8	Primary ML nested-information-set comparison	Eligible for primary discussion after author review.
ml_tail_metrics_per_model.parquet	64	Per-model diagnostics on each model's own valid OOS rows	Not a cross-model comparison and not a replacement primary ML table.
ml_tail_result_matrix.parquet	384	Restricted common-sample VaR-only and VaR-ES comparisons	Restricted evidence; direct quantile rows here are comparison anchors.

• ml_tail_metrics.parquet is the primary nested-information-set artifact. It contains the ML-tail rows that survived the strict common-sample gate in this run.
• ml_tail_metrics_per_model.parquet reports each implemented ML-tail model on its own valid OOS rows; it is useful for debugging coverage but is not a cross-model comparison table.
• ml_tail_result_matrix.parquet creates restricted common samples for VaR-only and VaR-ES comparisons across model families and within-model information-set increments.

4.5 All-model diagnostic scan

Suite	Model	Information set	Metric rows	OOS N mean+-sd	Breach mean+-sd	Abs cov err mean+-sd	Q loss mean+-sd	FZ loss mean+-sd	ES severity mean+-sd
benchmark_advanced	care_expectile_asymmetric_slope	Target history	2	653 +/- 36.7696	7.884% +/- 0.097%	2.884% +/- 0.097%	0.00138317 +/- 8.00061e-05	-3.57653 +/- 0.0638322	0.0082744 +/- 0.000559256
benchmark_advanced	care_expectile_sav	Target history	2	649.5 +/- 41.7193	7.658% +/- 1.250%	2.658% +/- 1.250%	0.00141367 +/- 5.95611e-05	-3.54112 +/- 0.0938805	0.00857223 +/- 0.000528859
benchmark_advanced	caviar_asymmetric_slope	Target history	2	506 +/- 2.82843	6.916% +/- 0.241%	1.916% +/- 0.241%	0.00152915 +/- 4.9109e-05	-3.49931 +/- 0.0468367	0.0100652 +/- 0.000926358
benchmark_advanced	caviar_sav	Target history	2	501 +/- 4.24264	6.086% +/- 0.372%	1.086% +/- 0.372%	0.00156273 +/- 6.81431e-06	-3.47395 +/- 0.093514	0.0116523 +/- 3.28833e-05
benchmark_advanced	gas_t_location_scale	Target history	2	722 +/- 0	5.679% +/- 0.979%	0.693% +/- 0.960%	0.00132955 +/- 3.20964e-05	-3.68092 +/- 0.0177958	0.00959201 +/- 0.00130427
benchmark_advanced	gas_t_pot_gpd	Target history	2	223 +/- 0	6.502% +/- 2.854%	2.018% +/- 2.124%	0.00153061 +/- 0.000359334	-3.41151 +/- 0.534257	0.009774 +/- 0.00427048
benchmark_baseline	ewma_vol_scaled	Target history	2	722 +/- 0	4.917% +/- 0.490%	0.346% +/- 0.118%	0.00137539 +/- 4.76049e-05	-3.65204 +/- 0.00647254	0.00921114 +/- 0.000813727
benchmark_baseline	garch_t	Target history	2	722 +/- 0	5.125% +/- 1.371%	0.970% +/- 0.176%	0.00132312 +/- 6.39034e-05	-3.70223 +/- 0.00162002	0.00974876 +/- 0.00161122
benchmark_baseline	gjr_garch_evt	Target history	2	722 +/- 0	5.956% +/- 0.196%	0.956% +/- 0.196%	0.00128162 +/- 6.87209e-05	-3.7274 +/- 0.0266288	0.00830943 +/- 0.000959208
benchmark_baseline	gjr_garch_t	Target history	2	722 +/- 0	5.679% +/- 1.959%	1.385% +/- 0.960%	0.00127999 +/- 8.26617e-05	-3.73159 +/- 0.0114919	0.00881549 +/- 0.00205961
benchmark_baseline	historical_quantile	Target history	2	722 +/- 0	6.233% +/- 0.979%	1.233% +/- 0.979%	0.00148919 +/- 1.83406e-05	-3.47372 +/- 0.0946715	0.0120717 +/- 9.68084e-05
benchmark_baseline	rolling_quantile	Target history	2	722 +/- 0	6.510% +/- 0.979%	1.510% +/- 0.979%	0.00148836 +/- 1.02111e-05	-3.476 +/- 0.068848	0.0115817 +/- 0.000143645
ml_tail	LGBM POT-GPD UniBM block-maxima shape	JP only	2	484 +/- 0	4.855% +/- 1.023%	0.723% +/- 0.205%	0.00143954 +/- 0.000108633	-3.54743 +/- 0.050522	0.0101296 +/- 0.000667455
ml_tail	LGBM POT-GPD UniBM block-maxima shape	JP + US close core	2	483 +/- 0	5.901% +/- 0.732%	0.901% +/- 0.732%	0.00102253 +/- 3.45998e-05	-4.02862 +/- 0.0642791	0.00630662 +/- 0.000886595
ml_tail	LGBM POT-GPD UniBM block-maxima shape	JP + US close core + JP proxy	2	473.5 +/- 0.707107	6.019% +/- 0.457%	1.019% +/- 0.457%	0.00101481 +/- 2.40341e-05	-4.07568 +/- 0.0784556	0.0063269 +/- 0.000530558
ml_tail	LGBM POT-GPD UniBM block-maxima shape	JP + US close core + JP proxy + Asia proxy	2	473.5 +/- 0.707107	6.125% +/- 0.606%	1.125% +/- 0.606%	0.00104182 +/- 6.17669e-05	-3.99333 +/- 0.0899169	0.00643634 +/- 0.00132792
ml_tail	LGBM POT-GPD plain MLE	JP only	2	484 +/- 0	4.649% +/- 1.315%	0.930% +/- 0.497%	0.00144417 +/- 0.000106677	-3.55083 +/- 0.03612	0.0106097 +/- 0.00111591
ml_tail	LGBM POT-GPD plain MLE	JP + US close core	2	482 +/- 0	5.498% +/- 0.440%	0.498% +/- 0.440%	0.00102429 +/- 3.8851e-05	-4.03154 +/- 0.069105	0.00653613 +/- 0.000836497
ml_tail	LGBM POT-GPD plain MLE	JP + US close core + JP proxy	2	473.5 +/- 0.707107	5.808% +/- 0.158%	0.808% +/- 0.158%	0.00101213 +/- 3.0664e-05	-4.0908 +/- 0.0864067	0.00623547 +/- 0.000519173
ml_tail	LGBM POT-GPD plain MLE	JP + US close core + JP proxy + Asia proxy	2	473 +/- 0	6.131% +/- 0.598%	1.131% +/- 0.598%	0.00103841 +/- 7.23005e-05	-4.00776 +/- 0.117572	0.00622715 +/- 0.0016854
ml_tail	LGBM direct quantile	JP only	2	554 +/- 0	8.664% +/- 1.021%	3.664% +/- 1.021%	0.00133825 +/- 9.25573e-05	-3.50275 +/- 0.0392189	0.00813676 +/- 0.00113574
ml_tail	LGBM direct quantile	JP + US close core	2	554 +/- 0	11.101% +/- 0.383%	6.101% +/- 0.383%	0.00118022 +/- 4.32589e-05	-3.61364 +/- 0.0922633	0.0064386 +/- 0.000247838
ml_tail	LGBM direct quantile	JP + US close core + JP proxy	2	527 +/- 0	11.670% +/- 0.134%	6.670% +/- 0.134%	0.00116259 +/- 6.88028e-05	-3.71317 +/- 0.220204	0.0059796 +/- 0.000662772
ml_tail	LGBM direct quantile	JP + US close core + JP proxy + Asia proxy	2	527 +/- 0	12.334% +/- 0.805%	7.334% +/- 0.805%	0.00117074 +/- 7.31616e-05	-3.68337 +/- 0.17568	0.00571666 +/- 0.000192238
ml_tail	LGBM location-scale empirical	JP only	2	508 +/- 0	4.823% +/- 0.696%	0.492% +/- 0.251%	0.00141628 +/- 9.98491e-05	-3.56404 +/- 0.0139556	0.00998686 +/- 0.000171526
ml_tail	LGBM location-scale empirical	JP + US close core	2	505.5 +/- 0.707107	6.133% +/- 0.009%	1.133% +/- 0.009%	0.00101756 +/- 4.15541e-05	-4.01111 +/- 0.0557055	0.00604578 +/- 0.000458165
ml_tail	LGBM location-scale empirical	JP + US close core + JP proxy	2	492.5 +/- 0.707107	6.295% +/- 0.296%	1.295% +/- 0.296%	0.00100557 +/- 2.51561e-05	-4.08481 +/- 0.0813422	0.0059098 +/- 0.000533713
ml_tail	LGBM location-scale empirical	JP + US close core + JP proxy + Asia proxy	2	492 +/- 0	6.707% +/- 0.575%	1.707% +/- 0.575%	0.00103557 +/- 7.0027e-05	-3.98452 +/- 0.104636	0.00590198 +/- 0.00152973
ml_tail	LGBM median/IQR POT-GPD UniBM block-maxima shape	JP only	2	554 +/- 0	3.430% +/- 0.255%	1.570% +/- 0.255%	0.00129247 +/- 0.000128665	-3.72329 +/- 0.0553233	0.0103895 +/- 0.000332801
ml_tail	LGBM median/IQR POT-GPD UniBM block-maxima shape	JP + US close core	2	553.5 +/- 0.707107	5.420% +/- 0.249%	0.420% +/- 0.249%	0.001005 +/- 6.42702e-05	-4.03849 +/- 0.105695	0.00680349 +/- 2.92911e-05
ml_tail	LGBM median/IQR POT-GPD UniBM block-maxima shape	JP + US close core + JP proxy	2	527 +/- 0	4.934% +/- 0.268%	0.190% +/- 0.094%	0.000977369 +/- 9.96524e-05	-4.07223 +/- 0.267913	0.00722562 +/- 0.000739001
ml_tail	LGBM median/IQR POT-GPD UniBM block-maxima shape	JP + US close core + JP proxy + Asia proxy	2	527 +/- 0	5.882% +/- 0.000%	0.882% +/- 0.000%	0.000981983 +/- 9.6245e-05	-4.07821 +/- 0.191496	0.00617459 +/- 0.00100583
ml_tail	LGBM median/IQR POT-GPD plain MLE	JP only	2	554 +/- 0	3.339% +/- 0.383%	1.661% +/- 0.383%	0.00129732 +/- 0.000121471	-3.71766 +/- 0.0543739	0.0104723 +/- 0.000488415
ml_tail	LGBM median/IQR POT-GPD plain MLE	JP + US close core	2	553.5 +/- 0.707107	5.330% +/- 0.121%	0.330% +/- 0.121%	0.00100641 +/- 5.8567e-05	-4.05822 +/- 0.117346	0.00677553 +/- 9.19624e-05
ml_tail	LGBM median/IQR POT-GPD plain MLE	JP + US close core + JP proxy	2	526 +/- 0	4.753% +/- 0.000%	0.247% +/- 0.000%	0.000972485 +/- 9.62715e-05	-4.06369 +/- 0.190447	0.00717965 +/- 0.000903805
ml_tail	LGBM median/IQR POT-GPD plain MLE	JP + US close core + JP proxy + Asia proxy	2	527 +/- 0	5.977% +/- 0.134%	0.977% +/- 0.134%	0.000981802 +/- 9.14751e-05	-4.07318 +/- 0.21113	0.00595077 +/- 0.000647845
ml_tail	LGBM median/MAD POT-GPD UniBM block-maxima shape	JP only	2	484 +/- 0	5.579% +/- 0.584%	0.579% +/- 0.584%	0.0014127 +/- 0.000111137	-3.60619 +/- 0.0108432	0.0101188 +/- 0.00167497
ml_tail	LGBM median/MAD POT-GPD UniBM block-maxima shape	JP + US close core	2	484 +/- 0	6.921% +/- 0.438%	1.921% +/- 0.438%	0.00109506 +/- 3.41604e-05	-4.06831 +/- 0.110891	0.00792169 +/- 0.000308992
ml_tail	LGBM median/MAD POT-GPD UniBM block-maxima shape	JP + US close core + JP proxy	2	473.5 +/- 0.707107	7.603% +/- 0.310%	2.603% +/- 0.310%	0.00103226 +/- 7.39019e-05	-4.16953 +/- 0.157826	0.00660286 +/- 0.000522774
ml_tail	LGBM median/MAD POT-GPD UniBM block-maxima shape	JP + US close core + JP proxy + Asia proxy	2	473.5 +/- 0.707107	7.497% +/- 0.161%	2.497% +/- 0.161%	0.00106577 +/- 8.54015e-05	-4.09025 +/- 0.239148	0.00711434 +/- 0.000949755
ml_tail	LGBM median/MAD POT-GPD plain MLE	JP only	2	484 +/- 0	5.165% +/- 0.000%	0.165% +/- 0.000%	0.00141016 +/- 0.000113401	-3.60412 +/- 0.009619	0.0105104 +/- 0.00101589
ml_tail	LGBM median/MAD POT-GPD plain MLE	JP + US close core	2	484 +/- 0	6.818% +/- 0.584%	1.818% +/- 0.584%	0.00109354 +/- 3.73982e-05	-4.05872 +/- 0.0856922	0.00790369 +/- 0.000394935
ml_tail	LGBM median/MAD POT-GPD plain MLE	JP + US close core + JP proxy	2	473.5 +/- 0.707107	7.287% +/- 0.758%	2.287% +/- 0.758%	0.00102391 +/- 8.23481e-05	-4.19814 +/- 0.158674	0.00660111 +/- 0.000331299
ml_tail	LGBM median/MAD POT-GPD plain MLE	JP + US close core + JP proxy + Asia proxy	2	473.5 +/- 0.707107	7.076% +/- 0.757%	2.076% +/- 0.757%	0.00105943 +/- 9.34222e-05	-4.09633 +/- 0.245078	0.0073057 +/- 0.000570338

• This table joins benchmark_metrics_per_model.parquet and ml_tail_metrics_per_model.parquet so all benchmark and LGBM tail-model variants are visible in one place.
• Mean and standard deviation are computed across registered metric rows for the same suite/model/information-set configuration; for most rows this summarizes left- and right-tail metrics.
• It is a diagnostic scan, not the formal cross-model comparison table. Cross-model claims still require common-sample result-matrix and DM evidence because valid dates and model gates can differ.

4.6 Restricted common-sample result matrix and DM evidence

Family	Axis	Loss	Rows	Common N	Date range	Joint exceptions
nested information sets	information_set_increment	var_coverage	64	471 to 527	2023-01-26 to 2026-05-21	34 to 80
nested information sets	information_set_increment	var_es_fz_loss	64	471 to 527	2023-01-26 to 2026-05-21	34 to 80
nested information sets	information_set_increment	var_quantile_loss	64	471 to 527	2023-01-26 to 2026-05-21	34 to 80
tail_model_family	model_family	var_coverage	64	472 to 484	2023-06-16 to 2026-05-21	47 to 70
tail_model_family	model_family	var_es_fz_loss	64	472 to 484	2023-06-16 to 2026-05-21	47 to 70
tail_model_family	model_family	var_quantile_loss	64	472 to 484	2023-06-16 to 2026-05-21	47 to 70

• The result matrix is the right place to compare direct quantile, location-scale empirical, plain POT-GPD, and the robust plain POT-GPD routes on their restricted common dates.
• It separates VaR-only losses from VaR-ES joint scoring, so VaR-only claims are not confused with ES claims.
• Restricted direct-quantile performance is only a comparison anchor for the tail-model family; it does not replace the primary direct-quantile evidence.
• DM records are emitted only where registered row-count and exception-count gates pass; otherwise the result matrix remains descriptive.

4.7 Stress and diagnostic windows

Suite	Rows	Window labels
benchmark	146	loss_top_decile
ml_tail	212	loss_top_decile, vix_top_decile

• Stress windows identify high-loss or high-volatility subsamples for two-sided risk diagnostics.
• These rows use reproducible full-sample classifiers in this first pass, so they should be described as diagnostics rather than a live stress classifier.
• They are useful for finding whether model behavior changes in difficult regimes before writing manuscript discussion.

4.8 Integrated interpretation and claim boundaries

Data and timing audit

• The gold timing map covers 2016-07-19 to 2026-05-22 and the combined clean start is 2018-06-20.
• No forecast-sample rows before 2018-06-20 enter the modeling evidence.
• The leakage check reports status pass_with_warnings with zero leakage failures and 611790 warnings.
• FRED vintage safety is recorded as False; FRED values use conservative release timing but remain current historical observations rather than ALFRED real-time vintages.

Baseline benchmarks and advanced econometric benchmarks

• benchmark_metrics.parquet reports 14 common-sample rows across 7 baseline benchmark model families and 2 tail side(s), while benchmark forecasts contain 15173 model-date rows.
• Baseline benchmark models are external target-history and econometric references; this section does not rank them.
• Advanced econometric benchmark rows are implemented for 6 model families and contribute 6509 nonblocking forecast rows; these rows are claim-gated diagnostics unless a manuscript table explicitly promotes them through the same sample and inference review.
• Baseline benchmark breach rates have a median of 0.0581717, within 2.5 percentage points of the nominal level, indicating reasonable coverage calibration relative to the ML-tail models whose breach rates are reported in the nested-information-set section.

Primary ML specifications across nested information sets

• ml_tail_metrics.parquet defines the primary ML specification comparison across nested information sets for this run.
• The primary ML artifact contains 4 information sets, 1 tail level(s), and 2 tail side(s); the retained primary ML rows are LGBM direct quantile.
• The implemented ML-tail registry is LGBM direct quantile, LGBM location-scale empirical, LGBM POT-GPD plain MLE, LGBM POT-GPD UniBM block-maxima shape, LGBM median/MAD POT-GPD plain MLE, LGBM median/MAD POT-GPD UniBM block-maxima shape, LGBM median/IQR POT-GPD plain MLE, LGBM median/IQR POT-GPD UniBM block-maxima shape, but the primary nested-information-set comparison should be read only from ml_tail_metrics.parquet.
• The nested information sets report downside-risk and upside-risk surfaces separately. The registered artifacts show different left/right patterns, and the generator does not assume that the two sides share the same economic mechanism.
• Coverage warning: all 8 primary ML rows exhibit VaR breach rates (0.0815939 to 0.129032) that exceed the nominal level by more than 2.5 percentage points. Quantile-loss and FZ-loss differences across the nested information sets must be interpreted in this context; lower loss scores may partly reflect less conservative VaR estimates rather than better conditional tail calibration.
• For left_tail / LGBM direct quantile / tail=0.950, the largest quantile-loss change occurs at the first information-set augmentation (adding U.S. close core); subsequent additions of Japan proxy and Asia proxy ETFs contribute diminishing incremental loss changes. This saturation pattern is descriptive and does not automatically reduce the value of the broader information set.
• The nested information sets are used to assess candidate incremental U.S.-close information under strict common-sample rules; they do not by themselves establish forecast improvement.

Restricted model-family comparison

• ml_tail_result_matrix.parquet contains restricted common-sample comparisons for 8 LightGBM tail-model families.
• The restricted common-N range is 471 to 527 and the joint-exception range is 34 to 80.
• Recorded claim scopes are restricted_model_comparison_not_primary; these rows are restricted evidence and cannot replace the primary ML nested-information-set comparison.
• The tail-model family comparison is severely sample-limited: the largest restricted common-N is 484 rows. No model-family ranking claim is supportable from this restricted sample; extended OOS coverage is needed before tail-model family ranking becomes meaningful.
• Result-matrix inference is recorded separately from the primary suite-level DM: restricted DM records include 208 gate-pass rows and 104 unavailable rows. These entries are restricted common-sample diagnostics, not primary model-family rankings.
• The result matrix is a matched-date diagnostic layer. It should not be worded as one family being better than another.

Coverage and inference gates

• Coverage review flags 8/8 primary ML rows with breach rates more than 2.5 percentage points from nominal coverage; Kupiec p-values fall below 0.05 in 8/8 reported rows and Christoffersen p-values fall below 0.05 in 0/8 reported rows.
• Model-eviction artifacts record 8 retained rows and 56 non-retained rows under the primary ML sample policy.
• Block-bootstrap DM artifacts are unconditional forecast-comparison diagnostics; any p-value should be read on average across the unconditional evaluation sample, not as condition-specific evidence.
• Loss differentials alone do not constitute an improvement claim; coverage, exception counts, sample gates, and inference status must be reviewed together.
• Result-matrix tail-event power flags and suite-level inference gates report 0 restricted rows with insufficient tail-event power and 0/18 unavailable DM inference rows.

Supporting diagnostics

• Supporting LaTeX diagnostic table files are present for 2/2 registered diagnostic families.
• ES severity diagnostics contain 86 finite rows with mean exceedance severity ranging from 0.00482029 to 0.0121402; this is conditional-on-exception evidence.
• Stress-window diagnostics contain 358 rows, and 24-check LGBM Murphy diagnostics contain 3200 rows.
• Feature-unavailability diagnostics contain 384 rows.
• Figure manifest references:
• Figure: market_timing_design (Source: manifest.json, config/research_config.json, panel/calendar_map.parquet; Claim scope: design_forecast_origin_not_causal_price_discovery; File: latex/figures/market_timing_design.png).
• Figure: coverage_breach_rates_left_tail (Source: metrics/benchmark_metrics.parquet, metrics/benchmark_metrics_per_model.parquet, metrics/ml_tail_metrics.parquet, metrics/ml_tail_metrics_per_model.parquet; Claim scope: coverage_diagnostic_not_primary_claim; File: latex/figures/coverage_breach_rates_left_tail.png).
• Figure: coverage_breach_rates_right_tail (Source: metrics/benchmark_metrics.parquet, metrics/benchmark_metrics_per_model.parquet, metrics/ml_tail_metrics.parquet, metrics/ml_tail_metrics_per_model.parquet; Claim scope: coverage_diagnostic_not_primary_claim; File: latex/figures/coverage_breach_rates_right_tail.png).
• Figure: cumulative_lgbm_a_anchor_fz_gain (Source: metrics/benchmark_loss_matrix.parquet, metrics/ml_tail_loss_matrix.parquet, forecasts/benchmark_forecasts.parquet, forecasts/ml_tail_forecasts.parquet; Claim scope: headline_lgbm_a_anchor_gjr_evt_and_information_increment_fz_gain; File: latex/figures/cumulative_lgbm_a_anchor_fz_gain.png).
• Figure: selected_model_performance_left_tail (Source: metrics/benchmark_metrics_per_model.parquet, metrics/ml_tail_metrics_per_model.parquet; Claim scope: selected_benchmark_vs_lgbm_main_figure_not_full_result_set; File: latex/figures/selected_model_performance_left_tail.png).
• Figure: selected_model_performance_right_tail (Source: metrics/benchmark_metrics_per_model.parquet, metrics/ml_tail_metrics_per_model.parquet; Claim scope: selected_benchmark_vs_lgbm_main_figure_not_full_result_set; File: latex/figures/selected_model_performance_right_tail.png).
• Figure: full_sample_var_overlay_left_tail (Source: forecasts/benchmark_forecasts.parquet, forecasts/ml_tail_forecasts.parquet; Claim scope: full_sample_var_overlay_fixed_selection_visual_diagnostic; File: latex/figures/full_sample_var_overlay_left_tail.png).
• Figure: full_sample_var_overlay_right_tail (Source: forecasts/benchmark_forecasts.parquet, forecasts/ml_tail_forecasts.parquet; Claim scope: full_sample_var_overlay_fixed_selection_visual_diagnostic; File: latex/figures/full_sample_var_overlay_right_tail.png).
• Figure: benchmark_murphy_left_tail (Source: metrics/benchmark_murphy.parquet; Claim scope: murphy_diagnostic_benchmark_baseline_common_grid; File: latex/figures/benchmark_murphy_left_tail.png).
• Figure: benchmark_murphy_right_tail (Source: metrics/benchmark_murphy.parquet; Claim scope: murphy_diagnostic_benchmark_baseline_common_grid; File: latex/figures/benchmark_murphy_right_tail.png).
• Figure: lgbm_24check_murphy_left_tail (Source: metrics/lgbm_24check_murphy.parquet, metrics/ml_tail_metrics_per_model.parquet, forecasts/ml_tail_forecasts.parquet; Claim scope: murphy_diagnostic_lgbm_24check_robust_ladder; File: latex/figures/lgbm_24check_murphy_left_tail.png).
• Figure: lgbm_24check_murphy_right_tail (Source: metrics/lgbm_24check_murphy.parquet, metrics/ml_tail_metrics_per_model.parquet, forecasts/ml_tail_forecasts.parquet; Claim scope: murphy_diagnostic_lgbm_24check_robust_ladder; File: latex/figures/lgbm_24check_murphy_right_tail.png).
• Figure: es_severity_left_tail (Source: metrics/benchmark_metrics.parquet, metrics/ml_tail_metrics.parquet, metrics/ml_tail_metrics_per_model.parquet; Claim scope: es_severity_diagnostic_not_model_selection_claim; File: latex/figures/es_severity_left_tail.png).
• Figure: es_severity_right_tail (Source: metrics/benchmark_metrics.parquet, metrics/ml_tail_metrics.parquet, metrics/ml_tail_metrics_per_model.parquet; Claim scope: es_severity_diagnostic_not_model_selection_claim; File: latex/figures/es_severity_right_tail.png).
• Figure: var_es_stress_overlay_2024_stress_episode (Source: forecasts/benchmark_forecasts.parquet, forecasts/ml_tail_forecasts.parquet; Claim scope: appendix_stress_overlay_illustration_not_validation; File: latex/figures/var_es_stress_overlay_2024_stress_episode.png).
• Figure: var_es_stress_overlay_2025_stress_episode (Source: forecasts/benchmark_forecasts.parquet, forecasts/ml_tail_forecasts.parquet; Claim scope: appendix_stress_overlay_illustration_not_validation; File: latex/figures/var_es_stress_overlay_2025_stress_episode.png).
• Figure: dm_heatmap_left_tail (Source: forecasts/benchmark_forecasts.parquet, forecasts/ml_tail_forecasts.parquet; Claim scope: post_24check_cross_suite_fz_dm_diagnostic; File: latex/figures/dm_heatmap_left_tail.png).
• Figure: dm_heatmap_right_tail (Source: forecasts/benchmark_forecasts.parquet, forecasts/ml_tail_forecasts.parquet; Claim scope: post_24check_cross_suite_fz_dm_diagnostic; File: latex/figures/dm_heatmap_right_tail.png).

Not yet claimed

• No hedge PnL, transaction-cost, or trading-alpha analysis is performed.
• Left-tail and right-tail outputs are both economic tail-risk surfaces for futures positions; neither side should be promoted beyond the sample, coverage, and inference gates without author review.
• The current evidence does not create an automatic model-selection statement; any manuscript claim still requires author review of sample gates, coverage, loss metrics, and inference diagnostics.

5. Figures, Tables, And Source Artifacts

This section merges the former figure/table placement page into the results snapshot. All generated figures and tables are listed with their intended interpretation. The words "supporting" and "diagnostic" describe claim scope; they do not mean the artifact is missing from this page.

5.1 Configuration robustness evidence

Field	Value
Source primary run	tailrisk_20160719_20260522_20260527T083659Z_commit_7f628ff4
Scope	paper
Primary-claim allowed	False
Selected LGBM models	LGBM POT-GPD plain MLE, LGBM POT-GPD UniBM block-maxima shape
Selected benchmark models	gjr_garch_evt
Selected information sets	JP + US close core + JP proxy
Job counts	evt_boundary_rows=6, evt_threshold=12, lgbm_capacity=8
Forecast rows	13096
Metric rows	26
Status	ok

Sensitivity family	Rows / classifications
LGBM capacity	8 rows (robust=8)
POT threshold	18 rows (boundary_diagnostic=6, robust=12)

• The primary design compares pre-specified point-in-time forecast specifications. Configuration sensitivity is post-24-check robustness evidence and is not used for model selection or the cross-suite FZ DM heatmap.
• The run is fixed to the post-24-check paper set. LGBM rows perturb capacity only for the pass-all C-information LGBM+EVT families, and POT threshold rows perturb those rows plus GJR-GARCH-EVT.
• Robustness classes describe conclusion stability versus the registered primary specification. They do not feed DM gates, promoted-model logic, or selected-model figures.

5.2 Generated table manifest

Table	Source artifacts	Claim scope	Tail side	File
tailrisk_predictor_block_coverage	panel/feature_coverage.parquet	main_text_predictor_block_coverage_information_transparency	None	latex/tables/tailrisk_predictor_block_coverage_table.tex
benchmark_metrics	metrics/benchmark_metrics.parquet	benchmark_common_sample_metric_table	None	latex/tables/benchmark_metrics_table.tex
benchmark_left_tail_risk	metrics/benchmark_metrics.parquet	left_tail_benchmark_risk_table	left_tail	latex/tables/benchmark_left_tail_risk_table.tex
benchmark_right_tail_risk	metrics/benchmark_metrics.parquet	right_tail_benchmark_risk_table	right_tail	latex/tables/benchmark_right_tail_risk_table.tex
ml_tail_metrics	metrics/ml_tail_metrics.parquet	ml_tail_nested_information_set_table	None	latex/tables/ml_tail_metrics_table.tex
ml_tail_left_tail_risk	metrics/ml_tail_metrics.parquet	left_tail_ml_tail_primary_risk_table	left_tail	latex/tables/ml_tail_left_tail_risk_table.tex
ml_tail_right_tail_risk	metrics/ml_tail_metrics.parquet	right_tail_ml_tail_primary_risk_table	right_tail	latex/tables/ml_tail_right_tail_risk_table.tex
tailrisk_model_inventory	config/research_config.json, metrics/benchmark_metrics_per_model.parquet, metrics/ml_tail_metrics_per_model.parquet	main_text_model_inventory_forecast_construction	None	latex/tables/tailrisk_model_inventory_table.tex
tailrisk_selected_model_performance	metrics/benchmark_metrics_per_model.parquet, metrics/ml_tail_metrics_per_model.parquet	selected_benchmark_vs_lgbm_main_figure_rows	None	latex/tables/tailrisk_selected_model_performance_table.tex
appendix_benchmark_all_models	metrics/benchmark_metrics_per_model.parquet	appendix_full_benchmark_results	None	latex/tables/appendix_benchmark_all_models_table.tex
ml_tail_promoted_tail_models	metrics/ml_tail_metrics_per_model.parquet, metrics/ml_tail_result_matrix_dm.parquet	side_specific_ml_tail_promotion_gate	None	latex/tables/ml_tail_promoted_tail_models_table.tex
appendix_lgbm_all_models	metrics/ml_tail_metrics_per_model.parquet	appendix_full_lgbm_results	None	latex/tables/appendix_lgbm_all_models_table.tex
tailrisk_es_severity	metrics/benchmark_metrics.parquet, metrics/ml_tail_metrics.parquet, metrics/ml_tail_metrics_per_model.parquet	es_severity_diagnostic_table	None	latex/tables/tailrisk_es_severity_table.tex
tailrisk_claim_scope	manifest.json, config/research_config.json	claim_boundary_reference_table	None	latex/tables/tailrisk_claim_scope_table.tex
ml_tail_result_matrix	metrics/ml_tail_result_matrix.parquet	restricted_model_comparison_table	None	latex/tables/ml_tail_result_matrix_table.tex
ml_tail_result_matrix_summary	metrics/ml_tail_result_matrix.parquet, metrics/ml_tail_result_matrix_dm.parquet	restricted_result_matrix_summary_table	None	latex/tables/ml_tail_result_matrix_summary_table.tex
tailrisk_dm_summary	metrics/ml_tail_result_matrix_dm.parquet	main_text_compact_dm_summary	None	latex/tables/tailrisk_dm_summary_table.tex
appendix_lgbm_configuration_sensitivity	sensitivity/metrics/lgbm_configuration_sensitivity_metrics.parquet	appendix_configuration_robustness_lgbm	None	sensitivity/latex/tables/appendix_lgbm_configuration_sensitivity_table.tex
appendix_evt_threshold_sensitivity	sensitivity/metrics/evt_threshold_sensitivity_metrics.parquet	appendix_configuration_robustness_evt_threshold	None	sensitivity/latex/tables/appendix_evt_threshold_sensitivity_table.tex

• The table manifest records the generated LaTeX table files, their source artifacts, and their claim scopes.
• Tables are paper-facing exports; the Markdown tables above are snapshot summaries for browser review.

5.3 Table interpretation guide

Results/Discussion role	Artifact	How to read it
Predictor block and coverage	tailrisk_predictor_block_coverage_table.tex	Data/Methods table showing source families, feature counts, examples, missingness, and model role; coverage is not timestamp admissibility.
Model inventory	tailrisk_model_inventory_table.tex	Methods table explaining model families, information sets, VaR construction, ES construction, and role; performance belongs elsewhere.
Benchmark floor summary	benchmark_metrics_table.tex	Results table for target-history and econometric benchmark calibration and loss evidence.
Benchmark tail-side details	benchmark_left_tail_risk_table.tex, benchmark_right_tail_risk_table.tex	Tail-specific benchmark rows for left and right risk surfaces.
ML information ladder	ml_tail_metrics_table.tex	Core nested-information-set table for direct LightGBM; read loss changes with coverage gates.
ML tail-side details	ml_tail_left_tail_risk_table.tex, ml_tail_right_tail_risk_table.tex	Tail-specific direct LightGBM information-set rows.
Selected model performance	tailrisk_selected_model_performance_table.tex	Deterministic selected-row summary after sample-size, coverage, FZ-loss, and quantile-loss gates.
Promoted tail rows	ml_tail_promoted_tail_models_table.tex	Locked side-specific promotion-gate rows; not a universal model-family ranking.
Full benchmark scan	appendix_benchmark_all_models_table.tex	Complete benchmark inventory supporting benchmark breadth.
Full LGBM scan	appendix_lgbm_all_models_table.tex	Complete per-model LightGBM scan; do not use as a raw leaderboard.
Restricted result matrix	ml_tail_result_matrix_table.tex, ml_tail_result_matrix_summary_table.tex	Restricted common-sample model-family comparison and summary.
Compact DM summary	tailrisk_dm_summary_table.tex	Headline paired inference table; negative loss differences favor the candidate.
ES severity	tailrisk_es_severity_table.tex	Conditional-on-exception severity diagnostic; not standalone model selection.
Claim boundary	tailrisk_claim_scope_table.tex	Reference table separating headline, restricted, diagnostic, and robustness claims.

Figure 1. Market Timing Design

• Key readings: the diagram defines JST event timing, the matched U.S.-close cutoff, and the OSE day-open target.
• OSE schedule note: pre-2024-11-05 hours use day close 15:15 JST and night session 16:30-05:30 JST; from 2024-11-05, JPX uses day close 15:45 JST and night session 17:00-06:00 JST, with day open still 08:45 JST.
• The OSE night close is timing context; the forecast origin is the matched U.S. cash close plus the data-availability lag.
• It is a session-alignment schematic, not a structural market-transmission diagram.

Figure: market_timing_design. Source: manifest.json, config/research_config.json, panel/calendar_map.parquet. Claim scope: design_forecast_origin_not_causal_price_discovery. Tail side: design. Run file: latex/figures/market_timing_design.png.

Figure 2. Opening-Gap Tail Motivation

• Key readings: the composite figure combines density, log survival, mean-excess, and Hill tail-index diagnostics for the raw opening-gap target.
• It motivates tail-risk modeling and does not validate any forecast model.

Figure: target_tail_motivation. Source: panel/modeling_panel.parquet. Claim scope: target_distribution_motivation_not_forecast_validation. Tail side: left_right_target_distribution. Run file: latex/figures/target_tail_motivation.png.

Figure 3. Cumulative FZ-Gain Diagnostics

• Key readings: upward movement means the candidate has lower cumulative FZ loss under the fixed anchor-loss-minus-candidate-loss convention.
• Each panel uses the corresponding A-only LGBM+EVT forecast as anchor.
• The GJR-GARCH-EVT line is an own-history benchmark reference; B/C/D lines show same-family information increments.

Figure: cumulative_lgbm_a_anchor_fz_gain. Source: metrics/benchmark_loss_matrix.parquet, metrics/ml_tail_loss_matrix.parquet, forecasts/benchmark_forecasts.parquet, forecasts/ml_tail_forecasts.parquet. Claim scope: headline_lgbm_a_anchor_gjr_evt_and_information_increment_fz_gain. Tail side: left_right. Run file: latex/figures/cumulative_lgbm_a_anchor_fz_gain.png.

Figure 4. Full Coverage Breach-Rate Diagnostics

• Key readings: bars report realized VaR exception rates against the nominal line.
• Read this first: exception-rate deviations set the boundary for any loss-based interpretation.

Figure: coverage_breach_rates_left_tail. Source: metrics/benchmark_metrics.parquet, metrics/benchmark_metrics_per_model.parquet, metrics/ml_tail_metrics.parquet, metrics/ml_tail_metrics_per_model.parquet. Claim scope: coverage_diagnostic_not_primary_claim. Tail side: left_tail. Run file: latex/figures/coverage_breach_rates_left_tail.png.

Figure: coverage_breach_rates_right_tail. Source: metrics/benchmark_metrics.parquet, metrics/benchmark_metrics_per_model.parquet, metrics/ml_tail_metrics.parquet, metrics/ml_tail_metrics_per_model.parquet. Claim scope: coverage_diagnostic_not_primary_claim. Tail side: right_tail. Run file: latex/figures/coverage_breach_rates_right_tail.png.

Figure 5. Selected Benchmark-vs-LGBM Performance

• Key readings: compact main-figure rows split models into two broad groups, Benchmark and LGBM.
• Within each tail and group, rows are selected by sufficient sample size, VaR coverage near 5%, then lower FZ loss and quantile loss.
• Full benchmark and LGBM per-model results are exported in full-result tables, so this figure is a readable summary rather than the full result set.

Figure: selected_model_performance_left_tail. Source: metrics/benchmark_metrics_per_model.parquet, metrics/ml_tail_metrics_per_model.parquet. Claim scope: selected_benchmark_vs_lgbm_main_figure_not_full_result_set. Tail side: left_tail. Run file: latex/figures/selected_model_performance_left_tail.png.

Figure: selected_model_performance_right_tail. Source: metrics/benchmark_metrics_per_model.parquet, metrics/ml_tail_metrics_per_model.parquet. Claim scope: selected_benchmark_vs_lgbm_main_figure_not_full_result_set. Tail side: right_tail. Run file: latex/figures/selected_model_performance_right_tail.png.

Figure 6. Full-Sample VaR Overlay Diagnostics

• Key readings: full-sample overlays show realized loss against a fixed benchmark-comparator VaR and the locked side-specific promoted ML-tail VaR.
• The benchmark line uses GJR-GARCH-EVT with GJR-GARCH-t fallback; the ML line is not selected by inspecting this plot.
• Treat the plot as a visual diagnostic. Formal validation remains the coverage, loss, DM, Murphy, and EVT evidence.

Figure: full_sample_var_overlay_left_tail. Source: forecasts/benchmark_forecasts.parquet, forecasts/ml_tail_forecasts.parquet. Claim scope: full_sample_var_overlay_fixed_selection_visual_diagnostic. Tail side: left_tail. Run file: latex/figures/full_sample_var_overlay_left_tail.png.

Figure: full_sample_var_overlay_right_tail. Source: forecasts/benchmark_forecasts.parquet, forecasts/ml_tail_forecasts.parquet. Claim scope: full_sample_var_overlay_fixed_selection_visual_diagnostic. Tail side: right_tail. Run file: latex/figures/full_sample_var_overlay_right_tail.png.

Figure 7. VaR/ES Stress-Window Overlays

• Supporting diagnostic: stress-window overlays illustrate threshold behavior in broad OOS stress episodes with left/right tails sharing each episode's x-axis.
• The LGBM overlays use information set C, the best-FZ row within the two 24-check LGBM+EVT families.
• They do not report hedge PnL, transaction-cost evidence, or trading performance.

Figure: var_es_stress_overlay_2024_stress_episode. Source: forecasts/benchmark_forecasts.parquet, forecasts/ml_tail_forecasts.parquet. Claim scope: appendix_stress_overlay_illustration_not_validation. Tail side: left_right_tail. Run file: latex/figures/var_es_stress_overlay_2024_stress_episode.png.

Figure: var_es_stress_overlay_2025_stress_episode. Source: forecasts/benchmark_forecasts.parquet, forecasts/ml_tail_forecasts.parquet. Claim scope: appendix_stress_overlay_illustration_not_validation. Tail side: left_right_tail. Run file: latex/figures/var_es_stress_overlay_2025_stress_episode.png.

Figure 8. DM Heatmaps

• Supporting diagnostic: heatmap cells report pairwise FZ-loss differences and one-sided DM p-values for the pass-all cross-suite model set.
• Rows are candidates, columns are anchors, and negative candidate-minus-anchor differences favor the row model.
• Each tail uses a strict global common sample across GJR-GARCH-EVT, LGBM plain MLE C, and LGBM UniBM C.

Figure: dm_heatmap_left_tail. Source: forecasts/benchmark_forecasts.parquet, forecasts/ml_tail_forecasts.parquet. Claim scope: post_24check_cross_suite_fz_dm_diagnostic. Tail side: left_tail. Run file: latex/figures/dm_heatmap_left_tail.png.

Figure: dm_heatmap_right_tail. Source: forecasts/benchmark_forecasts.parquet, forecasts/ml_tail_forecasts.parquet. Claim scope: post_24check_cross_suite_fz_dm_diagnostic. Tail side: right_tail. Run file: latex/figures/dm_heatmap_right_tail.png.

Figure 9. Benchmark Murphy Diagnostics

• Key readings: curves report target-history benchmark elementary-score diagnostics on a common grid.
• The plot is a scoring-family diagnostic, not a pairwise ranking statement.

Figure: benchmark_murphy_left_tail. Source: metrics/benchmark_murphy.parquet. Claim scope: murphy_diagnostic_benchmark_baseline_common_grid. Tail side: left_tail. Run file: latex/figures/benchmark_murphy_left_tail.png.

Figure: benchmark_murphy_right_tail. Source: metrics/benchmark_murphy.parquet. Claim scope: murphy_diagnostic_benchmark_baseline_common_grid. Tail side: right_tail. Run file: latex/figures/benchmark_murphy_right_tail.png.

Figure 10. 24-Check LGBM Murphy Diagnostics

• Key readings: curves report only the LGBM families that pass the full tail-by-information-set calibration screen.
• Interpret curve separation as scoring-family sensitivity evidence, not as a standalone model-selection rule.

Figure: lgbm_24check_murphy_left_tail. Source: metrics/lgbm_24check_murphy.parquet, metrics/ml_tail_metrics_per_model.parquet, forecasts/ml_tail_forecasts.parquet. Claim scope: murphy_diagnostic_lgbm_24check_robust_ladder. Tail side: left_tail. Run file: latex/figures/lgbm_24check_murphy_left_tail.png.

Figure: lgbm_24check_murphy_right_tail. Source: metrics/lgbm_24check_murphy.parquet, metrics/ml_tail_metrics_per_model.parquet, forecasts/ml_tail_forecasts.parquet. Claim scope: murphy_diagnostic_lgbm_24check_robust_ladder. Tail side: right_tail. Run file: latex/figures/lgbm_24check_murphy_right_tail.png.

Figure 11. ES Severity Diagnostics

• Key readings: bars report conditional-on-exception severity diagnostics.
• Severity is reported for risk interpretation but is not a standalone model-selection claim.

Figure: es_severity_left_tail. Source: metrics/benchmark_metrics.parquet, metrics/ml_tail_metrics.parquet, metrics/ml_tail_metrics_per_model.parquet. Claim scope: es_severity_diagnostic_not_model_selection_claim. Tail side: left_tail. Run file: latex/figures/es_severity_left_tail.png.

Figure: es_severity_right_tail. Source: metrics/benchmark_metrics.parquet, metrics/ml_tail_metrics.parquet, metrics/ml_tail_metrics_per_model.parquet. Claim scope: es_severity_diagnostic_not_model_selection_claim. Tail side: right_tail. Run file: latex/figures/es_severity_right_tail.png.

5.4 Source artifact index

Artifact	Path	Exists
manifest	reports/runs/tailrisk_20160719_20260522_20260527T083659Z_commit_7f628ff4/manifest.json	yes
data_vintage	reports/runs/tailrisk_20160719_20260522_20260527T083659Z_commit_7f628ff4/data_vintage.json	yes
modeling_panel	/Volumes/ExternalSSD/data/n225-open-gap-tail/gold/tp/tailrisk_20160719_20260522_20260527T083659Z_commit_7f628ff4/modeling_panel.parquet	yes
target_audit	/Volumes/ExternalSSD/data/n225-open-gap-tail/gold/tp/tailrisk_20160719_20260522_20260527T083659Z_commit_7f628ff4/target_audit.parquet	yes
calendar_map	/Volumes/ExternalSSD/data/n225-open-gap-tail/gold/tp/tailrisk_20160719_20260522_20260527T083659Z_commit_7f628ff4/calendar_map.parquet	yes
feature_coverage	/Volumes/ExternalSSD/data/n225-open-gap-tail/gold/tp/tailrisk_20160719_20260522_20260527T083659Z_commit_7f628ff4/feature_coverage.parquet	yes
leakage_summary	/Volumes/ExternalSSD/data/n225-open-gap-tail/gold/ls/tailrisk_20160719_20260522_20260527T083659Z_commit_7f628ff4/summary.json	yes
benchmark_status	reports/runs/tailrisk_20160719_20260522_20260527T083659Z_commit_7f628ff4/metrics/benchmark_status.json	yes
benchmark_metrics	reports/runs/tailrisk_20160719_20260522_20260527T083659Z_commit_7f628ff4/metrics/benchmark_metrics.parquet	yes
benchmark_metrics_per_model	reports/runs/tailrisk_20160719_20260522_20260527T083659Z_commit_7f628ff4/metrics/benchmark_metrics_per_model.parquet	yes
benchmark_forecasts	reports/runs/tailrisk_20160719_20260522_20260527T083659Z_commit_7f628ff4/forecasts/benchmark_forecasts.parquet	yes
benchmark_dm_inference	reports/runs/tailrisk_20160719_20260522_20260527T083659Z_commit_7f628ff4/metrics/benchmark_dm_inference.parquet	yes
ml_tail_status	reports/runs/tailrisk_20160719_20260522_20260527T083659Z_commit_7f628ff4/metrics/ml_tail_status.json	yes
ml_tail_metrics	reports/runs/tailrisk_20160719_20260522_20260527T083659Z_commit_7f628ff4/metrics/ml_tail_metrics.parquet	yes
ml_tail_metrics_per_model	reports/runs/tailrisk_20160719_20260522_20260527T083659Z_commit_7f628ff4/metrics/ml_tail_metrics_per_model.parquet	yes
ml_tail_forecasts	reports/runs/tailrisk_20160719_20260522_20260527T083659Z_commit_7f628ff4/forecasts/ml_tail_forecasts.parquet	yes
ml_tail_result_matrix	reports/runs/tailrisk_20160719_20260522_20260527T083659Z_commit_7f628ff4/metrics/ml_tail_result_matrix.parquet	yes
ml_tail_result_matrix_dm	reports/runs/tailrisk_20160719_20260522_20260527T083659Z_commit_7f628ff4/metrics/ml_tail_result_matrix_dm.parquet	yes
ml_tail_dm_inference	reports/runs/tailrisk_20160719_20260522_20260527T083659Z_commit_7f628ff4/metrics/ml_tail_dm_inference.parquet	yes
ml_tail_model_eviction	reports/runs/tailrisk_20160719_20260522_20260527T083659Z_commit_7f628ff4/metrics/ml_tail_model_eviction.parquet	yes
lgbm_24check_murphy	reports/runs/tailrisk_20160719_20260522_20260527T083659Z_commit_7f628ff4/metrics/lgbm_24check_murphy.parquet	yes
ml_tail_feature_unavailability	reports/runs/tailrisk_20160719_20260522_20260527T083659Z_commit_7f628ff4/metrics/ml_tail_feature_unavailability.parquet	yes
benchmark_stress_windows	reports/runs/tailrisk_20160719_20260522_20260527T083659Z_commit_7f628ff4/metrics/benchmark_stress_windows.parquet	yes
ml_tail_stress_windows	reports/runs/tailrisk_20160719_20260522_20260527T083659Z_commit_7f628ff4/metrics/ml_tail_stress_windows.parquet	yes
figure_manifest	reports/runs/tailrisk_20160719_20260522_20260527T083659Z_commit_7f628ff4/latex/figure_manifest.json	yes
table_manifest	reports/runs/tailrisk_20160719_20260522_20260527T083659Z_commit_7f628ff4/latex/table_manifest.json	yes
latex_dir	reports/runs/tailrisk_20160719_20260522_20260527T083659Z_commit_7f628ff4/latex/tables	yes
claim_scope_table	reports/runs/tailrisk_20160719_20260522_20260527T083659Z_commit_7f628ff4/latex/tables/tailrisk_claim_scope_table.tex	yes
es_severity_table	reports/runs/tailrisk_20160719_20260522_20260527T083659Z_commit_7f628ff4/latex/tables/tailrisk_es_severity_table.tex	yes
result_matrix_summary_table	reports/runs/tailrisk_20160719_20260522_20260527T083659Z_commit_7f628ff4/latex/tables/ml_tail_result_matrix_summary_table.tex	yes

• All paths above are local ignored artifacts; they are reproducible outputs, not tracked source files.
• Forecast/reporting rebuilds should read these artifacts and must not call vendor APIs.
• If this page is stale, rerun just snapshot after a completed just full or pass an explicit run id to the CLI snapshot command.