Future Work¶

This page records extensions that should remain outside the current paper unless they become necessary for interpretation. The present paper is a point-in-time out-of-sample forecast evaluation of OSE Nikkei 225 Futures opening-gap VaR and Expected Shortfall. It already covers benchmark models, ML tail models, left-tail and right-tail risk surfaces, coverage diagnostics, loss-based comparison, CPA regressions, Murphy diagrams, DST diagnostics, ES severity, and risk-trigger summaries.

Future work should therefore do one of two things: either sharpen the economic interpretation of the current evidence, or define a clearly separate paper.

Extension Summary¶

Extension	Research question	When to pursue
Intraday U.S.-close Nikkei reference mark	Does U.S. close information predict the opening-auction residual after conditioning on a Nikkei futures price observed at the U.S. cash close?	After licensed OSE, CME, SGX, or equivalent intraday Nikkei futures marks are available with reliable timestamps.
EVT and high-quantile evaluation	Does the LightGBM POT-GPD layer improve VaR-ES performance beyond direct LightGBM quantile forecasts, especially at more extreme quantiles?	After the 95% common-sample comparison is stable, and before any 97.5% claim is made.
Richer option-implied tail-risk predictors	Do option-implied measures beyond VIX close add incremental information for opening-gap tail risk?	After the core U.S. close information sets are settled and richer option data can be timestamped.
Night-session microstructure and absorption	How much U.S. information is incorporated during the OSE night session, and what remains for the day-session open?	After night-session OHLC or intraday OSE data are available with session-level timestamps.
Real-time pre-open risk monitoring	Can the historical forecast design be converted into a live pre-open risk-monitoring process?	Only after live feeds, vendor availability, failure handling, and alert logs are specified.
Submission reproducibility package	Can every table and figure be reproduced from documented commands and source manifests?	Before manuscript circulation.

1. Intraday U.S.-Close Nikkei Reference Mark¶

Question¶

Does U.S. close information predict residual OSE day-session opening risk after conditioning on a Nikkei futures price observed at the U.S. cash close?

Data Requirements¶

OSE intraday Nikkei 225 Futures prices, or a licensed equivalent.
CME or SGX Nikkei futures marks if they provide a reliable U.S.-close reference.
Source-specific timestamps and data-availability records.

Guardrails¶

Do not synthesize a U.S.-close Nikkei mark from daily data.
Do not use a mark observed after the forecast origin.
Keep the U.S.-close-mark target unavailable until a licensed intraday source exists.
Keep full settlement-to-open risk separate from U.S.-close-mark-to-open residual risk.

2. EVT and High-Quantile Evaluation¶

Question¶

Does the LightGBM standardized-loss POT-GPD layer improve VaR-ES performance relative to direct LightGBM quantile forecasts?

First Comparison¶

Compare 95% direct LightGBM quantile forecasts against 95% LightGBM POT-GPD forecasts.
Use common out-of-sample dates.
Evaluate VaR coverage, quantile loss, FZ joint VaR-ES loss, ES exceedance severity, and exception counts together.

Conditions for 97.5% Results¶

97.5% VaR/ES results should be promoted only if the evidence supports them:

sufficient common-sample observations;
sufficient out-of-sample exceptions;
stable GPD shape and scale estimates;
threshold-sensitivity checks;
interpretable ES severity diagnostics;
clear separation between left-tail and right-tail results.

Computing a 97.5% forecast is not, by itself, enough to make a 97.5% empirical claim.

3. Richer Option-Implied Tail-Risk Predictors¶

Question¶

Do option-implied measures beyond the VIX close add information about OSE opening-gap tail risk beyond ETFs, rates, FX, and the existing volatility-index controls?

Candidate Predictors¶

VVIX.
SKEW or related option-implied tail measures.
VIX futures term structure.
SPX option-implied tail or variance-risk-premium measures, subject to data access.

Guardrails¶

Record the observation timestamp and the practical availability time for each predictor.
Treat delayed historical series as historical predictors, not live pre-open data.
Test incremental value through pre-specified information-set additions or ablations.
Do not rely on feature-importance narratives without out-of-sample loss and coverage evidence.

4. Night-Session Microstructure and Absorption¶

Question¶

How much of the U.S. close signal is incorporated during the OSE night session, and what part remains for the next day-session open?

Candidate Measures¶

Night-session return.
Night-session range.
Last-hour night-session movement.
Night-session volume and liquidity measures.
Opening-auction residual gap after conditioning on the night-session close.

Guardrails¶

Do not interpret full settlement-to-open predictability as opening-auction residual predictability.
Keep settlement-to-open, close-to-open, night-close-to-open, and U.S.-close-mark-to-open targets in separate tables.
Report holiday, roll-window, SQ-window, and early-close sensitivity.

5. Real-Time Pre-Open Risk Monitoring¶

Question¶

Can the historical forecast design be converted into a live pre-open risk-monitoring process?

Required Additions¶

Live or near-live U.S. close predictors.
A live OSE, CME, SGX, or broker feed for Nikkei futures reference prices.
Vendor-availability timestamps and missing-data rules.
Alert logs, failure handling, and monitoring procedures.

Guardrails¶

Do not describe the historical backtest as production-ready.
Do not rely on J-Quants daily futures OHLC as a live pre-open feed.
Keep risk alerts separate from execution, transaction-cost, and trading-alpha claims.

6. Submission Reproducibility Package¶

Before manuscript circulation, prepare a reproducibility package with:

data-source manifest and access notes;
source as-of dates and hashes where permitted;
schema reports for raw, interim, and processed tables;
target audit report;
point-in-time feature checklist;
model configuration files;
table and figure reproduction commands;
smoke fixtures for reviewers without vendor data.

Submission criterion: every manuscript table and figure should map to one documented command and one output path.