{"status":"ok","message-type":"work","message-version":"1.0.0","message":{"indexed":{"date-parts":[[2026,10,9]],"date-time":"2026-10-09T17:50:58Z","timestamp":1791568258561,"version":"4.3.4"},"posted":{"date-parts":[[2026]]},"group-title":"SSRN","reference-count":0,"publisher":"Elsevier BV","license":[{"start":{"date-parts":[[2026,1,1]],"date-time":"2026-01-01T00:00:00Z","timestamp":1767225600000},"content-version":"unspecified","delay-in-days":0,"URL":"https:\/\/www.uspto.gov\/ip-policy\/copyright-policy\/copyright-basics"}],"content-domain":{"domain":[],"crossmark-restriction":false},"short-container-title":[],"abstract":"<jats:p>This study develops a global probabilistic forecasting framework for forecasting the specific conditional distributions of individual assets in multi-asset realized variance. Using data from 16 assets, we replace the parametric output of the DeepAR architecture, which assumes a prespecified distribution, with direct conditional-quantile forecasts. We also introduce a variance-state-dependent weighting scheme into the existing quantile-based continuous ranked probability score (CRPS) loss. Specifically, we adopt the standard deviation of realized variance over the previous 22 trading days (SD22) as a stability measure of the recent variance state and construct observation-specific loss weights. Observations with higher SD22, corresponding to relatively unstable variance states, are assigned greater loss weights during model training, encouraging the model to emphasize observations associated with unstable variance conditions. Based on this approach, we develop the Volatility-State-Weighted Quantile-DeepAR (VS-Q-DeepAR) model, and train it globally on observations from all assets using shared network parameters and asset-specific embeddings, thereby exploiting cross-asset information while preserving asset-level heterogeneity. Importantly, the model separately forecasts the conditional distribution of each asset rather than modeling the joint distribution across assets. Empirical results show that, in one-step-ahead forecasting of log realized variance, the proposed global quantile-based forecasting framework performs comparably to or slightly better than highly competitive asset-specific HAR-type benchmark models. Ablation experiments and sensitivity analysis of the weighting parameter show that the overall CRPS is minimized at \u03b7 = 0, indicating that volatility-state weighting does not necessarily improve overall distributional forecasting accuracy. In contrast, moderate weighting can improve the forecasting performance of selected upper-tail quantiles by placing greater emphasis on observations associated with unstable volatility states. Finally, we extend the proposed framework to conditional return-distribution forecasting and value-at-risk (VaR) analysis, where state-dependent weighting yields more pronounced improvements in tail-risk forecasting.<\/jats:p>","DOI":"10.2139\/ssrn.7590090","type":"posted-content","created":{"date-parts":[[2026,10,9]],"date-time":"2026-10-09T16:57:36Z","timestamp":1791565056000},"source":"Crossref","is-referenced-by-count":0,"title":["Global Deep Learning for Asset-Specific Distributional\u00a0Forecasting of Realized Variance"],"prefix":"10.2139","author":[{"given":"Jianlong","family":"Bao","sequence":"first","affiliation":[],"role":[{"vocabulary":"crossref","role":"author"}]},{"ORCID":"https:\/\/orcid.org\/0000-0001-8035-3276","authenticated-orcid":true,"given":"Takayuki","family":"Morimoto","sequence":"additional","affiliation":[],"role":[{"vocabulary":"crossref","role":"author"}]}],"member":"78","container-title":[],"original-title":[],"deposited":{"date-parts":[[2026,10,9]],"date-time":"2026-10-09T16:57:36Z","timestamp":1791565056000},"score":1,"resource":{"primary":{"URL":"https:\/\/www.ssrn.com\/abstract=7590090"}},"subtitle":[],"short-title":[],"issued":{"date-parts":[[2026]]},"references-count":0,"URL":"https:\/\/doi.org\/10.2139\/ssrn.7590090","relation":{},"subject":[],"published":{"date-parts":[[2026]]},"subtype":"preprint"}}