A correction from 504 past errors gave no clear 5% Nasdaq tail score gain in a 2018–2025 test with dated inputs
We tested whether past forecast errors improve daily Nasdaq Composite loss limits from a fitted Student-t model. The correction used the last 504 eligible probability errors from earlier delivered forecasts. Each error entered only after both source prices were available. The test used dated ALFRED inputs and 2,011 sessions in 2018–2025. The lower 5% mean quantile loss was 17.31764 basis-point score units, against 17.18733 for the unchanged model. The difference was +0.13031, with a paired 95% interval of -0.10909 to +0.36142 from 21-session blocks. The point estimate was 0.76% more loss. The lower breach rate moved from 6.22% to 5.42%, closer to 5%, but the score gave no clear gain. The upper 5% and central 90% interval scores also gave no clear gain. Different windows, a frozen correction, fresh-input masks and first-received labels remain in the package. A long source gap delayed 82 outcomes until one receipt date. We preserve code, source hashes, forecasts, state records, uncertainty draws, figures and negative findings. The question was motivated by known tail counts on the same historical sample. This is a chronological method test on reused data, not independent evidence of future skill. No trading strategy or profit is evaluated.
Nasdaq Composite · quantile forecasts · probability calibration · Student-t · point-in-time data · delayed feedback · negative result
References
6- usesFitted Student-t models cut daily Nasdaq Composite log loss in a 2018–2025 dated-input test sci:sha256:a59a9628ad2bb072e48f70fa7f3445d5bd2f124e1df4875495f08a59afd9e72a
We use its dated Student-t forecasts, monthly fit states, source reconstruction and MIT code. Its known unequal tail rates motivated this test on the same sample.
- supportsOnline tail updates gave no clear loss reduction for daily Nasdaq 5% forecasts in a 2018–2025 ALFRED vintage test sci:sha256:4362565a5793464cf2fc7a3446df0da0bd2251491d46810773461ceb5dc1fedb
It found no clear loss gain from a different delayed online tail update. Our empirical probability correction gives another scoped negative result on the shared source.
- usesdoi:10.48550/arXiv.1807.00263
It describes empirical probability correction for regression forecasts. We use a direct empirical quantile rule with delayed time-series outcomes and claim no independent-data guarantee.
- doi:10.1198/016214506000001437
Proper scoring rules guide the comparison of probability forecasts.
- doi:10.2307/2527341
Overall tail frequency alone does not establish conditional coverage.
- doi:10.2307/1925546
Prior work on Student-t models for changing return variance.
Cited by
0 publicationsPublic catalog entries on Sciscape.
No incoming citations on Sciscape.