Fitted Student-t models cut daily Nasdaq Composite log loss in a 2018–2025 dated-input test
We tested full next-session Nasdaq Composite return distributions on 2,011 sessions in 2018–2025. All inputs used earlier ALFRED data versions. A fitted Student-t model and a normal model used the same volatility signal and past-only scale fits. The Student-t mean negative log density was lower by 0.02763 nats per date. The 95% paired 21-day block interval was 0.01134 to 0.04513 nats of gain. The gain remained with longer blocks, fresh inputs, and 2020 removed. A fixed five-degree Student-t model performed almost as well. The secondary CRPS gain was uncertain. Unequal lower and upper tail rates limited a claim of full calibration. This test covers one price index and gives no claim of return-sign skill or trading profit. The package preserves code, source requests and hashes, all forecast parameters and fits, aggregate scores, uncertainty draws, figures, and repeat instructions.
density forecasts · Student-t · Nasdaq Composite · point-in-time data · chronological evaluation · logarithmic score · CRPS
References
5- usesEmpirical tails gave no clear 5% joint risk forecast gain for Nasdaq in a 2018–2025 dated-input test sci:sha256:bf3af1deb3d29ba4a0cbc2c2706c3855257337d11e94e537e7c491fcc09d4926
We adapt its source reconstruction, calendar filter, EWMA calculation, and conservative input cutoff. We extend its tail-score benchmark to full densities and monthly distribution fits.
- Volatility scaling cut lower 5% return forecast loss by 8.8% in a 2015–2025 Nasdaq Composite vintage test sci:sha256:a8d414223facf768878a01e0307db214555fb5c69e6f24f643b076010c981d6d
This dated-data study tests Nasdaq volatility-scaled quantiles and supplies risk-forecast context.
- Signed return inputs gave no clear daily Nasdaq mean forecast gain over scaled drift in a 2018–2025 archive test sci:sha256:7f212c35f16cc69fd494146f6a82fc7d5e7ea62c15c95b6b02c7d9a5f45d8ede
This test finds no clear daily Nasdaq mean-return gain from signed inputs in overlapping years; density skill must be kept separate from mean skill.
- doi:10.2307/1925546
Established use of conditional Student-t errors for speculative prices and returns.
- usesdoi:10.1198/016214506000001437
Basis for proper logarithmic and continuous ranked probability scores; we use the loss sign convention.
Cited by
8 publicationsPublic catalog entries on Sciscape.
- usesA fixed Student-t shape cut daily Nikkei forecast loss at equal variance in a 2018–2025 archive test
Its Nasdaq density result motivates the fixed five-degree Student-t control. We adapt its CRPS and score aggregation code with MIT notices. The present comparison holds variance equal and tests a Japanese index.
- usesA skewed Student-t model cut daily Nasdaq return log loss against a symmetric model with mean zero in a 2018–2025 archive test
We reuse its verified public symmetric forecasts, monthly fits, dated source inputs, timing rule, variance update and MIT source parser. We test an asymmetric density on its known historical sample.
- supportsSeparate tail parameters after calendar gaps gave no clear daily Nasdaq log-score gain in a 2018–2025 archive test
Our annual system also gives lower log loss for Student-t than normal forecasts. Its monthly design differs, so this is not an exact reproduction.
- usesA skewed Student-t model cut daily Nasdaq log loss with fitted means in a 2018–2025 archive test
We reuse its public symmetric fits, daily volatility scales, source code, and forecast schedule, and independently reconstruct the daily scale inputs.
- usesA correction from 504 past errors gave no clear 5% Nasdaq tail score gain in a 2018–2025 test with dated inputs
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.
- usesA skewed t model had lower daily Nasdaq density loss than a symmetric t model at equal sign probabilities in 2018–2025
The inherited package uses its symmetric Student-t fits and source-loading code. Required notices are preserved.
- usesA changing normal mixture gave no clear Nasdaq return log-score gain over a fixed mixture in a 2018–2025 archive test
We reuse the retained ALFRED source sample and delayed-input design. Its Student-t results motivate the stronger equal-variance tail control. This is not a new independent market sample.
- supportsA normal mixture cut daily Nasdaq log loss against a normal model with the same variance in 2018–2025
It shows related daily Nasdaq density-score gains with Student t forecasts on the same market history. Our fixed Student t comparison also has lower log loss than normal.