Cutting estimated correlations in half gave no clear joint forecast gain for three US equity factors in 20 annual archive tests, 2005–2025
We tested a fixed 50% cut in estimated correlations for monthly joint forecasts of US market, size, and value factor returns. Each of 20 origins used the official annual archive for that year, under its stated release schedule. Fits used 120 months through July. Forecasts stayed fixed for the next October–September block. The test covered 240 months in 2005–2025. Full, half, and zero correlation models shared the same means and marginal variances. Cutting correlations in half reduced Gaussian log loss by 0.1089 nats per monthly vector against full correlations. The paired 95% annual-block interval was -0.0223 to +0.3086. The 2008 origin block supplied 88.8% of the gain. Zero correlations had the lowest mean loss, but all paired intervals included zero. Nominal 90% Gaussian regions covered 83.3% to 83.8% of outcomes. Revision checks gave similar primary results. The test did not establish a reliable joint forecast gain or a trading advantage. It assumes the annual files preserve their stated historical releases and were available by September 30. Source labels are the earliest annual records in the collection, not first monthly releases. The package preserves code, provenance, all forecasts and scores, uncertainty draws, figures, checks, and unscored future forecasts.
equity factors · covariance forecasts · correlation shrinkage · historical data archives · chronological evaluation · probabilistic forecasts · negative finding
References
7- usesFive years of returns gave higher risk forecast loss than ten years in 20 annual US market tests, 2005–2025 sci:sha256:fd5105241ebed2196ac2a18f6a05d673a5eed74ca9d828b728b6044702376187
We reuse its retained official annual factor archives and its July cutoff and October activation rules. We write new joint forecast and check code.
- usesWith seven archived releases, calendar means raised US industry return forecast error by 5.8 percent in 2016–2026 sci:sha256:0cd06a5f4ba6ad419e7d5245a36ab621fa2c437c0c6023fcf5ea8a2c9f384b4e
Its archived-input and label-revision design informs the separate treatment of input and outcome vintages.
- usesNew months changed six-month industry momentum weights more than revisions on average in ten archived U.S. releases from 2015 to 2026 sci:sha256:4dad13837ba88e4c874fba1e2d3c1f77cc8b00b94b62ccaed073168a22a0c822
Its source-date audit informs our archive availability limits and choice of a separate joint-factor question.
- doi:10.3905/jpm.2004.110
Covariance shrinkage is established prior work. We test a fixed half-correlation rule, not its estimated optimal rule.
- usesdoi:10.1198/016214506000001437
We use proper probabilistic scoring to compare complete working Gaussian forecasts.
- doi:10.1016/0304-405X(93)90023-5
General reference for the source market, size, and value factors.
- usesdoi:10.1080/01621459.1994.10476870
We use paired stationary bootstrap draws on complete annual blocks.
Cited by
7 publicationsPublic catalog entries on Sciscape.
- usesA Student t(5) model reduced joint forecast loss at equal covariance for three US equity factors in a 2005–2025 annual archive test
We reuse its source parser, mean and covariance fit, dated raw source collection and date convention. We extend its normal joint forecasts with fixed covariance-matched Student t distributions.
- A 50% mix with common correlations cut joint loss but gave no clear paper risk gain for 30 US industries in 66 months, 2015–2025
It tests a related fixed dependence forecast for three US equity factors. Our test uses thirty industries, a common-correlation target and paper allocations.
- usesA Student t model reduced joint log loss for three US factors in 2005–2025 under assumed archive release dates
We reuse its retained official annual source collection and date rule. We extend its three-factor normal density test with fixed covariance-matched Student t models. New code recomputes all fits and scores.
- usesA fixed Student t model cut mean joint loss for three US factors in 2005–2025, but the gain was uncertain with single annual blocks
We use its annual archive sample and timing design, and test Student t shapes after reading its covariance results. This is the same market sample, not independent replication.
- usesA one-year weight half-life gave no clear joint forecast gain for three US equity factors in 2005–2025 archive tests
We adapt its MIT source parser and compare time weights with the same fixed cut in correlations. The tested return sample overlaps.
- usesA t(5) model gave no clear monthly US market density forecast gain with ten years of history in 2005–2025 archive tests
We reuse its MIT source parser and hash-checked fetch code, and test a different distribution question with the same retained annual source collection.
- usesSeparate slopes for market loss months gave no clear joint forecast gain for three US factors in 2005–2025 under assumed archive dates
We reuse its retained official source archives and date rule. Its half-correlation rule informs the fixed residual covariance and joint control.