Five published US portfolio return signals did not pass a Holm-corrected holdout test after their original samples to August 2026, and industry seasonality changed to a net loss
Question: do published monthly return-prediction rules that use only public US portfolio returns still earn a net return in the months after each paper's sample ended? Approach: we rebuilt four published rules and one fixed composite from the Kenneth French Data Library (CRSP build 202608) and AQR factors. The rules are cross-industry OLS post-LASSO (Rapach et al. 2019), industry momentum (Moskowitz and Grinblatt 1999), time-series factor momentum with 13 US factors (Ehsani and Linnainmaa 2022), industry same-calendar-month seasonality (Keloharju et al. 2016), and a composite of three of them. Each holdout starts in the month after the original sample and ends in August 2026. We fixed the protocol before we calculated holdout statistics. Signals use only earlier returns; costs are 10 bp one way per unit traded; intervals come from a stationary block bootstrap. Results: no rule passed the one-sided Holm-corrected test of net mean above zero (smallest adjusted p = 0.097). The LASSO rule replicated in 1970-2016 (6.7% a year gross, t = 4.7) but earned -4.8% net in 2017-2026 (95% CI -11.0 to 1.3), and its forecasts had an out-of-sample R2 of -6.1% against the prevailing mean. Industry momentum earned 5.6% net in 1995-2026 (CI 0.3 to 10.8), but its FF5+UMD alpha was 1.5% (t = 0.7). Factor momentum earned 1.0% net in 2016-2026 (CI -2.0 to 4.1), about 40% of its earlier mean. Industry seasonality earned 7.6% gross in 1947-2011 and -9.6% net in 2012-2026 (CI -15.2 to -4.2). Its retention ratio was -0.76 (CI -1.64 to -0.06). Without the 2026 months, its gross mean was -3.6% (t = -1.5). The composite earned -2.7% net. A comparison with an archived 2017 vintage showed that 78% of the industry-month returns since 1960 have been revised; the LASSO strategy returns on the two vintages correlate only 0.88. Limitations: three holdouts are only 116-128 months long, so minimum detectable effects are 3.8-11 percentage points a year; our LASSO rule is a reconstruction; portfolio returns are gross of their internal trading costs. Code, monthly strategy returns, every LASSO forecast, data hashes and the protocol are included.
return predictability · out-of-sample · post-publication decay · industry momentum · factor momentum · return seasonality · LASSO · replication · Fama-French data · data vintages · trading costs · holdout test
References
12- challengesdoi:10.3905/jfds.2019.1.3.009
Rule C1 reconstructs its OLS post-LASSO industry rotation; the holdout after its 2016 sample end gave a negative net return.
- doi:10.1111/0022-1082.00146
[Legacy relation: reproduces] Rule C2 rebuilds its 6-month industry momentum strategy on 30 French industries and tests it after July 1995.
- doi:10.1111/jofi.13131
[Legacy relation: reproduces] Rule C3 rebuilds its time-series factor momentum with 13 of its US factors and tests it after December 2015.
- challengesdoi:10.1111/jofi.12398
Rule C4 tests its same-calendar-month seasonality on industries after December 2011; the premium changed to a loss.
- doi:10.1016/j.jfineco.2007.02.003
Origin of the same-calendar-month seasonality signal.
- doi:10.1111/jofi.12365
Prior evidence of post-sample and post-publication decline in anomaly returns.
- doi:10.1111/fire.12300
Prior reexamination of factor momentum.
- usesdoi:10.1016/j.jeconom.2006.05.023
Clark-West test for nested forecast comparison.
- usesdoi:10.1080/01621459.1994.10476870
Stationary bootstrap for confidence intervals.
- usesdoi:10.1016/j.jfineco.2013.10.005
BAB factor (AQR data) is one of the C3 factors.
- usesdoi:10.1007/s11142-018-9470-2
QMJ factor (AQR data) is one of the C3 factors.
- usesdoi:10.1016/j.jfineco.2014.10.010
FF5 factors for alpha regressions and C3.
Cited by
15 publicationsPublic catalog entries on Sciscape.
- VIX from the prior session reduced 22-session S&P 500 variance forecast loss in a 2016–2025 test of 2026 data snapshots
Its data-version check informed our source records and stated limit. This study uses different data and a variance target.
- usesWith seven archived releases, calendar means raised US industry return forecast error by 5.8 percent in 2016–2026
We use its retained source files and provenance. Its revised-history seasonality loss and revision check motivate this separate dated-input replay.
- usesFixed industry momentum ranks gave no clear forecast gain across 66 US monthly outcomes in a 2015–2025 archive test
We reuse retained source files first described by this study. Its published return-signal tests give prior context.
- Prior-day large-stock returns stopped predicting next-day US small-stock index returns after 2000 and gave no tradable ETF forecast gain in 2006–2026
Related SCI holdout test of published monthly US portfolio return signals.
- In 2012–2026 the announcement-day beta premium in 35 US portfolios was +5.3 bp a day (95% CI −12 to +24); this neither confirms nor rejects the 1978–2011 value of +11.6 bp
Related SCI holdout test of published US portfolio return signals with French data.
- Past VIX values reduced weekly Nasdaq Composite variance forecast loss in 504 tests from 2016 to 2025
This SCI study reports weak monthly return forecast results and source revisions. It gives context for the use of dated archives here.
- usesNew months changed six-month industry momentum weights more than revisions on average in ten archived U.S. releases from 2015 to 2026
We reuse its retained industry source files and reference build, and extend its revision check by separating revisions from additional available months at archived capture times.
- usesVXN cut QLIKE loss for monthly Nasdaq-100 daily-return variance forecasts by 34% in a 2017–2026 vintage test
Its report of revised public portfolio returns motivates explicit input and outcome vintages. No data or forecasts are reused.
- usesArchived seasonal industry forecasts lost 10.9% a year after assumed costs across five 12-month tests
We reuse its retained source files and manifest. We extend its data-version check with forecasts fixed from archived files before each test.
- Signed return inputs gave no clear daily Nasdaq mean forecast gain over scaled drift in a 2018–2025 archive test
- In 2019–2025, winner equity factors beat loser factors in 43 non-US countries, but 12-month factor momentum gave no net gain over equal weights at 10 bp costs (153 JKP factors)
Related holdout test of published US portfolio return signals, including factor momentum on 13 US factors.
- Recent losses gave no clear added daily forecast gain in fixed Nasdaq Composite models in a 2018–2025 ALFRED test
Motivates chronological tests, simple return baselines, and explicit data-version records.
- Two seasonal means gave no clear US market return forecast gain in 240 months in 2005–2025 under assumed archive dates
Prior tests of published portfolio signals informed the question and revision warning.
- Adding prior overnight gaps raised ridge error for next-session open-to-close returns in SPY, QQQ and IWM in 2016–2025
Its catalog summary reports a chronological test of monthly portfolio signals and data revision risks.
- usesFive years of returns gave higher risk forecast loss than ten years in 20 annual US market tests, 2005–2025
Its results and data-revision check motivated the use of archived releases in the main test. No code was reused.