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)

Question: can past factor returns, or machine-learning forecasts from them, time 153 published equity factors well enough to beat an equal-weight factor portfolio after trading costs, in data that the factor-momentum papers did not use? Approach: we used JKP Global Factor Data (capped value weight, monthly, 2026-04 vintage). Before we loaded returns after 2018, we froze a protocol. A factor can enter the universe only after its publication year. We tested 11 rules: 12-month time-series and cross-sectional factor momentum, winners only, winners minus losers, 1-month, volatility-scaled, historical-mean sign, and pooled ridge and LightGBM forecasts. The baseline is equal-weight buy-and-hold. Costs are 10 bp for each dollar of stock traded, with 0 to 50 bp scenarios. The holdout is 2019-01 to 2025-12. After the first holdout, we froze an addendum before we downloaded country data. The addendum pools the same test over 43 non-US countries. Results: no pre-registered timing hypothesis was supported (Holm p >= 0.43 in the US, 0.56 in the pool). The US net alpha of 12-month factor momentum on buy-and-hold was 0.18% per month (t 1.24). This agrees with a parallel 23-factor SCI study. In the country pool, the net alpha was 0.04% per month (95% CI -0.08 to 0.15), and the minimum detectable effect was 0.16%. But the pre-registered winners-minus-losers gross spread was 0.31% per month (t 3.18, Holm p 0.002) and positive in 40 of 43 countries. Factor returns thus continued, but trading removed the gain. ML forecasts ranked factors with skill (holdout IC 0.13, t 3.6 ex-US), but high turnover gave significantly negative net alphas in the pool (LightGBM long-only -0.12% per month, t -4.6). A look-ahead universe of all factors made the 1972-2018 US spread larger (0.34% vs 0.23% for published factors only). Limitations: costs are scenarios, internal factor rebalancing and shorting costs are not modelled, 84 months, one data vintage. The package has code, frozen protocols, hashes, aggregate tables and figures. Forecasts, weights and monthly returns are in a retained dataset.

factor momentum · factor timing · out-of-sample test · pre-registration · JKP global factor data · trading costs · publication bias · international equity · LightGBM · holdout