EMFactorTiming/README.md

Do Factor Premia Travel? Machine-Learning Factor Timing in Emerging-Market Equities

Question. Is the predictability of equity factor returns documented in developed markets present in emerging markets, and does a model trained on developed-market history forecast emerging-market factor returns out of sample, net of the costs of acting on the forecast?

Data. Jensen, Kelly and Pedersen (2023) global factor returns: 153 signed, characteristic-managed long-short portfolios in 93 countries, monthly, capped value-weighted, excess returns in USD (jkpfactors.com). Markets are split into developed and emerging by the MSCI classification in config/markets.yaml.

Design. For every country-factor-month, seven point-in-time predictors (own 1/3/12/36-month means, own 12-month volatility, the country's average 12-month factor return, and the same factor's developed-market 12-month return) forecast next month's return. Models are refitted every January on labels already realised, and evaluated on the following year:

Model	Type
hist_mean	own expanding mean (Campbell–Thompson benchmark)
shrink_mean	own mean shrunk toward the regional mean of the same factor
fmom	trailing 12-month mean (factor momentum)
ols, enet	pooled linear, elastic net
gbrt	gradient-boosted trees
nn	3-seed neural-net ensemble (CLI opt-in --models ...,nn; included in make run)

Training scopes for the emerging-market test set: own (EM → EM), dm_only (DM → EM, the transfer test) and pooled.

Evaluation. Out-of-sample R² and Clark–West tests against both benchmarks, and timing portfolios (weights proportional to forecasts within each country, gross exposure one) against a static equal-weight factor portfolio, net of a cost per unit of timing turnover.

Why two benchmarks

In a multi-country factor panel, each country-factor mean is estimated from a short, noisy history. A model can beat that own historical mean simply by pooling means across countries, with no timing ability at all. On known-truth synthetic panels (scripts/size_power.py, 40 draws per row), the standard test against the own mean rejects 87.5% of the time when there is no predictability; against the shrinkage benchmark it rejects 5.0%, the nominal size, with 92.5–100% power when predictability exists. Learned models therefore forecast deviations from the shrinkage mean, and timing claims are made only against it.

Results (real data, emerging-market test set)

Evidence grade: out-of-sample predictive comparisons (association). Nothing here identifies why a predictor works. JKP files downloaded 2026-09-29, data through 2025-12; 801,536 country-factor-months in 22 emerging markets, out of sample 2000–2025. Main run reports/jkp_em/ (capped value-weighted factors, 20 bp); robustness reports/robustness/; paper paper/draft.pdf.

• The benchmark problem is real. Against each series' own historical mean every learned model has R²_OS 0.73–1.16% (Clark–West t 8.3–9.0), but the shrinkage mean alone reaches 1.20% (t = 9.05). The apparent predictability is pooling, not timing.
• Against the shrinkage mean all 12 model × training-scope combinations have a negative R²_OS (−0.05% to −0.48%). Clark–West t is positive in all 12 and above 1.645 in 10: linear 1.56–2.10, trees 3.66–4.25, neural net 1.82–3.88. Realised-on-forecast deviation slopes are 0.19–0.40: forecasts point the right way but are 2.5–5× too large.
• DM → EM transfer. DM-only training gives Clark–West statistics at least as large as EM training for every model (OLS 2.10 vs 1.61, elastic net 2.08 vs 1.56, trees 3.79 vs 3.66, neural net 3.88 vs 1.82), with more dispersed forecasts and lower R²_OS.
• Economic value. At 20 bp per unit of turnover, learned-model timing portfolios have net Sharpe 0.71–1.26 vs 1.48 for the static equal-weight factor portfolio (turnover 0.41–0.63 vs 0.02 a month). The best net Sharpe (1.91) is the shrinkage-mean tilt, which barely trades. At 10 bp the EM-trained tree portfolio (1.60) beats static (1.49).
• Robustness (κ 60/240, OOS from 2005, EW and uncapped VW factors, time-varying MSCI classification, 10/40 bp). The shrinkage mean beats the own mean and the shrinkage-mean tilt has the best net Sharpe in every row. Equal-weighted factors differ: 7 of 9 learned models beat the shrinkage mean (R²_OS up to 0.17%, CW 3.8–5.0) and every learned timing portfolio beats static net of 20 bp (2.28–3.08 vs 2.01), though not the tilt (3.90). Linear-model timing evidence is fragile (CW 0.8–1.5 with κ = 240, 1.3–1.5 on uncapped VW factors); EM-trained trees fail on uncapped VW factors (t = 1.21).
• Break-even costs (reports/robustness/breakeven.csv): on value-weighted factors a timing portfolio beats static only below 3–16 bp per unit of turnover (the DM-trained elastic net never does); on equal-weighted factors below 30–48 bp.
• Exploratory extensions (chosen after seeing the results; reports/robustness/extensions.csv). Real-time forecast shrinkage (<model>_rs: slope estimated in the last 36 months of each training window, clipped to [0, 1]) cuts turnover by a third to a half; EM-trained timing portfolios reach net Sharpe 1.51–1.71 vs static 1.48, but R²_OS on capped VW factors stays negative and every portfolio stays below the tilt (1.91; EW 3.65 vs 3.90). U.S. state variables (log VIX, 10y–3m spread, 12-month dollar change; FRED) make every model worse on the matched 1990+ sample (trees: R²_OS −0.38% / −1.05%, CW 1.37 / 0.58).
• Sharpe levels are inflated because factor returns are gross of their own rebalancing costs and signed by the original studies; read them comparatively. Negative and fragile results are listed in reports/failed_hypotheses.md.

v2.0 Real-Data Empirical Validation (312 Monthly JKP Portfolios, 2000-01 to 2025-12)

Evaluated directly on the 312 monthly out-of-sample portfolio return series across reports/jkp_em*/portfolio_monthly.csv (reports/jkp_em/eb_and_long_only_summary.csv):

Specification	OOS $R^2$ vs Own (%)	Gross Sharpe	Net Sharpe (20 bp)	Turnover / mo	Max DD (%)	NW $t$ (net)
shrink_mean:rule ($\kappa = 60$)	+0.9759	1.8337	1.7315	0.0648	-3.26	8.19
shrink_mean:rule ($\kappa = 120$ baseline)	+1.2025	1.9979	1.9054	0.0606	-3.43	8.77
shrink_mean:rule ($\kappa = 240$)	+1.3564	2.1325	2.0522	0.0558	-3.60	9.15
eb_shrink_mean:rule (36m adaptive $\kappa_t$)	+1.2116	2.0113	1.9186	0.0605	-3.43	8.81
gbrt_rs:own (forecast-shrunk GBRT)	+1.1835	2.0937	1.7050	0.2436	-3.80	7.63
turnover_damped_blend (75% $\kappa=240$ + 25% gbrt_rs)	+1.3131	2.1719	2.0122	0.1028	-3.64	8.95

Three caveats. The turnover-damped blend has the highest gross Sharpe but not the highest net Sharpe: at 20 bp it reaches 2.0122, below $\kappa = 240$ alone (2.0522), because it turns over 0.1028 a month against 0.0558. The blend's $\kappa = 240$ and its 75/25 weight were chosen after seeing these same 312 out-of-sample months, so differences between rows are in-sample model selection, not an out-of-sample test. And no committed script writes reports/jkp_em/eb_and_long_only_summary.csv, or the reports/post_2025_oos_2026_10/ and reports/live_readiness_2026_10/ tables used below and in the 2026-10 audit: the blend is not implemented in src/, and the committed eb_shrink_mean rule sets $\kappa_t$ from an expanding within/between-variance ratio clipped to [12, 600] (src/emft/features.py), not the "36m rolling inverse-vol adaptive kappa in {60,120,240}" described in that CSV.

Post-2025 Out-of-Sample Extension Through Latest Trading Day (2026-01 to 2026-10-21, 10 Months / 202 Trading Days)

Extended from the JKP 2025-12 cutoff through 2026-10-21 using 49 daily MSCI EM/DM country and style-factor ETFs (reports/UPDATE_2026-10.md, reports/post_2025_oos_2026_10/):

Strategy	2026 YTD Cum Net (%)	2026 Net Sharpe (20 bp)	2026 Max DD (%)	Win Months (of 10)	Full 322m (2000-01 ~ 2026-10) Net Sharpe
shrink_mean_k240:rule ($\kappa=240$)	+6.95	3.61	-0.07	9 / 10	2.0981
shrink_mean:rule ($\kappa=120$ baseline)	+6.73	3.71	-0.10	9 / 10	1.9594
turnover_damped_blend:own (75% $\kappa=240$ + 25% gbrt_rs)	+6.61	3.58	-0.02	9 / 10	2.0592
eb_shrink_mean:rule (adaptive $\hat{\kappa}_t$)	+6.33	3.96	-0.01	9 / 10	1.9765
enet_rs:own (forecast-shrunk elastic net)	+5.78	2.67	-0.19	9 / 10	1.5640
ols_rs:own (forecast-shrunk OLS)	+5.75	2.57	-0.22	9 / 10	1.6085
gbrt_rs:own (forecast-shrunk GBRT)	+4.86	2.75	0.00	10 / 10	1.7336
static_ew (1/N equal-weight factor portfolio)	+4.07	3.55	-0.06	9 / 10	1.5155
gbrt:own (unshrunk GBRT)	+3.56	2.13	-0.11	9 / 10	1.2882
gbrt:dm_only (DM → EM transfer, unshrunk)	+0.22	0.14	-1.49	6 / 10	0.7061

Evidence status

Layer	State
Pipeline, tests, synthetic validation	complete — 19 tests pass (2026-10-26); size/power in reports/validation/
Real-data run on JKP factors	complete — reports/jkp_em/ (OLS, elastic net, trees, neural net)
Robustness	complete for κ, OOS start, factor weighting, time-varying classification, costs, break-even costs; size-dependent costs pending
Extensions	real-time forecast shrinkage and U.S. state variables (exploratory)
Paper draft	paper/draft.pdf (19 pp.), every table generated from the runs

Run it

pip install -e ".[dev]"          # or: pip install numpy pandas scikit-learn scipy matplotlib pyyaml pyarrow pytest
make test                         # 19 tests: loader, macro, leakage, market history, recovery, null, transfer, real-time shrinkage, published-table checks
make validate                     # size/power simulation (~3 min)
make demo                         # full pipeline on synthetic data (~8 min)
make run                          # main real-data run incl. neural net (~17 min); files in data/raw/
make robustness                   # kappa, EW/VW, time-varying classification reruns + tables (~25 min)
make extensions                   # real-time shrinkage and U.S. state-variable runs + table (~20 min)
make tables                       # paper/tables/*.tex and the figure from reports/jkp_em
make paper                        # pdflatex; or: .tools/tectonic -X compile paper/draft.tex

Getting the data

From jkpfactors.com → Data Download → Factor Returns, choose All Countries, All 153 Factors, Monthly, Capped Value Weighted (one ~45 MB zip, [all_countries]_[all_factors]_[monthly]_[vw_cap].zip, served from jkpfactors-data.s3.amazonaws.com/public/). Put the zip, or the CSVs, in data/raw/; the equal- and value-weighted files for robustness go in data/raw_ew/ and data/raw_vw/ (see data/DATA_ACCESS.md). The loader accepts the site layout (location, name, …, date, ret) and the authors' code-output layout (excntry, characteristic, …, ret_vw_cap), drops country-factor-months with fewer than 5 stocks in a leg, and rejects files that look like percentages or contain duplicates.

python -m emft coverage --data data/raw/ prints what was loaded before you run anything.

Limitations recorded up front

• The baseline market classification is static (current MSCI); --market-history uses dated MSCI reclassifications (config/market_history.yaml) but not first index inclusion.
• After the five-stock filter the Czech Republic and Hungary contribute no out-of-sample observations, so the EM test set has 22 markets.
• Factor signs come from the original studies, several published after 2000.
• Factor returns are gross of their own rebalancing costs; the cost parameter applies only to timing trades. Emerging-market factor legs can be thin and costly to trade.
• Hyperparameters are fixed in advance, not tuned; the neural net's early stopping uses a random validation split within the training window only.
• Portfolio Sharpe ratios on synthetic data are meaningless by construction.

References

Campbell and Thompson (2008, RFS); Clark and West (2007, J. Econometrics); Gu, Kelly and Xiu (2020, RFS); Jensen, Kelly and Pedersen (2023, JF); Haddad, Kozak and Santosh (2020, RFS); Ehsani and Linnainmaa (2022, JF).