ShawnChamberlain's picture
Lock in 2026-10-28 fully settled UTC and US close trading results on personal homepage
9f99d54 verified
|
Raw History Blame Contribute Delete
2.49 kB
# Status
**Last updated:** 2026-09-30 · **Stage:** results, robustness and two exploratory extensions complete; paper compiled (19 pp.)
## Done
- Loader for both JKP layouts with unit, duplicate and thin-leg checks; a data directory
may hold the site's zip directly.
- Seven point-in-time predictors, own historical mean, regional shrinkage mean; static or
time-varying (dated MSCI reclassifications, `config/market_history.yaml`) regions.
- Expanding-window annual refits with three training scopes (EM, DM-only, pooled);
OLS, elastic net, gradient-boosted trees, neural-net ensemble; rules: own mean,
shrinkage mean, factor momentum.
- 19 tests pass as of 2026-10-26 (loader, macro loader, leakage, market history, signal recovery,
null, transfer, real-time forecast shrinkage, published JKP result tables).
Size/power: 87.5% null rejection vs the own mean, 5.0% vs the shrinkage mean.
- **Main run** `reports/jkp_em/` (JKP all countries, 153 factors, monthly, capped VW;
downloaded 2026-09-29, data to 2025-12; 22 EM markets; incl. neural net).
- **Robustness** `reports/robustness/`: κ 60 / 240, OOS from 2005, equal-weighted and
uncapped VW factors, time-varying classification, cost 10 / 40 bp.
- Paper `paper/draft.pdf` compiled with tectonic (`.tools/`); all tables and the figure
generated by `scripts/paper_tables.py` and `scripts/robustness.py`.
- Break-even trading costs (`scripts/breakeven.py`): ≤ 16 bp on value-weighted factors,
30–48 bp on equal-weighted.
- Exploratory extensions (`make extensions`): real-time forecast shrinkage (`<model>_rs`)
and U.S. state variables (FRED; `--macro`). Neither overturns the headline.
- `config/market_history.yaml` cross-checked against MSCI's Previous Market
Reclassifications.xlsx: all events match.
- Negative and fragile results: `reports/failed_hypotheses.md`.
## Headline (details in README)
Shrinkage mean beats every learned model on capped VW factors; Clark–West positive
(trees 3.66–4.25) but forecasts 2.5–5× too large; DM-trained ≈ EM-trained directionally;
no learned timing portfolio beats static net of 20 bp; the shrinkage-mean tilt is best in
every specification. Equal-weighted factors are the exception (learned models beat the
shrinkage mean and static; not the tilt).
## Next
1. Size-dependent trading costs (do the equal-weighted gains survive?).
2. State variables that vary across emerging markets (local currency, capital flows)
instead of U.S. aggregates.