Factor Investing & Factor Timing

Fri, 24 Apr 2026 00:00:00 +0000

Navigating the factor zoo

The factor zoo is large, the cross-section is noisy, and predictive signals decay fast once they are published. Building portfolios from factor forecasts is as much a problem of managing estimation error as it is of finding genuine premia, and the signals that survive out of sample tend to be those that respect the structure of international markets rather than treating every country as an independent laboratory.

This research agenda develops factor-selection and factor-timing methods that take two facts seriously. First, estimation error is the dominant source of risk in mean-variance portfolios constructed from factor forecasts — so regularisation, shrinkage, and clustering are first-order concerns, not afterthoughts. Second, international factor dynamics are structured by geography and market integration: local, regional, and global signals do not substitute for one another, and ignoring the non-local dimension discards information that is priced in the cross-section.

The work connects cross-sectional pricing, the time-series predictability of factor premia, and the machine-learning toolkit that handles the associated dimensionality. It is deliberately methodological — signal construction, bias correction, and loss-function design rather than the addition of yet another factor to the zoo.

Cross-country factor momentum

(with Pedro Barroso and Merlin Bartel) documents that a factor performing better in one country than in others tends to continue outperforming in the following months, producing positive returns that are not spanned by leading factor models. Specialised style mutual funds respond to these signals, but with delay — a limits-to-arbitrage pattern more pronounced in markets with higher information frictions.

Complementing this, (with Merlin Bartel and Joshua Traut) shows that regional and global factor-momentum signals dominate local ones in forecasting factor risk premia — overturning the common presumption that local factors are pricing-superior. The outperformance is strongest in regions with high structural integration, and non-local signals revive momentum investing in markets previously thought to lack it, such as Japan.

Factor-timing and the time-series dimension

uses multivariate distance measures computed within each factor’s own cross-section to predict the time-series dynamics of factor premia, beating a battery of standard predictors for six of seven Fama-French factors. This is also a bridge to the Uncertainty & Ambiguity agenda, where parameter uncertainty is itself the signal. (with Lars Kaiser, Review of Financial Economics 2021) extends cross-sectional signal construction to skewness and kurtosis, showing that higher moments add statistically and economically significant predictive power beyond cross-sectional volatility.

Estimation error and regularisation

(with Merlin Bartel) clusters assets into equally weighted sub-portfolios that feed into classical minimum-variance optimisation. The optimal cluster count approximates N/4 — a structural bias-variance sweet spot that beats both plug-in minimum variance and 1/N. This connects directly to the Risk-Adjusted Forecast Performance pillar via the Sharpe-ratio gap framework, where the same estimation-error logic governs how forecasts translate into realised portfolio performance.

Infrastructure: ffdownload and rqmoms

Two R packages support this research line. ffdownload ( , ) provides programmatic access to the full Fama-French data library — factors, portfolios, industry classifications — for fully reproducible empirical work. rqmoms ( ) implements quantile-based higher-moment estimators used in the higher-moments line of work.

Ongoing work

Active extensions include factor-momentum concentration and its interaction with limits to arbitrage, factor-zoo selection under explicit estimation-error penalties, and machine-learning-based factor timing with economically-aligned loss functions — the latter cross-referencing the Risk-Adjusted Forecast Performance pillar.

Collaborators

Pedro Barroso (Católica Lisbon), Merlin Bartel, Lukas Salcher (University of Liechtenstein), Joshua Traut (St. Gallen / SBF), Lars Kaiser (formerly University of Liechtenstein).

Higher Moments | Sebastian Stöckl