Statistical Factors vs Fundamental Factors (统计因子 vs 基本面因子)
Definition. Two complementary ways of building factor risk models: fundamental factors use economically motivated exposures (industry dummies, book-to-price, beta) constructed from company-level data; statistical factors use PCA / EM / autoencoders on the return covariance, with no economic priors.
The three families (from the article)
| Family | How loadings are built | How factor returns are built | Update frequency |
|---|---|---|---|
| Fundamental | Direct (industry, BP, liquidity, ...) | Cross-sectional regression → sample/EWMA covariance | Monthly/quarterly (Barra) |
| Macro | Time-series regression on observable macro returns (inflation, oil, rates) | The macro variable's return series | Quarterly |
| Statistical | PCA on sample covariance, or EM fit of low-rank-plus-diagonal | Latent; estimated jointly with loadings | Daily/weekly |
Tradeoffs
| Axis | Fundamental | Statistical |
|---|---|---|
| Interpretability | High (each factor has a name) | Low (PCs are linear combos) |
| Stability over time | High (loadings based on company attributes that change slowly) | Lower (loadings chase covariance regime) |
| Coverage of short-horizon themes | Weak (theme not in factor list) | Strong (any co-movement becomes a factor) |
| Risk-model use | Primary (Barra, Axioma) | Often secondary (PCA shrinkage) |
The Tzikas et al. (2026) synthesis
The paper's insight is that the two are not mutually exclusive. Keep the fundamental base model's loadings B for interpretability and stability, and augment it with a small number of transient statistical factors X that mop up whatever the fundamental set misses in the current market regime. This way you get the stability of fundamental + the responsiveness of statistical, with a clean structural separation: B stays fixed, X is re-learned.
Operational implication
- Fundamental B: stable, slow update → business-as-usual risk reporting.
- Transient X: fast update (EWMA, daily) → real-time risk adjustments during regime shifts.
If you want post-hoc interpretability for the transient factors, the paper suggests two routes:
- Cross-sectional correlation of X's columns with known themes (AI, energy, geopolitics, value/growth, etc.).
- Use an LLM to summarize the assets with the largest X loadings.
Connections
- Risk Model Enhancement via Transient Statistical Factors — the method that explicitly combines the two.
- Barra risk model — the canonical fundamental base.
- PCA for covariance estimation — the classical statistical-factor approach.
- Factor Themes (13 因子主题分类) — 另一条切分线: 不问因子从哪来, 而问 153 个特征该归并成几组 (Jensen-Kelly-Pedersen 2023 的 1-\rho + Ward 层次聚类)。
Reference
Tzikas et al. (2026), see summaries/202606-mle-transient-factor-mining/202606-mle-transient-factor-mining.