1.3.6 Advanced Extensions

Basic Prerequisites: 03 · Factor Construction and 05 · Failure Modes.

---

1. Intuition & Practical Objective

Every extension of the 3-factor model answers the same question: which systematic risks are priced, and what is the right factor universe? This page is the launchpad: it shows the three directions the field took - (1) the five-factor model (add profitability and investment, FF 2015), (2) statistical / latent factors (let PCA discover the factors from the covariance instead of from fundamentals, Tsay §9.4–9.5), and (3) the factor zoo / q-factor consolidation (hundreds of candidate characteristics collapsing to a handful of robust factors). It then hands off to the Fundamentals folder where the accounting characteristics live.

Why these three first? The 5-factor model is the natural FF family extension and is already linked from Quantitative Fundamental Investing. Statistical factors (PCA) are the opposite philosophy - data-driven rather than characteristic-driven - and reveal the hidden dimension structure. The factor zoo is the empirical endpoint: too many factors, so how many are real?

---

2. Mathematical Ground Truth & Derivations

2.1 The five-factor model (FF 2015, eq. 5)

Adding profitability and investment to the three-factor regression:

$$R_{it}-R_{ft}={\alpha }_i+b_{i}MK{T}_t+s_i{\text{SMB}}_t+h_i{\text{HML}}_t+r_i{\text{RMW}}_t+c_i{\text{CMA}}_t+{\epsilon }_{it}.$$

Constructed (as on 03 · Factor Construction) by 2×3 independent sorts on size × {B/M, OP, Inv}. The paper is rejected on the GRS test but is “an acceptable description of average returns for applied purposes”; its main residual failure is small, low-profitability, high-investment stocks (which earn lower average returns than the model predicts). FF 2015 stresses the result is not sensitive to the factor definitions - 2×3, 2×2, and 2×2×2×2 variants give the same conclusions. Note the dividend-discount-model logic: price is the PV of expected dividends, so high B/M (cheap), high profitability, and low investment all mechanically imply higher expected returns (FF2015 eq. 1–3).

2.2 Statistical factor models via PCA (Tsay §9.4–9.5; ESL Ch 14)

The orthogonal factor model (Tsay eq. 9.16):

$$r_t-\mu =\beta f_t+{\epsilon }_t,\text{Cov}(f)=I,\text{Cov}(\epsilon )=D,f\perp \epsilon ,$$

so that (Tsay eq. 9.17) ${\Sigma }_r=\beta {\beta }^{\prime }+D$. PCA finds the loadings from the eigen-decomposition of the covariance/correlation matrix: the principal components are $y_i=e_i^{\prime }r$ (eigenvectors of $\Sigma$), with $\text{Var}(y_i)={\lambda }_i$, and loadings $\beta =\sqrt{{\lambda }_j}{e}_j$ (Tsay eq. 9.19). The communality of asset $i$ is $c_i^2={\sum }_j{\beta }_{ij}^2$ and its unique/specific variance is ${\sigma }_i^2$, with $\text{Var}(r_{it})=c_i^2+{\sigma }_i^2$ (Tsay §9.5). The proportion of variance explained by the first $m$ PCs is ${\sum }_{j\le m}{\lambda }_j/{\sum }_j{\lambda }_j$ - the metric for choosing how many latent factors.

Choosing the number of factors: the scree/CPV rule (cumulative proportion of variance), Connor–Korajczyk (no significant drop in cross-sectional residual variance as $m\to m+1$), and the Bai–Ng information criteria $C_{p1},C_{p2}$ (Tsay §9.6.1). Example (Tsay): 40 stocks, $T=36$: CK picks $m=1$, Bai–Ng picks $m=6$, 6 factors explain ~89.4%.

2.3 The factor zoo and its consolidation

The zoo is the empirical failure of the small-factor premise: hundreds of published return-predictive characteristics (Green–Hand–Zhang count ~100, ~24 genuinely independent). Two consolidation directions:

• q-factor model (Hou, Xue & Zhang 2015): a production-based model whose four factors (market, size, investment, ROE) digest most anomalies through one economic lens - investment is the bridge to expected returns because a firm invests until marginal benefit equals discount rate.
• Statistical compression (PCA / APCA): let the data reveal the factor dimension structure rather than imposing characteristics. APCA (Connor–Korajczyk) works when $k>T$: eigen-analysis of the $T\times T$ inner product ${\hat{\Omega }}_T=(1/k)(R-1_T{\bar{r}}^{\prime })(R-1_T{\bar{r}}^{\prime }{)}^{\prime }$ (Tsay §9.6).

---

3. Computational Implementation - PCA statistical factor extraction

We reproduce the statistical-factor view on a 5-asset, 228-obs universe driven by two latent factors (market-like + tech-vs-financial), recover the eigenvalues, loadings, communalities, and the cumulative-proportion metric for factor-number choice. Stdlib only.

Two latent factors explain 91.2% of total variance. The loadings tell the economic story: PC1 loads positively on all five (the market component), while PC2 separates assets 0–2 (positive, “tech”) from assets 3–4 (negative, “financials”) - the industrial/rotation component, exactly the pattern Tsay documents for IBM/HPQ/INTC/JPM/BAC (2 PCs explain ~74% there). Communality 0.927 means the two factors capture 92.7% of asset 4’s variance. The eigenvalue drop (3.09, 1.47 vs 0.16, 0.15, 0.13) is the scree signal that $m=2$.

---

4. Failure Modes & First-Principles Breakdowns

1. The 5-factor model is rejected, not accepted. FF 2015 fails the GRS joint test and specifically misprices small, low-profitability, high-investment stocks. “Five factors” is a pragmatic description of average returns, not the true pricing kernel. Add factors and you approach the zoo; keep few and you misprice segments.
2. Statistical factors are rotation-invariant and unlabeled. PCA loadings are identified only up to orthogonal rotation (Tsay §9.5.2); the “market” and “tech-vs-financial” labels above are interpretations, not outputs. Varimax/quartimax rotation helps but does not guarantee economic meaning.
3. Latent factors ≠ traded factors. A statistical factor is a statistical object; it is not directly a tradeable long-short portfolio until you build a mimicking portfolio on it. Bridging latent structure to tradeable factors requires the construction step.
4. The factor zoo is a multiple-testing problem. With ~100 characteristics tested, spurious premiums are guaranteed by chance. The correct number of factors is an econometric choice (CK, Bai–Ng) AND a discipline choice (Harvey–Liu–Zhu’s $t>3$ hurdle, Deflated Sharpe) - see Backtesting Hygiene.

---

5. References

• Fama & French, “A Five-Factor Asset Pricing Model” (JFE, 2015)
• Tsay, Analysis of Financial Time Series
• Hastie et al., The Elements of Statistical Learning
• Hou, Kewei; Xue, Chen & Zhang, Lu, “Digesting Anomalies: An Investment Approach” (RFS, 2015)
• Green, Hand & Zhang, “The Characteristics That Provide Independent Information About Average U.S. Monthly Stock Returns” (RFS, 2017)
• Harvey, Liu & Zhu, ”… and the Cross-Section of Expected Returns” (RFS, 2016)

---

6. Connected Graph Bridges

• Back: 05 · Failure Modes · Index Hub
• Fundamentals (the characteristics): Quantitative Fundamental Investing · Value & Profitability · Quality & F-scores
• Statistical side: Linear Algebra & Matrices (PCA/SVD) · Machine Learning & Alt Data (tree-based factor ranking, high-dimensional selection)
• Multiple testing: Backtesting Hygiene & Deflated Sharpe
• Sibling: Momentum Factors