# Chapter 17: Appendix: Reference Material _Previous: [Chapter 16: Practical Considerations](16-practical-considerations.md)_ --- ## 17.1 Notation summary | Symbol | Dimensions | Meaning | First used | | -------------------------------- | --------------------------- | ------------------------------------------------------------------------------------------------------------------------------ | -------------------------------------------------------------------------- | | $N$ | scalar | number of assets | [Ch. 1](01-introduction.md) | | $K$ | scalar | number of factors | [Ch. 1](01-introduction.md) | | $T$ | scalar | number of time periods | [Ch. 1](01-introduction.md) | | $r$ | $N \times 1$ | asset returns over one period | [Ch. 2](02-the-factor-model-equation.md) | | $X$ | $N \times K$ | factor exposure (loading) matrix | [Ch. 2](02-the-factor-model-equation.md) | | $f$ | $K \times 1$ | factor returns over one period | [Ch. 2](02-the-factor-model-equation.md) | | $\epsilon$ | $N \times 1$ | specific (idiosyncratic) returns | [Ch. 2](02-the-factor-model-equation.md) | | $F$ | $K \times K$ | factor covariance matrix | [Ch. 2](02-the-factor-model-equation.md) | | $\Delta$ | $N \times N$ diag. | specific variance matrix, $\Delta_{ii} = \sigma^2_{\epsilon_i}$ | [Ch. 2](02-the-factor-model-equation.md) | | $\Sigma$ | $N \times N$ | asset covariance, $\Sigma = XFX^\top + \Delta$ | [Ch. 2](02-the-factor-model-equation.md) | | $w, w_p, w_b, w_a$ | $N \times 1$ | weights: generic, portfolio, benchmark, active ($w_p - w_b$) | [Ch. 2](02-the-factor-model-equation.md), [Ch. 9](09-risk-attribution.md) | | $x, x_p, x_b, x_a$ | $K \times 1$ | factor exposures of a portfolio, $x = X^\top w$ | [Ch. 2](02-the-factor-model-equation.md), [Ch. 9](09-risk-attribution.md) | | $\sigma_p, \sigma_a$ | scalar | portfolio volatility, tracking error | [Ch. 2](02-the-factor-model-equation.md) | | $d_{ik}$ | scalar | raw descriptor value, stock $i$, factor $k$ | [Ch. 3](03-factors-and-exposures.md) | | $\mu_k, \sigma_k$ | scalar | standardization location (cap-weighted) and scale (equal-weighted) | [Ch. 3](03-factors-and-exposures.md) | | $S(f)$ | scalar | least-squares objective, (weighted) sum of squared residuals minimized to get $\hat f$ | [Ch. 6](06-estimating-factor-returns.md) | | $W$ | $N \times N$ diag. | regression weights (convention: $\propto \sqrt{\text{cap}}$) | [Ch. 6](06-estimating-factor-returns.md) | | $C$ | $1 \times K$ (or more rows) | constraint matrix, $Cf = 0$ | [Ch. 6](06-estimating-factor-returns.md) | | $R$ | $K \times (K - c)$ | restriction matrix, feasible $f = Rg$ | [Ch. 6](06-estimating-factor-returns.md) | | $P$ | $K \times N$ | pure factor portfolio matrix, $\hat f = Pr$ | [Ch. 6](06-estimating-factor-returns.md), [Ch. 7](07-factor-portfolios.md) | | $H$ | $N \times N$ | OLS projection ("hat") matrix, $H = X(X^\top X)^{-1}X^\top$ | §17.2 | | $\Omega$ | $N \times N$ | general specific-return covariance in GLS; reduces to $\Delta$ when diagonal | §17.2 | | $\beta, \beta_{p,h}$ | scalar | time-series beta to a factor; portfolio beta to a hedge instrument | [Ch. 1](01-introduction.md), [Ch. 12](12-hedging.md) | | $\gamma$ | scalar / $K \times 1$ | Fama–MacBeth risk premium; Lagrange multiplier in the characteristic-portfolio derivation | [Ch. 7](07-factor-portfolios.md) | | $h$ | varies | hedge notionals, characteristic portfolio weights | [Ch. 7](07-factor-portfolios.md), [Ch. 12](12-hedging.md) | | $\lambda$ | scalar | risk aversion ([Ch. 11](11-portfolio-construction.md)), EWMA decay ([Ch. 8](08-risk-model-assembly.md)), context disambiguates | [Ch. 8](08-risk-model-assembly.md), [Ch. 11](11-portfolio-construction.md) | | $b$ | scalar | bias statistic, $\mathrm{std}(r_t/\hat\sigma_{t-1})$ | [Ch. 8](08-risk-model-assembly.md), [Ch. 14](14-model-evaluation.md) | | $\mathrm{MCR}_i, \mathrm{CTR}_i$ | scalar | marginal contribution $(\Sigma w)_i/\sigma$, contribution $w_i \cdot \mathrm{MCR}_i$ | [Ch. 9](09-risk-attribution.md) | Conventions: returns are arithmetic and in decimal unless a table is marked %. Risk numbers are annualized unless noted. "Exposure" always means a column-standardized or dummy loading per [Chapter 3](03-factors-and-exposures.md). ## 17.2 Refresher: the least-squares family in one place **OLS:** $\min_f (r - Xf)^\top(r - Xf) \Rightarrow \hat f = (X^\top X)^{-1} X^\top r$. Fitted values $X\hat f = Hr$ with the projection ("hat") matrix $H = X(X^\top X)^{-1}X^\top$: symmetric, idempotent ($H^2 = H$), projecting onto the column space of $X$. Residuals $(I - H)r$ are orthogonal to that space: $X^\top \hat\epsilon = 0$. **WLS:** With positive diagonal $W$: $\min_f (r - Xf)^\top W (r - Xf) \Rightarrow \hat f = (X^\top W X)^{-1} X^\top W r$. Equivalent to OLS on $\tilde r = W^{1/2} r$, $\tilde X = W^{1/2} X$. Best linear unbiased when $W \propto \mathrm{Cov}(\epsilon)^{-1}$ (Aitken/GLS). The $\sqrt{\text{cap}}$ convention approximates this for equities ([Ch. 6](06-estimating-factor-returns.md)). **GLS:** General $\mathrm{Cov}(\epsilon) = \Omega$: $\hat f = (X^\top \Omega^{-1} X)^{-1} X^\top \Omega^{-1} r$. In the factor-model context $\Omega = \Delta$ (diagonal), so GLS = WLS with $W = \Delta^{-1}$. **Covariance algebra used throughout:** For conformable constant matrices $A, B$ and random vectors $u, v$: $\mathrm{Cov}(Au) = A\,\mathrm{Cov}(u)\,A^\top$; $\mathrm{Cov}(Au + Bv) = A\,\mathrm{Cov}(u)A^\top + B\,\mathrm{Cov}(v)B^\top$ when $\mathrm{Cov}(u,v) = 0$. These two lines, applied to $r = Xf + \epsilon$, _are_ the derivation of $\Sigma = XFX^\top + \Delta$. **EWMA:** Weights $\lambda^s$ on lag $s$. Half-life $h \leftrightarrow \lambda = 2^{-1/h}$. Recursive update $\hat F_t = \lambda \hat F_{t-1} + (1-\lambda)\tilde f_t \tilde f_t^\top$ (normalized form). Effective sample size $\approx (1+\lambda)/(1-\lambda) \approx 2.89h$ for large $h$. **Eigendecomposition/PCA:** Symmetric $\Sigma = Q\Lambda Q^\top$, $Q$ orthonormal, $\Lambda$ diagonal with $\lambda_1 \ge \dots \ge \lambda_N \ge 0$ for PSD. PCA: factor $j$'s exposures = $q_j$, factor variance = $\lambda_j$. The rank-$K$ truncation is the best rank-$K$ approximation in Frobenius norm (Eckart–Young). [Marchenko–Pastur](https://en.wikipedia.org/wiki/Marchenko%E2%80%93Pastur_distribution) noise edge for aspect ratio $N/T$: $\lambda_{\pm} = \sigma^2 (1 \pm \sqrt{N/T})^2$. ## 17.3 Derivation collection **D1 - Constrained WLS ([Ch. 6](06-estimating-factor-returns.md)):** Minimize $S(f) = (r - Xf)^\top W (r - Xf)$ s.t. $Cf = 0$ ($c$ independent rows). _Restriction form:_ pick $R$ ($K \times (K-c)$) whose columns span the null space of $C$ (so $CR = 0$ and any feasible $f = Rg$). Substitute: $S(Rg)$ is unconstrained in $g$, and by the WLS formula with design $XR$: $\hat g = (R^\top X^\top W X R)^{-1} R^\top X^\top W r$, $\hat f = R\hat g$. _Lagrangian form:_ $\mathcal{L} = S(f) + 2\lambda^\top C f$. Stationarity gives the bordered system $\begin{pmatrix} X^\top W X & C^\top \\ C & 0\end{pmatrix}\begin{pmatrix}\hat f\\ \lambda\end{pmatrix} = \begin{pmatrix}X^\top W r\\ 0\end{pmatrix}$. Same solution, and the multiplier $\lambda$ prices the constraint. **D2 - Pure factor portfolios, $PX = I$ ([Ch. 7](07-factor-portfolios.md)):** Unconstrained: $P = (X^\top W X)^{-1}X^\top W \Rightarrow PX = (X^\top W X)^{-1}(X^\top W X) = I_K$. Row $k$ of $P$ is a portfolio with exposure vector = row $k$ of $PX$ = $e_k^\top$: unit own-factor, zero others. Constrained: $P = R(R^\top A R)^{-1} R^\top A$ with $A = X^\top W X$. Then $PXRg = Rg$ for all $g$, identity on the feasible subspace. Style rows are exactly pure, market/industry rows carry the constraint's structure ([Ch. 7](07-factor-portfolios.md) table). **D3 - Euler risk decomposition ([Ch. 9](09-risk-attribution.md)):** $\sigma(w) = (w^\top \Sigma w)^{1/2}$ is positively homogeneous of degree 1. Euler's theorem for homogeneous functions: $\sigma = \sum_i w_i \partial\sigma/\partial w_i$. Directly: $\partial \sigma/\partial w = \Sigma w / \sigma$, so $\sum_i w_i (\Sigma w)_i / \sigma = w^\top \Sigma w / \sigma = \sigma$. ∎ Factor-space version: with $\sigma(x) = (x^\top F x)^{1/2}$ (factor block), contributions $x_k (Fx)_k$ sum to $x^\top F x$. **D4 - Characteristic portfolio ([Ch. 7](07-factor-portfolios.md)):** $\min_h h^\top \Sigma h$ s.t. $x^\top h = 1$. Lagrangian $h^\top \Sigma h - 2\gamma(x^\top h - 1)$. Stationarity $\Sigma h = \gamma x \Rightarrow h = \gamma \Sigma^{-1} x$. The constraint fixes $\gamma = 1/(x^\top \Sigma^{-1} x)$: $h_x = \Sigma^{-1}x / (x^\top \Sigma^{-1} x)$, with minimized variance $1/(x^\top \Sigma^{-1}x)$. The GLS pure portfolio for factor $k$ solves the same problem with the _added_ constraints of zero exposure to the other factors. Stacking those constraints and applying D1's bordered system shows the GLS ($W = \Delta^{-1}$) regression rows solve it: estimation efficiency = portfolio efficiency. **D5 - Woodbury identity for $\Sigma^{-1}$ ([Ch. 11](11-portfolio-construction.md)):** For $\Sigma = \Delta + XFX^\top$ with $\Delta, F$ invertible: $\Sigma^{-1} = \Delta^{-1} - \Delta^{-1}X(F^{-1} + X^\top \Delta^{-1} X)^{-1} X^\top \Delta^{-1}$. Verify by multiplication: $\Sigma \cdot [\text{RHS}] = I + X F X^\top \Delta^{-1} - (X + XFX^\top\Delta^{-1}X)(F^{-1} + X^\top \Delta^{-1}X)^{-1}X^\top \Delta^{-1}$. Factor $XF$ from the middle term: $X + XFX^\top \Delta^{-1} X = XF(F^{-1} + X^\top \Delta^{-1} X)$, so the middle term collapses to $XFX^\top \Delta^{-1}$, cancelling. ∎ Cost: $O(NK^2 + K^3)$ vs. $O(N^3)$. **D6 - Minimum-variance hedge ratio ([Ch. 12](12-hedging.md)):** $\mathrm{Var}(r_p + h\, r_h) = \sigma_p^2 + 2h\,\mathrm{Cov}(r_p, r_h) + h^2 \sigma_h^2$. Minimize over $h$: $h^* = -\mathrm{Cov}(r_p, r_h)/\sigma_h^2 = -\beta_{p,h}$. Multi-instrument: $h^* = -\mathrm{Cov}(r_H)^{-1}\mathrm{Cov}(r_H, r_p)$, the population regression coefficients of the portfolio on the instruments. Through the model, $\mathrm{Cov}(r_H) = X_h F X_h^\top + \Delta_h$ and $\mathrm{Cov}(r_H, r_p) = X_h F x_p (+ \text{specific overlap})$. ## 17.4 Glossary - **Active return/risk:** portfolio minus benchmark return. Volatility thereof (tracking error). - **Alpha:** expected return not explained by factor exposures. In construction, the forecast vector fed to an optimizer. - **Bias statistic:** std of realized returns standardized by forecast volatility. The calibration score of a risk model. - **Characteristic portfolio:** minimum-variance portfolio with unit exposure to a characteristic: $\Sigma^{-1}x / x^\top\Sigma^{-1}x$. - **Coverage universe:** all assets the model assigns exposures and risk to (cf. estimation universe). - **Descriptor:** a raw measurable per-stock quantity (B/P, 12-1 return) before standardization. Factors blend one or more. - **Estimation universe:** the curated asset set on which factor returns are estimated. - **Exposure (loading):** a stock's sensitivity to a factor: dummy (industry/country) or standardized z-score (style). - **Factor-mimicking/pure factor portfolio:** the long–short portfolio (row of $P$) whose return is the estimated factor return. Unit own-exposure, zero other-exposure. - **Factor return:** per-period payoff to unit exposure of a factor, estimated by cross-sectional regression. - **Half-life:** lag at which an EWMA weight halves. The responsiveness dial of covariance estimation. - **Idiosyncratic/specific risk:** return variance unique to a stock. Diagonal of $\Delta$. Diversifiable. - **Information coefficient (IC):** cross-sectional correlation between a signal and forward returns. In alpha research, the number that matters is the IC of the factor-residualized signal. - **Linked assets:** multiple listings of one issuer (ADR, share classes) sharing factor exposures and correlated specifics. - **Pure factor portfolio:** see factor-mimicking portfolio. - **Restriction matrix:** basis of a constraint's null space, converting constrained to unconstrained regression. - **Style factor:** continuous characteristic-based factor (value, momentum, size, quality…). - **Tracking error:** annualized volatility of active return. - **VIF (variance inflation factor):** $1/(1-R^2)$ of one exposure regressed on the others. The redundancy gauge for candidate factors. - **Winsorization:** clipping extreme descriptor values before standardization. - **Z-score:** standardized exposure: (descriptor − cap-weighted mean) / equal-weighted std. ## 17.5 The mini example: complete dataset Everything below reproduces every number in Chapters [2](02-the-factor-model-equation.md)–[15](15-modifying-the-model.md) (NumPy, deterministic, no randomness). The full script is reproduced on its own page: [Mini Example Source Code](18-mini-example-source-code.md). **Universe, descriptors, month-1 data, portfolios:** | Stock | Industry | Cap $bn | B/P | Mom (12-1) | Spec vol (ann) | $r_1$ (%) | $w_p$ | | --------- | -------- | ------: | ---: | ---------: | -------------: | --------: | ----: | | AXIOM | Tech | 150 | 0.15 | +0.32 | 18% | +4.2 | 0.10 | | BINARY | Tech | 80 | 0.25 | +0.18 | 22% | +2.8 | 0.08 | | CIPHER | Tech | 40 | 0.45 | −0.05 | 30% | +0.5 | 0.10 | | DIGIT | Tech | 10 | 0.60 | +0.40 | 38% | +6.0 | 0.03 | | EVERGREEN | Fin | 120 | 0.85 | +0.06 | 16% | +0.8 | 0.22 | | FIDELIS | Fin | 60 | 0.95 | −0.02 | 20% | −0.6 | 0.14 | | GUARDIAN | Fin | 20 | 1.10 | −0.12 | 28% | −1.8 | 0.06 | | HARVEST | Cons | 90 | 0.40 | +0.10 | 17% | +1.2 | 0.15 | | INDIGO | Cons | 30 | 0.55 | +0.02 | 26% | +2.0 | 0.08 | | JUNIPER | Cons | 15 | 0.70 | −0.08 | 32% | −0.5 | 0.04 | Benchmark $w_b$ = cap weights (total cap 615). SIZE descriptor = ln(cap). Style standardization: cap-weighted mean, equal-weighted std ([Ch. 3](03-factors-and-exposures.md). Resulting $X$ tabulated there). Regression weights $\propto \sqrt{\text{cap}}$, normalized. Constraint: cap-weighted industry factor returns sum to zero, industry cap weights (0.4553, 0.3252, 0.2195). **Factor covariance $F$ (annualized):** Volatilities: MKT 16%, TECH 9%, FIN 7%, CONS 5%, VALUE 4%, MOM 6%, SIZE 4%. Correlations: | | MKT | TECH | FIN | CONS | VALUE | MOM | SIZE | | ----- | --: | ---: | ----: | ----: | ----: | ----: | ----: | | MKT | 1 | 0.10 | −0.05 | −0.10 | −0.20 | 0.05 | 0.15 | | TECH | | 1 | −0.40 | −0.30 | −0.35 | 0.30 | 0.05 | | FIN | | | 1 | −0.10 | 0.40 | −0.15 | 0.00 | | CONS | | | | 1 | 0.05 | −0.05 | −0.05 | | VALUE | | | | | 1 | −0.45 | 0.10 | | MOM | | | | | | 1 | 0.05 | | SIZE | | | | | | | 1 | (Symmetric, eigenvalues all positive: PSD verified in the script.) **Stipulated later-month factor returns** ([Ch. 10](10-performance-attribution.md). CONS set by the constraint): $f_2$ = (−2.0, −1.5, +1.0, +1.63, +1.2, −0.8, +0.5)%. $f_3$ = (+3.0, +0.8, −0.5, −0.919, −0.6, +1.0, −0.3)%. Active specific returns months 2–3: +0.30%, −0.10%. Constant exposures assumed across the quarter. **Hedge instruments** ([Ch. 12](12-hedging.md)): index future = benchmark exposures (1, 0.4553, 0.3252, 0.2195, 0, 0, 0). Small-cap future = (1.05, 0.35, 0.30, 0.35, 0.10, −0.05, −1.20). Both specific-risk-free. **Alpha-research candidate** ([Ch. 13](13-alpha-research.md)): raw descriptor = book-to-price − 0.30·(12-1 momentum) + a fixed per-stock idiosyncratic term, standardized to exposure $a$. Regressing $a$ on $X$ gives the spanning results below. **Key computed results (cross-chapter checkpoints):** month-1 factor returns $f_1$ = (1.821, 0.768, −1.282, 0.306, 0.548, 1.962, 0.046)%. Portfolio/benchmark/active vols 17.55% / 18.14% / 5.42%. Active exposures (0, −0.145, 0.095, 0.051, 0.385, −0.332, −0.275). Quarter attribution VALUE +0.46% MOM −0.72% specific +0.24% total −0.102%. Optimization TE 5.42->4.65% at 25.2% turnover. Hedge (−0.759, −0.229) giving 17.55->8.14%. Alpha-research candidate: corr with VALUE/MOM/SIZE +0.82/−0.44/−0.50, spanned fraction 0.884, IC vs $r_1$ −0.49 raw / −0.13 residualized, signal long–short −0.63% = factor −0.46% + specific −0.17%. ## 17.6 Annotated bibliography **Foundations** - **Sharpe (1964)**, "Capital Asset Prices," _JF_: the one-factor beginning. Beta as the first exposure. - **Ross (1976)**, "The Arbitrage Theory of Capital Asset Pricing," _JET_: multi-factor pricing without specifying the factors. The license under which all factor models operate. - **Fama & MacBeth (1973)**, "Risk, Return, and Equilibrium," _JPE_: the two-pass cross-sectional methodology ([Ch. 7](07-factor-portfolios.md)). Still the standard premia test. - **Chen, Roll & Ross (1986)**, "Economic Forces and the Stock Market," _JB_: the canonical macroeconomic factor model ([Ch. 4](04-model-types.md)). - **Fama & French (1992, 1993)** _JF/JFE_; **(2015)** _JFE_; **Carhart (1997)** _JF_: the style-factor canon: size and value, the three-factor model, profitability/investment, momentum. The sorted-portfolio construction of [Ch. 7](07-factor-portfolios.md). - **Rosenberg (1974)**, "Extra-Market Components of Covariance," _JFQA_: the founding paper of the fundamental cross-sectional architecture this primer centers on. **Books** - **Grinold & Kahn**, _Active Portfolio Management_ (2nd ed.): the practitioner bible: characteristic portfolios, the fundamental law, IR-based thinking. The source of [Ch. 7](07-factor-portfolios.md)'s optimization view and [Ch. 11](11-portfolio-construction.md)'s alpha discipline. - **Connor, Goldberg & Korajczyk**, _Portfolio Risk Analysis_: the most rigorous book-length treatment of factor risk models per se, all three families. - **Qian, Hua & Sorensen**, _Quantitative Equity Portfolio Management_: cross-sectional modeling and construction with worked detail. - **Litterman et al.**, _Modern Investment Management_: risk decomposition and budgeting culture (Goldman's quantitative tradition). **Methods** - **Ledoit & Wolf (2003, 2004):** shrinkage covariance estimation ([Ch. 8](08-risk-model-assembly.md)): "Honey, I Shrunk the Sample Covariance Matrix." - **Newey & West (1987)**, _Econometrica_: autocorrelation-consistent covariance. The horizon-scaling fix of [Ch. 8](08-risk-model-assembly.md). - **Shanken (1992)**, _RFS_: errors-in-variables correction for Fama–MacBeth ([Ch. 7](07-factor-portfolios.md)). - **Menchero (2000s, various):** multi-period attribution linking ([Ch. 10](10-performance-attribution.md)), **Carino (1999):** the log-linking algorithm. - **Black & Litterman (1992)**, _FAJ_: equilibrium-anchored expected returns ([Ch. 11](11-portfolio-construction.md)). - **Michaud (1989)**, "The Markowitz Optimization Enigma," _FAJ_: error maximization named and shamed ([Ch. 11](11-portfolio-construction.md)). - **Harvey, Liu & Zhu (2016)**, _RFS_, "...and the Cross-Section of Expected Returns": the factor zoo's multiple-testing reckoning ([Ch. 16](16-practical-considerations.md)). **Hou, Xue & Zhang (2020)**, _RFS_: the replication audit. - **Kelly, Pruitt & Su (2019)**, _JFE_: IPCA, **Gu, Kelly & Xiu (2020)**, _RFS_: ML asset pricing ([Ch. 16](16-practical-considerations.md) directions). **Practitioner references** - **MSCI Barra model handbooks** (USE4, GEM3 and successors): full disclosure of a production fundamental model: descriptor recipes, estimation universes, regression weights, specific-risk blending. The single best way to see every choice in Chapters [3](03-factors-and-exposures.md)–[8](08-risk-model-assembly.md) made concretely, with parameters. - **Axioma/SimCorp research papers:** practical treatments of alpha alignment ([Ch. 11](11-portfolio-construction.md)), statistical-vs-fundamental hybrids ([Ch. 4](04-model-types.md)), and bias-statistic methodology ([Ch. 14](14-model-evaluation.md)). - **Menchero, Orr & Wang (MSCI)**, "The Barra US Equity Model (USE4)" research notes: readable bridge between the handbooks and the academic literature. --- _Next: [Mini Example Source Code](18-mini-example-source-code.md)_