Ordinary Least Squares
Ordinary Least Squares (OLS) is the most widely used statistical estimation technique, which finds the linear relationship between a dependent variable and one or more independent variables by minimizing the sum of squared differences between observed data points and the fitted regression line. In finance, OLS underpins factor model estimation, beta calculation, and alpha identification.
Key takeaways
- OLS minimizes the sum of squared residuals to produce the Best Linear Unbiased Estimator (BLUE) under the Gauss-Markov assumptions.
- The OLS beta coefficient in a market model regression measures the systematic risk (market sensitivity) of a security.
- OLS results are reliable only when key assumptions hold: linearity, homoscedasticity, no autocorrelation, and no multicollinearity.
- In finance, heteroscedasticity (non-constant variance) and autocorrelation of residuals frequently violate OLS assumptions, requiring corrections.
- Factor models (CAPM, Fama-French) are estimated via OLS regression of security returns against factor returns over historical periods.
Explanation
Ordinary Least Squares regression is the workhorse of empirical finance, providing a systematic way to quantify the linear relationship between variables and test hypotheses about financial market behavior. The OLS estimator finds the coefficients β that minimize the sum of squared residuals: Σ(yᵢ − β₀ − β₁x₁ᵢ − ... − βₖxₖᵢ)², producing coefficient estimates that are linear in the observations, unbiased under classical assumptions, and efficient (minimum variance among all linear unbiased estimators) — the Gauss-Markov theorem guarantees these properties when the classical linear model assumptions are satisfied.
In the canonical market model (a one-factor version of CAPM), OLS regression of a security's excess returns against the market's excess returns produces the security's beta (systematic risk): rᵢ − rf = α + β(rₘ − rf) + ε. The OLS beta estimate — the slope coefficient — measures how much the security's excess return moves for each unit of market excess return. The alpha (intercept) measures risk-adjusted outperformance relative to the CAPM prediction. The R² of the regression quantifies the fraction of the security's variance explained by market movement.
The Gauss-Markov assumptions that guarantee OLS optimality are: (1) the model is correctly specified and linear in parameters; (2) the independent variables are not perfectly multicollinear; (3) the error term has zero conditional mean; (4) homoscedasticity (constant error variance); and (5) no serial correlation in errors. Financial return data routinely violates conditions (4) and (5): volatility clustering (GARCH effects) produces heteroscedasticity, and momentum and mean-reversion produce autocorrelation in returns. When these assumptions are violated, OLS estimates remain unbiased but are no longer efficient, and standard errors are incorrect — leading to faulty inference about statistical significance.
For the Fama-French three-factor model and its extensions, OLS is applied in a multivariate context: regressing security returns against the market factor, the SMB (small-minus-big) factor, and the HML (high-minus-low book-to-market) factor simultaneously. The partial regression coefficients for each factor represent the security's loading on that specific systematic risk dimension. In the information ratio framework, the breadth of independent bets and the individual signal IC (information coefficient, estimated via OLS correlation) combine in Grinold's Fundamental Law of Active Management: IR ≈ IC × √Breadth.
Cointegration testing builds on OLS: the Engle-Granger methodology runs OLS regression of one integrated time series on another and tests whether the residuals are stationary. If they are, the series are cointegrated — they share a long-run equilibrium relationship even if they individually wander without bounds. This is the statistical foundation for pairs trading: finding two cointegrated stocks, estimating their equilibrium relationship via OLS, and trading the spread when it deviates from equilibrium.
Formula
β = (XᵀX)⁻¹Xᵀy; Minimizes: Σ(yᵢ − ŷᵢ)²; Single factor: β = Cov(r_stock, r_market) / Var(r_market)
Example
A quantitative analyst estimates the Fama-French three-factor model for a U.S. large-cap equity fund using monthly returns over 60 months. The OLS regression of fund excess returns on market, SMB, and HML factor returns produces: α (Jensen's alpha) = 0.15% per month (t-statistic = 2.1, statistically significant at 5%), β_market = 0.92 (well-diversified, near-market exposure), β_SMB = −0.18 (slight large-cap tilt, expected), β_HML = 0.31 (value tilt). The R² is 0.87, meaning 87% of the fund's return variance is explained by the three factors. The OLS standard errors are corrected for heteroscedasticity using White's robust estimator after the Breusch-Pagan test finds evidence of non-constant error variance in the regression residuals. The positive, significant alpha suggests genuine skill, but only 0.15% × 12 = 1.8% annualized — modest after the three-factor risk adjustment.
Related terms
Alpha Autocorrelation Beta Breadth Cap Cointegration Correlation Equity Factor Model Fama French Three Factor Model Fundamental Law Of Active Management Information Coefficient