Why Estimate Beta From Historical Returns?
Equity beta measures how an asset’s returns have moved in relation to market returns. In cost-of-equity analysis, beta is an input for describing systematic risk: risk associated with movements in the market rather than risk unique to one company. Historical beta estimation uses paired observations of a stock’s returns and a market benchmark’s returns to estimate that relationship. The estimate is informative, but it is not a direct observation of a fixed, known characteristic. Its value depends on the data, period, and statistical model used.
The calculation requires a consistent series of return observations. Each stock return must correspond to the market return over the same interval, and both series must use the same return convention. For the standard CAPM regression, use excess returns: each asset’s return less the risk-free return for the same period. This tutorial focuses on estimating and interpreting beta, not on converting that estimate into a cost of equity.
The Regression Model and OLS Estimate
For observation \(t\), let \(R_{i,t}\) be the equity return, \(R_{m,t}\) the market return, and \(R_{f,t}\) the risk-free return, all measured over the same interval. Define excess returns \(y_t=R_{i,t}-R_{f,t}\) and \(x_t=R_{m,t}-R_{f,t}\). The market-model regression is
Here, \(\alpha\) is the regression intercept, \(\beta\) is the slope, and \(\varepsilon_t\) is the residual: the part of the observed equity excess return not accounted for by the fitted line. OLS chooses the intercept and slope that minimize the sum of squared residuals. With \(\bar{x}\) and \(\bar{y}\) denoting the sample means, the estimates are
The denominator must be positive; otherwise, all market excess returns are identical and the slope cannot be estimated. The slope is dimensionless when both return series use the same units. If returns are entered as percentages, the slope is unchanged by expressing both series as decimal returns instead. The intercept, however, has return units.
Why the OLS Slope Has This Form
For any proposed slope \(b\), the intercept that minimizes squared residuals is \(a=\bar{y}-b\bar{x}\). This centers the fitted regression around the sample means. The residuals can then be written as \((y_t-\bar{y})-b(x_t-\bar{x})\). Minimizing their sum of squares gives the slope formula.
Because \(Q(b)\) is a quadratic with positive coefficient \(\sum (x_t-\bar{x})^2\), this solution is its minimum when the market returns vary. The derivation also shows why beta depends on how stock and market returns vary together, relative to the variation in market returns.
Worked Example: Estimating Beta From Four Annual Observations
Method. Regress the equity’s annual excess returns \(y_t\) on the market’s annual excess returns \(x_t\), using OLS. The following hypothetical observations are percentages; the risk-free return has already been subtracted separately for each year.
| Year | Market excess return \(x_t\) | Equity excess return \(y_t\) |
|---|---|---|
| 1 | -2 | -1 |
| 2 | 0 | 1 |
| 3 | 2 | 5 |
| 4 | 4 | 7 |
First calculate the means: \(\bar{x}=(-2+0+2+4)/4=1\) percentage point and \(\bar{y}=(-1+1+5+7)/4=3\) percentage points. The centered market values are \(-3,-1,1,3\); the centered equity values are \(-4,-2,2,4\). Thus,
Therefore, \(\hat{\beta}=28/20=1.40\). The intercept is \(\hat{\alpha}=3-(1.40)(1)=1.60\) percentage points. The fitted line is \(\hat{y}=1.60+1.40x\).
Check. At \(x=-2,0,2,4\), the fitted values are \(-1.20,1.60,4.40,7.20\). Subtracting these from the observed values gives residuals \(0.20,-0.60,0.60,-0.20\). They sum to zero, and their products with the centered market values sum to \((-3)(0.20)+(-1)(-0.60)+(1)(0.60)+(3)(-0.20)=0\), as required for an OLS fit with an intercept. Beta says the fitted equity excess return changes by 1.40 percentage points for a one-percentage-point market excess-return change; it does not say every observation lies on the fitted line.
Common mistake avoided. The slope is not the ratio of the two sample means. It uses centered cross-products divided by centered market squared deviations; the intercept is then recovered from the means.
Interpreting the Slope and Intercept
A beta estimate above 1 indicates that the fitted equity excess return has moved more than the market excess return, on average in the regression relationship. A beta between 0 and 1 indicates a smaller fitted movement in the same direction. A negative estimate indicates an estimated opposite-direction relationship. These are interpretations of the fitted slope, not guarantees about future returns or statements that the company is riskier in every respect. Beta concerns market-related movement; it does not capture all company-specific risk.
The intercept is not beta. In an excess-return regression, \(\hat{\alpha}\) is the fitted equity excess return when the market excess return is zero. It is an estimated regression intercept, not automatically a forecast or a measure of persistent abnormal performance. Its interpretation depends on the model and the chosen sample.
Worked Example: Matching Quarterly Observations
Method. Use four hypothetical quarterly excess-return pairs for a listed infrastructure fund. Both returns are measured over the same calendar quarter and stated in percentage points.
| Quarter | Market excess return \(x_t\) | Fund excess return \(y_t\) |
|---|---|---|
| 1 | -2 | -3 |
| 2 | -1 | 0 |
| 3 | 1 | 1 |
| 4 | 2 | 5 |
The market mean is \(\bar{x}=(-2-1+1+2)/4=0\). The fund mean is \(\bar{y}=(-3+0+1+5)/4=0.75\). Since the market mean is zero, the centered market values are the original values. The cross-product sum is \((-2)(-3)+(-1)(0)+(1)(1)+(2)(5)=17\), and the market squared-deviation sum is \(4+1+1+4=10\). Thus \(\hat{\beta}=17/10=1.70\), and \(\hat{\alpha}=0.75-(1.70)(0)=0.75\) percentage points.
The estimate describes a quarterly-return relationship because the observations are quarterly. The beta itself is not multiplied by four or by the square root of four to convert it to an annual beta. Such scaling rules, used for other quantities under particular assumptions, do not convert a regression slope into an annualized beta. A regression using annual observations could produce a different estimate because it uses different data and a different aggregation.
Check. The fitted returns at \(x=-2,-1,1,2\) are \(-2.65,-0.95,2.45,4.15\). The observed-minus-fitted residuals are \(-0.35,0.95,-1.45,0.85\). They sum to zero, and their products with the corresponding market values sum to \(0.70-0.95-1.45+1.70=0\). The regression conditions for an OLS fit with an intercept are satisfied.
Common mistake avoided. Pairing a fund return from one quarter with a market return from a different quarter would estimate a relationship between mismatched periods. Frequency labels alone are not enough; observation dates and interval boundaries must align.
Sample Window, Frequency, and Estimation Uncertainty
The sample window is the span of time covered by the observations. A longer window can provide more observations and include a wider range of market conditions, but it may combine periods in which the company’s business mix, leverage, or market relationship changed. A shorter window can better reflect recent conditions, but may be dominated by a small number of unusual observations. Neither choice is universally best. The window should be selected deliberately and reported with the estimate.
Return frequency creates a related trade-off. Monthly observations usually provide more data points than annual observations over the same span, but higher-frequency returns can contain more short-term noise and may be affected by trading patterns or nonsynchronous price observations. When estimating beta, use matched periods, a consistent return definition, and a clearly identified market benchmark. Do not mix, for example, monthly equity returns with annual market returns, or simple returns for one series with a different return convention for the other.
An OLS slope calculated from historical data is subject to sampling uncertainty. Under the usual linear-regression assumptions that the errors have mean zero conditional on the market returns, are independent across observations, and have a common finite variance, an estimated standard error for the slope is
The \(n-2\) denominator reflects the two fitted regression parameters, intercept and slope. When the errors are normally distributed as well, a conventional \(t\)-interval for the slope uses \(n-2\) degrees of freedom. If the assumptions fail, such as when residual variability changes over time or observations are dependent, this standard error and interval may not be reliable without an appropriate adjustment. A point estimate alone does not communicate this uncertainty.
Worked Example: Quantifying Uncertainty in a Beta Estimate
Method. Suppose a hypothetical regression uses five monthly excess-return pairs, with market values \(x=(-2,-1,0,1,2)\) and equity values \(y=(-3,0,0,3,4)\), all in percentage points. Assume, for this calculation, independent errors with constant variance and a linear relationship.
The means are \(\bar{x}=0\) and \(\bar{y}=(-3+0+0+3+4)/5=0.8\). The cross-product sum is \(6+0+0+3+8=17\), and \(S_{xx}=\sum(x_t-\bar{x})^2=4+1+0+1+4=10\). Hence \(\hat{\beta}=17/10=1.70\), and \(\hat{\alpha}=0.8\).
The fitted values are \(-2.6,-0.9,0.8,2.5,4.2\). The residuals, observed minus fitted, are \(-0.4,0.9,-0.8,0.5,-0.2\). Squaring and adding gives \(\operatorname{SSE}=0.16+0.81+0.64+0.25+0.04=1.90\). The residual variance estimate is \(1.90/(5-2)=0.6333\) percentage-points-squared. Therefore,
For illustration, using the 95% two-sided \(t\) critical value \(3.182\) with 3 degrees of freedom gives a margin of \(3.182(0.252)\approx0.802\). The corresponding interval is \(1.70\pm0.802\), or approximately \([0.898,2.502]\). These endpoints are rounded to three decimals. The interval is wide because there are only five observations; it should not be mistaken for a precise measurement of a permanent beta.
Check. The residuals sum to zero, and their products with \(x\) sum to \((-2)(-0.4)+(-1)(0.9)+(0)(-0.8)+(1)(0.5)+(2)(-0.2)=0\). Also, the standard error is positive because the residual sum of squares and market squared-deviation sum are positive.
Common mistake avoided. Dividing the residual sum of squares by \(n\), rather than by \(n-2\), ignores the two fitted parameters and understates the variance estimate in this example.
Practical Checks Before Reporting Beta
A reported historical beta is reproducible only if the analyst states enough detail to identify what was estimated. In practice, report the equity and market benchmark, the sample window, return frequency, return definition, and whether excess returns were used. Also inspect the data for unmatched dates, outliers, and changes in the business or capital structure that could affect whether the historical relationship is informative for the intended analysis.
- Confirm matching. Each equity return and benchmark return must cover the same interval; use the risk-free return for that same interval when calculating excess returns.
- Keep units consistent. If one series is expressed in percentages and the other in decimals, convert them to a common unit before regression.
- Distinguish the estimates. The slope is beta; the intercept is not a second beta or the equity’s average return.
- Describe estimation risk. Different windows, frequencies, and benchmarks can produce different estimates. A standard error or sensitivity comparison helps reveal that dependence.
Historical beta is therefore an estimate conditional on a model and a dataset, not a timeless company constant. The slope summarizes a particular sample’s fitted relationship. Sound interpretation requires attention to how the data were constructed and how much uncertainty remains.
Check Your Understanding
For each problem, use the OLS regression with an intercept unless the question says otherwise. Return values are in percentage points.
- A hypothetical insurer has market excess returns \((-2,0,2)\) and equity excess returns \((-3,1,3)\). Calculate \(\hat{\beta}\) and \(\hat{\alpha}\). (Answer: \(\hat{\beta}=1.50\), \(\hat{\alpha}=0.33\) percentage points.)
- For four matched monthly observations, market excess returns are \((-1,0,1,2)\), and stock excess returns are \((-2,1,2,4)\). Find the OLS slope and intercept. (Answer: \(\hat{\beta}=1.90\), \(\hat{\alpha}=0.30\) percentage points.)
- An analyst has monthly stock returns for January through June but market returns for February through July. What must be corrected before estimating beta, and why? (Answer: Align the return intervals so each stock observation is paired with the market return over the same month; otherwise the estimated slope does not represent the intended matched-period relationship.)
- A regression has \(n=12\), \(\operatorname{SSE}=30\), and \(\sum(x_t-\bar{x})^2=48\). Calculate the estimated slope standard error. (Answer: \(\sqrt{(30/10)/48}=0.25\).)
- A monthly regression estimates beta as \(1.25\). A candidate multiplies it by 12 to report an annual beta. Is that justified by the regression model? (Answer: No. Beta is a slope and is not annualized by multiplying by 12; an annual-frequency regression may yield a different estimate from annual observations.)