Quant Analytics: from scratch to pro · Statistics for markets · Advanced
Regression, beta and alpha
Regressing an asset's returns on a benchmark's returns splits them into a part that moves with the benchmark (beta), a part that does not (alpha and noise), and tells you how much of the variation the benchmark explains. The estimates are noisy, and regressing prices instead of returns produces false relationships.
In this lesson
- Compute beta as covariance over variance and alpha as the intercept
- Read R-squared as the share of variance explained and derive the residual volatility from it
- Estimate how uncertain a beta and an alpha are
- Explain why regressions of price levels on price levels mislead
Read first: How sure is an average? Standard error and confidence intervals, Correlation and its limits
Most investors want to know two things about an asset relative to a benchmark. How much does it move when the benchmark moves? And is there anything left over that the benchmark does not explain? Regression answers both, and it also shows how little of the answer is certain.
The regression
The simplest model fits asset returns to benchmark returns:
R_asset = alpha + beta x R_market + e
For excess returns (above the risk-free rate) the same form is used. Beta is the slope: beta = covariance(asset, market) / variance(market), which equals correlation x (sigma_asset / sigma_market). Alpha is the intercept, the average return left after beta's share is taken out. The residual e is what neither explains. The R-squared is the share of the asset's variance explained by the market, and in a single-variable regression equals the squared correlation.
Fit it on returns, not prices, and say which benchmark you used. The beta against an index of large shares differs from the beta against a bond index.
Worked example
Five days of returns, in percent. Market: 1, -2, 3, 0, -1. Asset: 2, -3, 4, 1, -2.
- Sample covariance 5.4, market variance 3.7, so
beta = 5.4 / 3.7= 1.46. - Means are 0.2 (market) and 0.4 (asset), so
alpha = 0.4 - 1.46 x 0.2= 0.11% a day. - R-squared is
5.4^2 / (3.7 x 8.3)= 94.95%, using the asset variance of 8.3.
Five points are only for the arithmetic. The estimates below show how much real data it takes.
Worked example: from volatilities and correlation
An asset has a volatility of 24%, the market 16%, and the correlation is 0.7.
- Beta:
0.7 x 24 / 16= 1.05. - R-squared:
0.7^2= 49%. - Residual volatility:
24% x sqrt(1 - 0.49)= 17.14%. This part does not move with the market.
Half of the variance is the market. The rest is specific to the asset and is what diversification across many assets removes (see why diversification works).
Now alpha. Suppose the asset returned 9% a year, the market 7%, and the risk-free rate is 2%. Alpha is (9% - 2%) - 1.05 x (7% - 2%) = 1.75% a year.
How uncertain are the estimates
Alpha's standard error is about the residual volatility divided by sqrt(n). For ten years, 17.14% / sqrt(10) = 5.42%, so the t-statistic is 0.32. A positive alpha of 1.75% a year is statistically invisible. Alphas of that size need decades to confirm, for the reasons in standard error.
Beta is better known, but not exactly. Its standard error is about (residual vol / market vol) / sqrt(n). With 60 monthly observations that is 0.14, so a 95% interval for the beta of 1.05 runs from 0.78 to 1.32. This is why a rolling beta, re-estimated on a moving window, wanders even when the true beta is stable. Longer windows reduce noise but react more slowly to genuine change.
Regressing prices: the spurious regression
Granger and Newbold (1974) showed that regressing one independent random walk on another often gives a high R-squared and a seemingly significant slope, though there is no relationship by construction. Trending levels share a direction by accident, and the usual standard errors assume away exactly the persistence that causes it.
The rule: regress returns on returns. If a relationship between levels is claimed, it needs methods built for it, such as cointegration tests. Warning signs are an R-squared near one with residuals that are themselves trending.
Other limits
- Beta is not fixed. It shifts with the regime and with the asset's own leverage.
- Alpha depends on the benchmark chosen. Add a size or value factor and part of it moves into betas on those factors (see Fama and French, 1993).
- Outliers pull a least-squares line. One crash day can set the slope.
- In a crisis correlations rise, so a calm-period beta can understate what an asset does in a sell-off.
Common mistakes
- Regressing prices on prices.
- Reading R-squared as a measure of usefulness. A low value says the benchmark explains little, not that the asset is bad.
- Reporting an alpha without its standard error.
- Fitting beta on a short window and treating it as a constant.
- Comparing betas computed against different benchmarks or different return frequencies.
In practice
State the benchmark, the return frequency and the window, and report each estimate with its interval. If alpha is positive, ask whether it is larger than twice its standard error, and what else besides skill could produce it. Analytics on FinUties applies measures like these to the markets it covers, so they are at hand when you read a chart.
Standard error of beta against the number of observations
Worked numbers
Each result is recomputed from the formula shown.
| What | Formula | Result |
|---|---|---|
| sample covariance of the five-day example | 5.4 | |
| sample variance of market returns in the example | 3.7 | |
| beta of the five-day example | cov5 / var5 | 1.46 |
| alpha of the example, in percent per day | 0.4 - beta5 * 0.2 | 0.11% |
| R-squared of the five-day example | cov5**2 / (var5 * 8.3) | 94.95% |
| beta from correlation and volatilities | 0.7 * 0.24 / 0.16 | 1.05 |
| R-squared | 0.7**2 | 49% |
| residual volatility | 0.24 * sqrt(1 - r2) | 17.14% |
| alpha from mean returns (asset 9%, risk-free 2%, market 7%) | (0.09 - 0.02) - beta * (0.07 - 0.02) | 1.75% |
| standard error of alpha, 10 years | resid / sqrt(10) | 5.42% |
| t-statistic of alpha | alpha / se_alpha | 0.32 |
| standard error of beta, 60 monthly observations | (resid / 0.16) / sqrt(60) | 0.14 |
| lower end of beta interval | beta - 1.96 * se_beta60 | 0.78 |
| upper end of beta interval | beta + 1.96 * se_beta60 | 1.32 |
Check yourself
The covariance of an asset with the market is 5.4 and the market variance is 3.7 (both in squared percent). What is beta?
5.4 / 3.7 = 1.46. The asset has moved about 1.46 percent for each 1 percent in the market.
An asset has a volatility of 24%, the market 16%, and the correlation is 0.7. What are beta and R-squared?
Beta = 0.7 x 24 / 16 = 1.05. R-squared = 0.7^2 = 0.49, so the market explains 49% of the variance.
With that asset, what is the residual volatility?
24% x sqrt(1 - 0.49) = 17.14%. This is the part that does not move with the market.
Over 10 years the fitted alpha is 1.75% a year with a residual volatility of 17.14%. Is it distinguishable from zero?
The standard error is 17.14% / sqrt(10) = 5.42%, so t = 0.32. Not at all.
Why is a regression of one trending price on another unreliable?
Two unrelated series that both wander upward or downward show high R-squared and apparently strong significance. This is the spurious regression of Granger and Newbold. Regress returns, which are far closer to stationary.
Key terms
Sources
- The Econometrics of Financial Markets. Princeton University Press (John Y. Campbell, Andrew W. Lo, A. Craig MacKinlay), 1997. Accessed 9 Nov 2026
- Spurious Regressions in Econometrics. Journal of Econometrics 2(2), 111-120 (C. W. J. Granger, P. Newbold), 1974. Accessed 9 Nov 2026
- Common risk factors in the returns on stocks and bonds. Journal of Financial Economics 33(1), 3-56 (Eugene F. Fama, Kenneth R. French), 1993. Accessed 9 Nov 2026