Low Beta Is Not Low Risk: Calculate Beta, Correlation and Alpha Before Comparing Stocks

A stock has a beta of just 0.4. Is it safer than the market? Not necessarily. Give the stock annual volatility of 40%, the market volatility of 20%, and a correlation of 0.2: its beta is exactly 0.4. Low sensitivity to that particular benchmark can coexist with much higher total volatility.

The useful question is not whether every risk statistic should be high or low. It is which part of returns the statistic describes and which exposures remain outside it. A short, reproducible regression makes the distinction concrete—and shows why beating an index is not automatically evidence of skill.

Beta is a sensitivity, not a maximum loss

Let y be a stock’s return minus the same-period risk-free return, and x the benchmark’s return minus that risk-free return. A single-factor regression with an intercept is y=α+βx+ε. Beta is the slope, alpha the intercept, and epsilon the period’s unexplained residual. Regress matched excess returns, not price levels or a mixture of monthly returns and annual interest rates.

Estimate beta as Cov(x,y)/Var(x), equivalently ρ×σy/σx. Covariance and variance must use the same paired observations and denominator convention. Correlation measures the strength of a linear association; beta also incorporates relative volatility. Beta can exceed one, whereas correlation cannot.

The opening example gives 0.2×40%/20%=0.4. In a one-regressor ordinary least-squares fit with an intercept, sample R² equals squared correlation: only 4% in that example. This is an explained share of sample variation, not a 4% probability of losing money. The unexplained exposure has not disappeared.

Nor is beta a daily multiplication guarantee. A stock with beta 1.2 need not fall 1.2% when its benchmark falls 1%. An earnings shock, refinancing problem or liquidity event can make the residual more important than the market-related component.

Reproduce beta of 1.2 and alpha of 0.2% from four observations

The table contains hypothetical monthly excess returns, in percent. Risk-free returns have already been deducted. Both series use the same currency, valuation cutoff and total-return treatment of distributions. Four observations demonstrate arithmetic; they are not a real product record or adequate evidence of investment skill.

The mean of x is zero and the mean of y is 0.2%. Working in percentage-point units, Σ(x −mean x)²=10 and Σ(x −mean x)(y −mean y)=12. Thus beta is 12/10=1.2, and alpha is mean y −beta×mean x=0.2%. That is 0.2 percentage points per month, not 20% and not an annual return.

Fitted returns are −2.2%, −1.0%, 1.4% and 2.6%; residuals are 1%, −1%, −1% and 1%. Residuals sum to zero, as does their cross-product with x. Those are consistency checks for the least-squares fit with an intercept, not demonstrations of forecasting power.

The stock’s centered sum of squares is 18.4, of which 14.4 is explained and 4 remains residual. R² is therefore 14.4/18.4, approximately 78.26%; correlation is about 0.8847. Sample standard deviations are √(10/3)=1.8257% for the market and √(18.4/3)=2.4766% for the stock. A beta of 1.2, correlation of 0.8847 and volatility of 2.4766% describe different properties.

With four observations and two fitted coefficients, there are only two residual degrees of freedom. A positive intercept and an apparently high R² do not establish persistent outperformance. Evaluation requires representative data, estimation uncertainty and observations not used to select the model.

Outperformance is not the same thing as alpha

Suppose the benchmark’s excess return in a month is 5% and the stock’s is 6%. The stock beats it by one percentage point. But with an applicable beta of 1.2, market exposure alone corresponds to 1.2×5%=6%. The raw lead does not, on its own, demonstrate added value from stock selection.

Regression alpha is conditional on a chosen model and sample. Introduce an appropriate industry, size or style factor and some apparent alpha may become identifiable factor exposure. Repeatedly changing benchmarks, dates and fee treatment until the intercept turns positive is not a valid test.

Costs also affect what the investor receives. If a precisely fixed fee subtracts 0.1 percentage points each month from the worked example, every y shifts down by that amount. Beta and residuals stay unchanged; alpha falls to 0.1%. Real costs that vary with turnover, assets or performance need not behave this neatly. Gross, pretax alpha is not a substitute for a net outcome.

Keep historical regression separate from forward-looking asset-pricing assumptions. For illustration only, a 3% annual risk-free rate, an expected market return of 8% and beta of 1.2 imply 3%+1.2×(8%−3%)=9% under the capital asset pricing model. The 8% input is an assumption, not a future fact discovered by the model. Adding the four-month sample’s monthly alpha to that annual estimate would mix horizons and unsupported forecasts.

Why two data providers can report different betas

Start with the benchmark. A technology stock measured against a broad equity index, its industry index or a foreign market answers a different question each time. Divergent results need not mean someone made an arithmetic error. Currency treatment matters too: converting only one series introduces an inconsistent exchange-rate exposure.

Then check frequency and dates. Daily, weekly and monthly estimates react differently to nonsynchronous trading, suspensions, illiquidity and stale prices. Matching overseas closes simply because they share a calendar date can understate contemporaneous covariance. Missing observations are not zero returns.

Check whether the business or holdings have changed. Acquisitions, a different business mix, balance-sheet leverage or portfolio turnover can make an old estimate a poor description of current exposure. Rolling windows reveal instability, but the latest window is not automatically the best forecast. Specify the rule before selecting favorable results.

Finally, stress what the regression leaves out. Mechanically retaining the example’s 0.2% intercept and 1.2 slope, a benchmark excess return of −5% gives a fitted −5.8%. Add a separate residual shock of −6 percentage points and the outcome becomes −11.8%. This is an explicit scenario, not a loss ceiling or probability forecast. It reveals the exposure omitted by a beta-only calculation.

Match the statistic to the decision

For sensitivity to a particular equity market, request beta together with its benchmark, frequency, dates, sample size and estimation interval. For the volatility or losses a holder might experience, inspect total volatility, drawdowns, concentration and concrete stress scenarios. A small beta cannot answer that second question alone.

For a proposed addition to an existing portfolio, calculate correlation, covariance and component risk contributions using the actual holdings. A stock’s beta against one index does not replace a portfolio risk budget. Historical negative beta does not guarantee a hedge will keep working in a crisis.

For a manager’s claimed skill, separate raw outperformance, factor exposure and the after-cost intercept. Then examine uncertainty and out-of-sample stability. Positive alpha is a reason to investigate, not the end of the investigation.

A usable risk report lets the reader reproduce the estimate from matched returns and identify conditions under which its interpretation fails. A beta without its benchmark, observation period or uncertainty is not enough to label an investment safe or dangerous.