What Is Multiple Linear Regression (MLR)? Definition and Examples

Search Fintrev

Search our guides, reviews and glossary.

Multiple linear regression (MLR) is a statistical method that estimates how two or more explanatory variables, taken together, relate to a single outcome. It fits one linear equation to the data and assigns each explanatory variable its own coefficient. Each coefficient measures how much the outcome is expected to change when that variable moves by one unit and every other variable in the model stays the same. MLR is also called multiple regression, multivariable linear regression or a multilinear model. It extends ordinary least squares (OLS) regression from one predictor to several.

What Is Multiple Linear Regression?

Simple linear regression links one dependent variable to one independent variable, such as a stock’s return to the market’s return. But an oil producer’s share price responds to the broad market, crude prices and interest rates, and a currency pair responds to rate differentials, commodity prices and shifts in risk appetite. MLR puts these candidate drivers into one equation so their effects can be estimated at the same time.

The method is called multiple because it uses several explanatory variables, not several outcomes. It is linear because the outcome is modeled as a weighted sum of the explanatory variables plus a constant and an error term.

How Multiple Linear Regression Works

The multiple linear regression formula

For each observation i in a data set of n observations, the model is:

yᵢ = β₀ + β₁xᵢ₁ + β₂xᵢ₂ + … + βₚxᵢₚ + εᵢ

The components of the MLR formula are:

• yᵢ: the dependent variable, the outcome being explained or predicted.
• xᵢ₁ … xᵢₚ: the p explanatory variables for that observation.
• β₀: the intercept (constant term), the predicted value of y when every explanatory variable equals zero.
• β₁ … βₚ: the slope coefficients, one per explanatory variable.
• εᵢ: the error term, the part of y the model does not capture. Once the model is estimated, the observed errors are called residuals: actual value minus predicted value.

The β values in the equation are the true but unknown population parameters. Running the regression produces estimates of them, usually written with a hat (β̂). In matrix form, the least-squares estimates are β̂ = (XᵀX)⁻¹Xᵀy, where X holds the explanatory data (with a column of ones for the intercept) and y holds the outcomes. Inverting that matrix by hand is tedious beyond two or three variables, so software does the work.

How the coefficients are estimated

Ordinary least squares chooses the coefficients that make the sum of squared residuals as small as possible. Squaring keeps positive and negative misses from canceling out and penalizes large misses more heavily than small ones. With one predictor the fit is a line, with two a flat plane, and with more a hyperplane that works the same way.

Running a multiple regression step by step

1. Define the outcome, for example a monthly percentage return rather than a price level.
2. Choose explanatory variables with a plausible economic link to the outcome, not every series that happens to be available.
3. Align the data to the same frequency and time stamps, handling missing values consistently.
4. Estimate the model in software: Excel’s LINEST function or Analysis ToolPak, R’s lm(), or Python’s statsmodels or scikit-learn.
5. Read the output: coefficients, their significance and the overall fit.
6. Check the assumptions using the residuals and the correlations among predictors.
7. Test out of sample on data the model was not fitted on.

Interpreting multiple linear regression

A typical regression output reports several figures, and each answers a different question.

• Coefficient (β̂): the estimated change in y for a one-unit change in that variable, other variables held constant. The sign gives the direction and the size gives the magnitude, in the units of the data.
• Standard error: how precisely the coefficient is estimated. A large standard error relative to the coefficient means the estimate could plausibly be far from its reported value.
• t-statistic and p-value: the t-statistic is the coefficient divided by its standard error. The p-value is the probability of seeing a result at least that extreme if the true coefficient were zero. A small p-value (0.05 is a common threshold) suggests the variable adds explanatory power, but statistical significance is separate from economic significance: a tiny coefficient can be significant in a large sample, and significance says nothing about cause.
• R-squared: the proportion of the variation in y explained by all the predictors together, from 0 (none) to 1 (error-free prediction within the sample).
• Adjusted R-squared: R-squared corrected for the number of predictors, calculated as 1 - (1 - R²)(n - 1)/(n - p - 1). Adding a useless variable always nudges R-squared up but can push adjusted R-squared down.
• F-statistic: tests whether the predictors, as a group, explain the outcome better than a model with only an intercept. An F-test comparing the full model with one that excludes a subset of variables tests whether that subset matters.

The “holding others constant” interpretation is what separates MLR from running several simple regressions. If oil prices and the stock market tend to rise together, a simple regression of an energy stock on oil alone would pick up some of the market’s effect and credit it to oil. MLR separates the two, provided both are in the model.

Why Is Multiple Linear Regression Important?

MLR does two jobs that simple regression cannot. First, it controls for other factors. Comparing a fund’s return with its benchmark tells you little if the fund also leans toward small companies. Including a size factor shows how much of the return the tilt explains. Second, it predicts with more information. Several relevant inputs usually explain more of an outcome than one, as long as they are genuinely related to it.

In finance, the clearest applications are factor models.

• CAPM beta comes from a simple regression of an asset’s excess return on the market’s excess return.
• The Fama-French three-factor model, published by Eugene Fama and Kenneth French in 1993, turns that into a multiple regression by adding a size factor (small minus big, SMB) and a value factor (high minus low book-to-market, HML). These account for the historical outperformance of small-cap and value stocks relative to CAPM, which helps judge whether a manager added value beyond known style exposures. Later versions added profitability, investment and momentum factors.
• Hedge ratios for cross hedges are often estimated by regressing the exposure’s price changes on those of the hedging instrument. Adding further variables shows how much of the exposure a single hedging instrument leaves uncovered.
• Economic and currency models relate exchange rate moves to interest rate differentials, inflation gaps, commodity prices or risk measures.
• Risk and credit analysis relates losses, spreads or default measures to macroeconomic variables.

Example of Multiple Linear Regression

Hypothetical example. An analyst wants to explain the monthly return of an oil company’s stock using three variables: the stock market index’s monthly return, the monthly percentage change in crude oil, and the monthly change in the 10-year Treasury yield in percentage points. After fitting the model on several years of monthly data, the software returns these illustrative estimates:

Return = 0.2 + 1.1 × (Market return) + 0.3 × (Oil change) - 0.8 × (Yield change)

Assume the model has an R-squared of 0.62 and an adjusted R-squared of 0.60, and that all three coefficients are statistically significant.

Reading the coefficients:

• Intercept (0.2): in a month when the market, oil and yields are all unchanged, the model predicts a 0.2% return.
• Market (1.1): with oil and yields held constant, each 1 percentage point of market return is associated with 1.1 points of stock return.
• Oil (0.3): with the market and yields held constant, each 1% rise in crude is associated with a 0.3 point rise in the stock’s return.
• Yields (-0.8): with the other two held constant, each 1 percentage point rise in the 10-year yield is associated with a 0.8 point lower return.

Now plug in a scenario: the market rises 2%, oil rises 5% and the yield rises 0.25 percentage points.

0.2 + (1.1 × 2) + (0.3 × 5) - (0.8 × 0.25) = 0.2 + 2.2 + 1.5 - 0.2 = 3.7%

That is the model’s predicted return for the month. If the stock actually returned 4.5%, the residual for that month is 4.5 - 3.7 = 0.8 percentage points. An R-squared of 0.62 means the three variables explain about 62% of the variation in monthly returns over the sample. The other 38% sits in the residuals: company news, sector moves and noise the model does not capture.

The units matter. If yields had been entered in basis points instead of percentage points, the coefficient would be 100 times smaller even though the relationship is identical.

Limitations of Multiple Linear Regression

The assumptions often fail in financial data. The standard OLS results depend on:

• a relationship that is linear in the coefficients;
• errors that are independent of each other (no autocorrelation);
• errors with constant variance (homoscedasticity);
• no perfect linear relationship among the explanatory variables;
• for small-sample tests and confidence intervals, errors that are roughly normally distributed.

Market returns tend to have fat tails and skewness, so the bell curve assumption behind textbook p-values can understate how often extreme residuals occur. Volatility clusters, which violates constant variance. Daily returns can show autocorrelation. Each of these mainly distorts the standard errors, so a variable can look significant when it is not. Robust (heteroscedasticity-consistent) standard errors, Newey-West corrections and robust regression methods that down-weight outliers are common responses.

Multicollinearity. When two predictors move closely together, the model struggles to separate their effects. Coefficients become unstable, standard errors balloon, and signs can flip between samples. The overall fit may still be fine, but individual coefficients become unreliable. The variance inflation factor (VIF) measures how much a coefficient’s variance is inflated by correlation with the other predictors.

Omitted variable bias. If a factor that drives the outcome is left out and is correlated with an included variable, the included variable’s coefficient absorbs part of its effect and is biased. Adding variables reduces this risk, which is one reason MLR is preferred over simple regression. But only variables you actually include can be controlled for.

Correlation is not causation. A regression measures association, and a strong coefficient is no proof that the explanatory variable causes the outcome. Both might be driven by a third factor, or the causation might run the other way.

Overfitting. With enough variables, a model can fit historical data closely while describing noise. R-squared will keep rising as predictors are added. Adjusted R-squared, information criteria such as AIC and BIC, and out-of-sample testing guard against this.

Spurious regression with trending data. Regressing one price level on another, when both trend over time, can produce a high R-squared and significant coefficients for unrelated series. Working with returns or changes, or testing for cointegration, addresses this.

Unstable relationships. Coefficients estimated over one period describe that period. Correlations between assets, and between markets and macro variables, shift with policy regimes and crises.

Extrapolation. Predictions outside the range of the data used to fit the model, such as an oil move far larger than any in the sample, rest on the assumption that the linear relationship continues to hold there.

Simple vs. Multiple vs. Multivariate Regression

• Simple linear regression: one explanatory variable, one outcome. Linear regression with two independent variables is already a multiple regression.
• Multiple linear regression: two or more explanatory variables, one outcome.
• Multivariate regression: strictly, two or more dependent variables modeled together, such as the returns of several currency pairs on the same set of factors. In everyday usage people sometimes say “multivariate” when they mean multiple regression, so check which is meant.
• Nonlinear and generalized models: logistic and probit regression handle yes/no outcomes such as default or no default. They use several predictors but are not linear regression.

The “linear” in MLR refers to the coefficients, not necessarily to straight-line shapes in the variables. A model with x and x² as predictors (quadratic or polynomial regression) produces a curve, yet it can still be estimated by ordinary least squares because the outcome is a linear combination of the coefficients. A model in which a coefficient appears as an exponent, for example, is genuinely nonlinear and needs different estimation methods.

FAQs

The Bottom Line

Multiple linear regression matters to analysts who need to separate the effects of several drivers on one outcome, from factor models and hedge ratios to currency and credit studies. A good in-sample fit can still describe a relationship that is unstable or not causal. Users should check the assumptions and predictor correlations, compare models with adjusted R-squared, and test predictions on data the model was not fitted on.