If you’re preparing for a quantitative analyst interview at Akuna Capital or any top-tier trading firm, you’ll almost certainly be tested on your understanding of regression analysis and, specifically, the assumptions underlying Ordinary Least Squares (OLS). OLS is a foundational technique in statistics and quantitative finance, central to everything from risk modeling to strategy backtesting. This article thoroughly explores the OLS assumptions, why they matter, and how they relate to both theory and practice—vital knowledge for any aspiring quant.

Akuna Capital Quantitative Analyst Interview Question: Assumptions of Ordinary Least Squares (OLS)

Table of Contents

• Overview of Ordinary Least Squares (OLS)
• Why Do OLS Assumptions Matter?
• List of OLS Assumptions
• Assumption 1: Linearity in Parameters
• Assumption 2: Random Sampling
• Assumption 3: No Perfect Multicollinearity
• Assumption 4: Zero Conditional Mean
• Assumption 5: Homoscedasticity
• Assumption 6: Normality of Errors (for inference)
• Consequences of Violating OLS Assumptions
• Diagnosing and Testing OLS Assumptions
• OLS in Quantitative Finance and Interviews
• Conclusion

Overview of Ordinary Least Squares (OLS)

Ordinary Least Squares (OLS) is the most common method for estimating the parameters of a linear regression model. In its simplest form, OLS seeks to fit a line through a set of data points such that the sum of the squared differences between the observed values and those predicted by the model is minimized.

The classic simple linear regression model can be written as:

$$ y_i = \beta_0 + \beta_1 x_i + \epsilon_i, \quad i = 1, 2, ..., n $$

Where:

• \(y_i\) = dependent variable (response variable)
• \(x_i\) = independent variable (predictor)
• \(\beta_0\) = intercept
• \(\beta_1\) = slope coefficient
• \(\epsilon_i\) = error term (residual) for observation \(i\)

For multiple regression with \(k\) predictors:

$$ y_i = \beta_0 + \beta_1 x_{i1} + \beta_2 x_{i2} + ... + \beta_k x_{ik} + \epsilon_i $$

OLS estimates the coefficients \(\beta_0, \beta_1, ..., \beta_k\) by minimizing:

$$ \sum_{i=1}^n (y_i - \hat{y}_i)^2 = \sum_{i=1}^n (y_i - \beta_0 - \beta_1 x_{i1} - ... - \beta_k x_{ik})^2 $$

Why Do OLS Assumptions Matter?

The OLS estimator is elegant and intuitive, but its desirable properties—such as unbiasedness, efficiency, and consistency—only hold when certain statistical assumptions are satisfied. These assumptions ensure that the results you obtain from regression analysis are valid and meaningful.

If one or more assumptions are violated, the OLS estimates can be biased, inefficient, or misleading. In quantitative finance, where decisions often rely on subtle statistical relationships, ignoring these assumptions can lead to poor trading strategies or risk models.

List of OLS Assumptions

The classical linear regression model makes several key assumptions. These are often summarized as follows:

• Linearity in parameters
• Random sampling
• No perfect multicollinearity
• Zero conditional mean (exogeneity)
• Homoscedasticity (constant error variance)
• Normality of errors (for inference)

Let’s examine each assumption in detail.

Assumption 1: Linearity in Parameters

Definition

The OLS regression model assumes that the relationship between the dependent variable and the independent variables is linear in the parameters. This doesn’t mean the relationship must be linear in the variables themselves; it means the model is a linear function of the coefficients \(\beta_j\).

Mathematically:

$$ y_i = \beta_0 + \beta_1 x_{i1} + \beta_2 x_{i2} + ... + \beta_k x_{ik} + \epsilon_i $$

The key is that the betas (\(\beta_j\)) appear linearly.

Implications

• Allows for transformations of variables (e.g., log, square root) as long as the model is linear in \(\beta_j\).
• Does not mean you can’t model nonlinear relationships—just that you must express them in a way that is linear in the parameters.

Examples

What Happens If Violated?

• If the model is nonlinear in parameters, OLS cannot be used directly. You must use nonlinear regression techniques.
• Misspecification can lead to biased or inconsistent estimates.

Assumption 2: Random Sampling

Definition

The data used in the regression analysis must be collected via random sampling from the population. Each observation \((x_i, y_i)\) is independently and identically distributed (i.i.d.).

Implications

• Ensures that the sample is representative of the population.
• Avoids selection bias, which can invalidate inferences about the population.

Examples

• Randomly choosing stocks for a portfolio regression meets this assumption.
• Using only stocks from a particular sector (e.g., only tech stocks) may violate randomness if you want to generalize to all equities.

What Happens If Violated?

• If the data is not random (e.g., time series with autocorrelation, or biased sampling), OLS estimators may be biased or inconsistent.
• In finance, time series data often exhibit autocorrelation, so specialized techniques (e.g., time series models, robust standard errors) are required.

Assumption 3: No Perfect Multicollinearity

Definition

None of the independent variables are exact linear functions of each other. That is, no predictor can be written as a perfect linear combination of the others:

$$ x_{j} = a_1 x_1 + a_2 x_2 + ... + a_{j-1} x_{j-1} + a_{j+1} x_{j+1} + ... + a_k x_k $$

Implications

• Ensures that each coefficient can be estimated uniquely.
• Prevents the design matrix \(X\) from being singular (i.e., not invertible).

Examples

• If \(x_2 = 2 x_1\), perfect multicollinearity exists, and OLS cannot estimate unique coefficients for \(x_1\) and \(x_2\).
• Including both “age” in years and “age” in months as predictors in the same regression is an example of perfect multicollinearity.

What Happens If Violated?

• OLS estimation fails; the matrix \(X^T X\) is not invertible.
• Statistical software usually drops one of the collinear variables automatically, but this should be handled deliberately during model design.

Near Multicollinearity (Imperfect Multicollinearity)

If variables are highly but not perfectly correlated, OLS can still be estimated, but:

• Coefficient estimates may have large standard errors.
• Interpretation of individual coefficients becomes difficult.

Assumption 4: Zero Conditional Mean (Exogeneity)

Definition

The expected value of the error term, conditional on the independent variables, is zero:

$$ E[\epsilon_i | X] = 0 $$

This is also known as the exogeneity assumption.

Implications

• Ensures that all relevant information for predicting \(y\) is captured by the regressors \(X\).
• Guarantees that OLS estimates are unbiased: \(E[\hat{\beta}_j] = \beta_j\).

Examples

• If you omit a relevant variable that is correlated with both the dependent variable and one or more independent variables, this assumption is violated (omitted variable bias).
• If there is simultaneity (e.g., price and demand are determined together), exogeneity fails.
• If measurement error exists in the regressors, this can also violate the assumption.

What Happens If Violated?

• The OLS estimators become biased and inconsistent.
• Leads to unreliable inference and predictions.

Mathematical Illustration

Suppose the true model includes a variable \(z\) that is omitted from the regression:

$$ y = \beta_0 + \beta_1 x + \beta_2 z + \epsilon $$

But you estimate:

$$ y = \beta_0 + \beta_1 x + u $$

Where \(u = \beta_2 z + \epsilon\). If \(x\) and \(z\) are correlated, then \(E[u|x] \neq 0\), violating the zero conditional mean assumption.

Assumption 5: Homoscedasticity (Constant Error Variance)

Definition

The variance of the error term is constant across all levels of the independent variables:

$$ Var(\epsilon_i | X) = \sigma^2 $$

Implications

• Ensures that OLS estimators are efficient (i.e., have the smallest possible variance among all linear unbiased estimators, per the Gauss-Markov theorem).
• Standard errors, hypothesis tests, and confidence intervals are valid only if this assumption holds.

Examples

• In finance, the variance of returns is often not constant over time (volatility clustering)—a classic case of heteroscedasticity.
• Regressing income on education may show larger residuals for high-income individuals than for low-income ones (heteroscedasticity).

What Happens If Violated?

• OLS coefficient estimates remain unbiased, but they are no longer efficient (not minimum variance).
• Standard errors are incorrect, leading to invalid hypothesis tests and confidence intervals.

Detecting Heteroscedasticity

Common tests include Breusch-Pagan and White tests. Visual inspection of residual plots can also be useful.

Addressing Heteroscedasticity

• Use robust (heteroscedasticity-consistent) standard errors (e.g., White’s correction).
• Transform variables (e.g., log transformation).
• Use weighted least squares.

Assumption 6: Normality of Errors (for inference)

Definition

The error terms are normally distributed, especially important for small sample sizes:

$$ \epsilon_i \sim N(0, \sigma^2) $$

Implications

• • Normality is not required for OLS estimators to be unbiased or consistent.
  • It is required for the validity of t-tests, F-tests, and construction of confidence intervals in small samples.
  • As the sample size increases, the Central Limit Theorem (CLT) allows inference even when errors are not perfectly normal.

  Examples

  • Financial returns often have fat tails (leptokurtosis) and may not be perfectly normal.
  • In cross-sectional data, errors can be skewed or have outliers, violating normality.

  What Happens If Violated?

  • OLS estimators remain unbiased and consistent.
  • Standard errors, p-values, and confidence intervals may be unreliable in small samples.
  • In large samples, the CLT often mitigates the problem, but extreme non-normality (e.g., due to outliers) can still affect results.

  Detecting Non-Normality

  • Visual inspection: Q-Q plots of residuals.
  • Formal tests: Shapiro-Wilk, Jarque-Bera, Anderson-Darling.

  Addressing Non-Normality

  • Transform response variable (e.g., log, square root).
  • Use robust inference methods (e.g., bootstrapping).
  • Remove or winsorize outliers, if justified.

  ---

  Consequences of Violating OLS Assumptions

  Understanding the consequences of assumption violations is vital for both interviews and real-world analysis. Below is a summary:

  Assumption	Violation Consequence	Remedies
  Linearity	Biased, inconsistent estimators; model misspecification.	Transform variables, use nonlinear regression.
  Random Sampling	Biased, inconsistent estimators; invalid inference.	Improve data collection, use time series models if needed.
  No Perfect Multicollinearity	Cannot estimate unique coefficients; standard errors inflate under high (but not perfect) multicollinearity.	Remove or combine collinear variables.
  Zero Conditional Mean	OLS estimators are biased and inconsistent (e.g., omitted variable bias, endogeneity).	Add omitted variables, use instrumental variables.
  Homoscedasticity	Unreliable standard errors, invalid inference, less efficient estimators.	Use robust standard errors, transform variables, weighted least squares.
  Normality	Invalid inference in small samples, unreliable p-values and confidence intervals.	Apply transformations, bootstrapping, nonparametric methods.

  ---

  Diagnosing and Testing OLS Assumptions

  A skilled quantitative analyst must know how to check if OLS assumptions hold. Here’s how you can test these in practice:

  1. Linearity

  • Plot residuals vs. fitted values. Non-random patterns may suggest nonlinearity.
  • Try adding polynomial or interaction terms.

  2. Random Sampling

  • Understand your data source: is it a random sample, time series, or panel data?
  • For time series, check for autocorrelation (see below).

  3. Multicollinearity

  • Compute Variance Inflation Factors (VIF) for each regressor.

    
    import statsmodels.api as sm
    from statsmodels.stats.outliers_influence import variance_inflation_factor
    
    X = sm.add_constant(X)  # Add intercept
    vif = [variance_inflation_factor(X.values, i) for i in range(X.shape[1])]
    print(vif)
        

  • VIFs above 5 or 10 indicate high multicollinearity.

  4. Zero Conditional Mean (Exogeneity)

  • This is hard to test directly; check for omitted variable bias conceptually.
  • Instrumental variables (IV) regression can help if you suspect endogeneity.

  5. Homoscedasticity

  • Plot residuals vs. fitted values to check for "fan" or "cone" shapes.
  • Apply Breusch-Pagan or White’s test:

    
    from statsmodels.stats.diagnostic import het_breuschpagan
    
    bp_test = het_breuschpagan(residuals, X)
    print("Breusch-Pagan p-value:", bp_test[1])
        

  6. Normality

  • Q-Q plot of residuals:

    
    import scipy.stats as stats
    import matplotlib.pyplot as plt
    
    stats.probplot(residuals, dist="norm", plot=plt)
    plt.show()
        

  • Shapiro-Wilk test:

    
    from scipy.stats import shapiro
    stat, p = shapiro(residuals)
    print("Shapiro-Wilk p-value:", p)
        

  ---

  OLS in Quantitative Finance and Interviews

  Understanding OLS assumptions is crucial for quant interviews and real-world financial modeling. Here’s how this knowledge applies at Akuna Capital and similar firms:

  • Strategy Backtesting: OLS is often used to estimate factor exposures or pricing models (e.g., CAPM, Fama-French). Violating assumptions (e.g., autocorrelation in returns) can lead to overfitting or false signals.
  • Risk Modeling: Estimating betas or sensitivities to market factors requires careful attention to heteroscedasticity and multicollinearity.
  • Interview Coding: You may be asked to implement OLS or diagnose a dataset for assumption violations in Python or R.
  • Case Studies: Expect questions such as “What is omitted variable bias? How would you test for heteroscedasticity? What does perfect multicollinearity mean in a portfolio context?”

  Sample Interview Question and Solution

  Question: Suppose you are regressing daily returns of a stock on market returns and find that the residuals show clear volatility clustering. Which OLS assumption is violated? How would you address this?

  Answer: The assumption of homoscedasticity (constant error variance) is violated due to volatility clustering (heteroscedasticity). To address this, use robust standard errors (e.g., White's correction) or model volatility explicitly using GARCH models.

  ---

  Conclusion

  The assumptions of Ordinary Least Squares regression are foundational to quantitative analysis, especially in high-stakes environments like Akuna Capital interviews. Mastery of these assumptions—not just their definitions, but also their practical implications and remedies—will set you apart as a thoughtful, rigorous candidate.

  In summary, OLS requires:

  • Linearity in parameters
  • Random sampling
  • No perfect multicollinearity
  • Zero conditional mean / exogeneity
  • Homoscedasticity
  • Normality of errors (for inference)

  Violating these assumptions can undermine your model’s reliability, so always diagnose, test, and address potential problems. By doing so, you’ll not only ace your Akuna Capital quantitative analyst interview but also build robust, reliable models in your quantitative finance career.

  ---

  References and Further Reading

  • Wooldridge, J. M. (2016). Introductory Econometrics: A Modern Approach. Cengage Learning.
  • James, G., Witten, D., Hastie, T., & Tibshirani, R. (2013). An Introduction to Statistical Learning. Springer.
  • Stock, J. H., & Watson, M. W. (2019). Introduction to Econometrics. Pearson.
  • Akuna Capital Careers: Official Interview Preparation
  • Statsmodels Documentation: Python Statsmodels