Cross-Sectional Factor Model Estimation and Diagnostics

Worked Solution

How to Think About It: This is the bread and butter of quantitative equity research -- the Fama-MacBeth two-pass approach and its diagnostics. The key idea: each day you run a cross-sectional regression of that day's stock returns on their factor exposures to extract implied factor returns. Then you treat the resulting time series of factor returns as data and ask whether each factor earns a statistically significant premium. The subtlety is that a single day's cross-sectional regression is noisy (hundreds of stocks, but the $R^2$ is typically low), so you need to aggregate across many days and be careful about how you compute standard errors.

Key Insight: The cross-sectional regression gives you one $\hat{\beta}_t$ per day. Statistical significance comes from the time-series behavior of the $\hat{\beta}_t$ sequence, not from any single day's regression.

The Method:

*Part (i): Estimation and $t$-statistics*

1. Each day $t$, run the cross-sectional OLS regression of $r_t$ on $X_t$:

$$\hat{\beta}_t = (X_t^\top X_t)^{-1} X_t^\top r_t$$

This gives you a $K$-vector of estimated factor returns for that day.

1. Collect the time series $\{\hat{\beta}_t\}_{t=1}^T$ over $T$ trading days.

1. For each factor $k$, compute the sample mean and standard error:

$$\bar{\beta}_k = \frac{1}{T} \sum_{t=1}^T \hat{\beta}_{k,t}, \quad \text{SE}(\bar{\beta}_k) = \frac{\hat{\sigma}_k}{\sqrt{T}}$$

where $\hat{\sigma}_k$ is the sample standard deviation of $\{\hat{\beta}_{k,t}\}$.

1. The $t$-statistic is:

$$t_k = \frac{\bar{\beta}_k}{\text{SE}(\bar{\beta}_k)} = \frac{\bar{\beta}_k \sqrt{T}}{\hat{\sigma}_k}$$

1. If the daily factor returns are serially correlated, use Newey-West standard errors with an appropriate lag window instead of the plain $\hat{\sigma}_k / \sqrt{T}$.

*Part (ii): Multicollinearity and instability diagnostics*

• Multicollinearity detection:
• Compute the condition number of $X_t^\top X_t$ each day. A condition number above 30 signals problematic collinearity.
• Equivalently, compute variance inflation factors (VIFs) for each factor. A VIF above 5-10 indicates that the factor's exposure is nearly a linear combination of others.
• Track the eigenvalue spectrum of $X_t^\top X_t$ over time. If the smallest eigenvalue is close to zero relative to the largest, certain factor return estimates will be unstable.

• Unstable loadings detection:
• Run rolling-window regressions (e.g., 60-day windows) and plot the time series of each $\hat{\beta}_{k,t}$. Large swings or regime changes in the factor return series indicate instability.
• Apply a formal structural break test (e.g., CUSUM or Bai-Perron) to each factor's return series.
• Compute the rolling correlation between factors. If correlations between factors change substantially over time, the model may be misspecified or the factors may be redundant in certain regimes.

*Part (iii): Nonlinearity*

• Structural drawback: A linear model assumes that a one-unit increase in, say, Value exposure always has the same marginal effect on returns, regardless of the level of Value or interactions with other factors. In practice, factor returns often exhibit threshold effects (e.g., Value works differently in the top decile vs. bottom decile) and interaction effects (e.g., cheap stocks with strong momentum behave differently from cheap stocks with weak momentum).

• Testing for nonlinearity:
• Add squared factor exposure terms $X_{k,t}^2$ or interaction terms $X_{j,t} \cdot X_{k,t}$ to the cross-sectional regression and test whether these additional coefficients are jointly significant using an $F$-test.
• Alternatively, sort stocks into quantile buckets by factor exposure and compare the actual mean return per bucket to the linear model's prediction. If the relationship is linear, the residuals across buckets should be roughly zero; systematic curvature indicates nonlinearity.
• A Ramsey RESET test applied to the cross-sectional regression is another standard diagnostic: regress residuals on powers of the fitted values and test for significance.

Practical Considerations:

• Because exposures $X_t$ are standardized, the estimated $\hat{\beta}_{k,t}$ has a natural interpretation: it is the return spread (in the same units as $r_t$) between a stock one standard deviation above the mean on factor $k$ and a stock at the mean, holding other factors constant.
• Outlier stocks can dominate a single day's cross-sectional regression. Winsorizing returns or exposures at the 1st/99th percentile each day is standard practice.
• In production, GLS (using an estimated residual covariance) often replaces OLS to improve efficiency, but OLS is the starting point.

Answer: Estimate factor returns daily via cross-sectional OLS, then compute Fama-MacBeth $t$-statistics from the time series of daily estimates (with Newey-West corrections if needed). Diagnose multicollinearity with VIFs or condition numbers and instability with rolling-window analysis or structural break tests. The main linear limitation is the assumption of constant marginal effects; test for nonlinearity by adding polynomial or interaction terms and checking their joint significance.