Conditional beta using LASSO adds value

We know that betas are not stable. They are time varying and may change based on fundamental factors associated with the firm as well as changing risk across the business cycle, yet we do not seem to focus on conditional betas. In the new paper, "Conditional Betas", the authors focus on comparing several econometric and machine-learning method to test asset pricing models and market anomalies. If we better measure or control beta, we should see fewer market anomalies and an increase in explanatory power for asset models. These tests will allow us to measure the relative valued-added of choosing better machine-learning technics. 

The comparisons shows that firm-level LASSO method provide improvement over more traditional econometric techniques. The LASSO method is a technique that add a penalty term to a linear regression that forces or encourages to some coefficient to be zero and focuses on the most important values. The penalty is one the absolute value of the slope. (Ridge regression will have a penalty based on the squared value of the slope.) If we penalize the marginal coefficients, we will reduce overfitting and make the overall regression easier to interpret. This will be helpful when there are a lot of potential predictors.

A LASSO regression is easy to interpret and generate and when applied to panel data, it generates coefficients for the three and four factor models that make intuitive sense.