Challenge
The study examined whether financial fundamentals could statistically explain company market capitalisation while remaining interpretable enough to support meaningful conclusions and avoid the opacity of a purely predictive model.
My contribution
With a co-author, I built and tested OLS, Ridge, LASSO and Elastic Net models using Bloomberg data from 181 OMXSGI companies between 2010 and 2019.
Approach
Prepared and log-transformed the dataset, winsorizing observations at the first and ninety-ninth percentiles across variables including EBIT, assets, debt, ROIC, cash, EPS growth and earnings.
Tested heteroskedasticity and multicollinearity using Breusch-Pagan, robust standard errors and VIF diagnostics.
Compared regularized models with a reduced OLS specification, balancing predictive fit against interpretability.
Outcome
LASSO achieved the lowest prediction error, while the reduced OLS model was selected for interpretation and achieved a holdout MSE of 0.387; earnings was the strongest driver, EPS growth was counterintuitively negative and debt dropped out as insignificant.
Deliverables
- Cleaned Bloomberg dataset
- OLS regression model
- Regularized model comparison
- Academic thesis
