Machine Learning & Econometrics: Capital Structure and Wage Prediction

Two empirical projects from the Machine Learning & Econometrics course in Fordham University's MS in Quantitative Finance. Both pull their data programmatically from the source, so nothing is downloaded by hand. Each one runs end to end in a single notebook.

	Question	Data	Methods	Write-up
1	What determines how much debt a firm carries?	Compustat via WRDS API, 36,718 firm-years	Pooled OLS, industry/year and firm fixed effects, clustered SEs	Report · PDF · Notebook
2	How well can personal characteristics predict wages?	2020 ACS via IPUMS API, 875,856 workers	OLS, forward stepwise, Lasso, Ridge, random forest	Report · PDF · Notebook

Key findings

Capital structure. Larger firms, firms with tangible assets and firms in high-leverage industries borrow more. Profitable and cash-rich firms borrow less, consistent with the pecking-order theory. Pooled OLS explains 21–25% of leverage. Within-firm changes explain only 3%, because leverage is highly persistent.

Wage prediction. A random forest gives the best out-of-sample fit (test R² 0.595 vs. 0.558 for OLS) by capturing interactions between hours, education and occupation. Lasso and Ridge do not improve on OLS, because with 160k observations and 154 features cross-validation chooses essentially no penalty.

Repository layout

notebooks/
  01_capital_structure.ipynb   WRDS query -> sample -> descriptives -> OLS / FE models
  02_wage_prediction.ipynb     IPUMS extract -> sample -> 5 models -> evaluation
reports/
  01_capital_structure_report.md / .pdf
  02_wage_prediction_report.md / .pdf
  figures/                     charts produced by the notebooks
data/                          created locally; licensed data is never committed

Reproducing the results

git clone https://github.com/SGhuman123/Fordham-ML-Econometrics-Project.git
cd Fordham-ML-Econometrics-Project
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
cp .env.example .env            # then fill in your credentials
jupyter lab notebooks/

Credentials (both free for academic users):

• WRDS: requires a WRDS account with Compustat access. Set WRDS_USERNAME in your environment and store the password once with python -c "import wrds; wrds.Connection().create_pgpass_file()".
• IPUMS: create an API key at https://account.ipums.org/api_keys and set IPUMS_API_KEY in .env. The extract takes a few minutes to build on the IPUMS servers.

On the first run, each notebook downloads its data to data/raw/ and caches it for later runs. Notebook 1 runs in under a minute, and notebook 2 in about two minutes on a laptop. All random steps are seeded.

Data licences. Compustat data are licensed through WRDS, and IPUMS terms prohibit redistributing the microdata, so data/ is excluded from the repository. If you reuse the ACS extract, please cite IPUMS USA (Ruggles et al., IPUMS USA, Minneapolis, MN: IPUMS).

Tools

Python · pandas · scikit-learn · statsmodels · linearmodels · matplotlib · WRDS API · IPUMS API