🎓 Final Project

📋 Project Overview

🔬 Track 1: Empirical Replication & Extension

🎯 Six Paper Options

# Download Option A: econml package
pip install econml
from econml.datasets import fetch_401k
data = fetch_401k()

# Download Option B: Federal Reserve
# https://www.federalreserve.gov/econres/scfindex.htm

import numpy as np
from sklearn.linear_model import LassoCV
from sklearn.model_selection import KFold

def dml_plr(Y, D, X, ml_g, ml_m, n_folds=5):
    """
    Double Machine Learning for Partially Linear Model
    Y = θ*D + g(X) + ε
    
    TODO: 
    1. Split data into K folds
    2. Train ml_g (outcome model) and ml_m (propensity model)
       on training set, predict on test set
    3. Compute residuals: Ỹ = Y - ĝ(X), D̃ = D - m̂(X)
    4. Estimate θ = mean(Ỹ * D̃) / mean(D̃²)
    5. Compute standard errors
    
    Returns: theta, standard_error
    """
    n = len(Y)
    Y_hat = np.zeros(n)
    D_hat = np.zeros(n)
    
    kf = KFold(n_splits=n_folds, shuffle=True, random_state=42)
    
    for train_idx, test_idx in kf.split(X):
        # YOUR CODE HERE
        pass
    
    return theta, se

# Download Option A: Dehejia-Wahba
# https://users.nber.org/~rdehejia/data/.nswdata2.html

# Download Option B: causaldata package
pip install causaldata
from causaldata import lalonde

class CausalTree:
    """
    Key difference from standard tree:
    - Standard: minimize prediction MSE
    - Causal: maximize |τ_left - τ_right| (treatment effect heterogeneity)
    """
    
    def fit(self, X, Y, D):
        """
        X: covariates, Y: outcome, D: treatment
        
        TODO: Implement recursive partitioning that maximizes
        treatment effect differences between leaves
        """
        pass
    
    def _treatment_effect(self, Y, D):
        """τ = E[Y|D=1] - E[Y|D=0]"""
        # YOUR CODE HERE
        pass

# Option A: Card-Krueger replication
# https://github.com/tyleransom/DiD-example

# Option B: QCEW data
# https://www.bls.gov/cew/

def compute_att_gt(df, group, time):
    """
    Compute ATT(g,t) following Callaway & Sant'Anna (2021)
    
    Group g: units first treated at time g
    Compare to never-treated or not-yet-treated
    
    ATT(g,t) = [E[Y_t|g] - E[Y_g|g]] - [E[Y_t|control] - E[Y_g|control]]
    
    TODO: Implement group-time ATT calculation
    """
    pass

def event_study_plot(df):
    """
    Plot dynamic treatment effects by event time
    X-axis: Time relative to treatment (-5 to +5)
    Y-axis: Treatment effect with confidence intervals
    """
    pass

# Option A: synthdid package
pip install synthdid
from synthdid import get_data
california = get_data('california_prop99')

# Option B: Ortega database
# https://github.com/NiclasOrtG/the-synthetic-control-group

from scipy.optimize import minimize

def synthetic_control(Y_target, Y_donors):
    """
    Find optimal weights w that minimize ||Y_target - Y_donors @ w||²
    Subject to: sum(w) = 1, w >= 0 (convex combination)
    
    Y_target: (T_pre,) pre-treatment outcomes for treated unit
    Y_donors: (T_pre, J) pre-treatment outcomes for donor pool
    
    Returns: optimal weights (J,)
    """
    def objective(w):
        return np.sum((Y_target - Y_donors.T @ w) ** 2)
    
    # TODO: Use scipy.optimize.minimize with constraints
    pass

# Option A: econml
from econml.datasets import fetch_jtpa

# Option B: MDRC (requires application)
# https://www.mdrc.org/

def doubly_robust_scores(Y, D, X):
    """
    Compute doubly robust scores:
    Γ = μ₁(X) - μ₀(X) + D(Y - μ₁(X))/e(X) - (1-D)(Y - μ₀(X))/(1-e(X))
    
    where μ₁, μ₀ are outcome models and e(X) is propensity score
    
    TODO: Implement with cross-fitting
    """
    pass

def learn_policy(X, dr_scores, budget=0.5):
    """
    Learn policy π(X) that maximizes welfare subject to budget
    
    max E[π(X) * Γ]  s.t. E[π(X)] ≤ budget
    
    Solution: Treat if DR score > threshold
    """
    pass

# Option A: EPU index (monthly)
# https://www.policyuncertainty.com/

# Option B: News articles (requires ProQuest/Factiva)
# Python: newspaper3k, beautifulsoup

# Option C: GDELT project
# https://www.gdeltproject.org/

UNCERTAINTY_TERMS = ['uncertain', 'uncertainty', 'risk', 'volatile']
POLICY_TERMS = ['policy', 'regulation', 'legislation', 'government']
ECONOMIC_TERMS = ['economy', 'growth', 'recession', 'inflation']

def count_epu_articles(articles):
    """
    Article counts as EPU if it contains:
    - At least one uncertainty term AND
    - At least one policy term AND
    - At least one economic term
    
    EPU Index = (EPU articles / Total articles) × Normalization
    """
    pass

📚 Track 2: Literature Review

Getting Started

1. Pick a Specific Domain: Narrow your focus to a concrete application. For example:
   • ML in Policy: Search for "Causal ML" or "Algorithmic Fairness" (e.g., Mullainathan & Spiess, 2017).
   • ML in Finance: Look into "Predictive Asset Pricing" or "Credit Scoring" (e.g., Gu, Kelly, & Xiu, 2020).
2. The Snowball Search: Use the bibliography of your "seed paper" to find foundational work (Backward Search) and use Google Scholar's "Cited by" to find recent 2025–2026 developments (Forward Search).
3. Categorize by Approach: Group papers by whether they focus on Prediction (E[Y|X]) or Causal Inference (e.g., Double ML).
4. Identify the Gap: Note what is missing. For instance, does a finance paper lack a structural interpretation, or does a policy paper ignore selection bias?
5. Synthesize: Don't just list papers. Compare them—explain how newer ML methods improve upon traditional econometric benchmarks.

Report Structure (Track 2)

Grading Rubric (Track 2)

📊 Report Requirements

Suggested Structure

\documentclass[11pt]{article}
\usepackage[margin=1in]{geometry}
\usepackage{graphicx, amsmath, natbib}

\title{Replication and Extension of [Paper Title]}
\author{Your Name\\ECON6083}
\date{\today}

\begin{document}

\maketitle

\begin{abstract}
Brief summary of paper, replication approach, key findings, and extension. 
(150-200 words)
\end{abstract}

\section{Introduction} (1-2 pages)
Motivation, research question, overview of replication and extension

\section{Literature and Background} (1-2 pages)
Related literature, theoretical framework, institutional details

\section{Data} (1-2 pages)
Data sources, sample construction, summary statistics (Table 1)

\section{Empirical Strategy} (2-3 pages)
Model specification, identification assumptions, estimation details

\section{Replication Results} (3-4 pages)
Main results, robustness checks, comparison to original paper

\section{Extension} (2-3 pages)
Motivation, methodology, results, interpretation

\section{Conclusion} (0.5-1 page)
Summary, limitations, future research

\bibliographystyle{aer}
\bibliography{references}

\appendix
\section{Additional Results}
\section{Code}

\end{document}

Grading Rubric (Track 1)

🛠️ Technical Resources

Required Packages

pip install numpy pandas matplotlib scikit-learn scipy statsmodels

Optional Packages

pip install econml              # DML, Causal Forests
pip install synthdid            # Synthetic Control
pip install linearmodels        # Panel data
pip install transformers        # BERT for text
pip install spacy nltk jieba    # Text processing

📅 Timeline