📊 ERP-OFDI Model - Chinese Investments

🌟 About the Project

Pipeline to analyze Chinese outbound investment (OFDI) and the impact of the ERP (Ecological Redlines Policy) implemented in 2017.

Data source: Dr. Derek Scissors, PhD - Stanford University / American Enterprise Institute

Documentation

• Pipeline and Implementation:

  • docsPython.pdf
  • docsR.pdf They include full details about implementation, method, and technical specs.

• Metrics Interpretation:

  • itpt.pdf It covers COFDI-ERP analysis with regression, classification, time series, causality, and other metrics.

🔄 Project Structure

📁 data/ → Stage 1: Data Load

Files: data.csv (transactions 2005-2024) + naturalearth_lowres.zip (maps)

• Process: Load CSV → Clean data → Convert numeric columns → Filter valid years
• Result: DataFrame organized and ready for analysis

📁 code/ → Main Run

Files: code.py (orchestration) + requirements.txt (dependencies)

• Command: python .\code\code.py
• Function: Runs all 11 steps automatically (Python)

📁 results/ → Outputs

🎨 r_style_plots/ → Stage 3: Visualizations

Generated plots:

• r_style_world_maps.png → World maps colored by investment
• r_style_investment_by_sector.png → Sector evolution in 3 phases
• r_style_greenfield_donuts.png → Donut charts Greenfield vs M&A
• python_style_boxplot_greenfield.png → Comparative statistical analysis

📁 eda/ → Stage 4: Exploratory Analysis

Files:

• summary_statistics.csv → Means, medians, standard deviations
• eda_target_distribution_log_vs_raw.png → Value distribution
• eda_total_investment_over_time.png → Full time series
• eda_h1_sector_shift_summary.csv → Sector shifts after ERP

📁 phase_analysis/ → Stage 5: Phase Analysis GG

Files:

• summary_interaction_ols_by_phase.csv → Interaction effects
• summary_mediation_ols_by_phase.csv → Mediation effects
• Phases: GG 1.0 (2005-2012) | GG 2.0 (2013-2016) | GG 3.0+ERP (2017-2024)

📁 models/regression/ → Stage 6: Predictive Models

Files:

• regression_models_summary_otimizada.csv → Performance XGBoost vs LightGBM
• Metrics: MAPE | Log RMSE
• Target: Predict log_Valor_USD (investment value)

📁 models/classification/ → Stage 7: Binary Classification

Files:

• classification_models_summary_otimizada.csv → AUC-ROC | Accuracy
• Target: Alvo_Adaptativo (1=high value, 0=low value)
• Models: XGBoost, LightGBM, CatBoost with hyperparameter tuning

📁 models/timeseries/ → Stage 8: Time Series

Files:

• arima_forecast.png → ARIMA forecasts
• prophet_forecast.png → Facebook Prophet forecasts
• prophet_forecast_data.csv → 5-year projection data
• Goal: Predict future investment trends

📁 models/causal/ → Stage 9: Causal Analysis

Files:

• dml_summary_att.csv → ERP causal effect via Double Machine Learning
• markov_prob.png → Regime change probabilities
• markov_switching_summary.txt → Regime model details
• Question: Did the ERP policy cause changes in investments?

📁 models/shap/ → Stage 10: Interpretability

Files:

• shap_summary_plot.png → Impact of each variable on the model
• shap_importance_plot.png → Feature importance
• Key variables: valor_roll_mean_3 (moving average) | Sector | Region

---

🛠️ Stage 2: Feature Engineering

Created features:

• valor_roll_mean_2/3 → Moving averages over 2/3 periods
• post_ERP → Post-2017 flag
• Fase_GG → Three Going Global phases
• Alvo_Binario → High/low value classification
• policy_interaction → Time × policy interaction

---

📋 Summary of the 11 Automatic Steps (Python)

1. Data Load → Cleaning and validation
2. Feature Engineering → Create ML variables
3. Visualizations → 20+ professional plots (via code.py and code.R)
4. EDA → Statistical exploratory analysis
5. GG Phase Analysis → OLS models by period
6. Regression → Predict investment value (XGBoost/LightGBM)
7. Classification → Identify high value (AUC-ROC)
8. Time Series → ARIMA & Prophet
9. Causal Models → Double ML & Markov Switching
10. SHAP → Model interpretability
11. Consolidation → Save all results

---

🚀 Execution

# 1. Install Python dependencies
python -m pip install -r requirements.txt

# 2. Run Python pipeline
python ./code/code.py

# 3. Run R code for advanced plots
Rscript ./code/code.R

# 4. View results
#    Python: ./results/
#    R: ./results/R/