🧠 Predict2Protect – Company Failure Prediction Challenge

This repository contains my solution for the Predict2Protect Machine Learning Competition, where the goal was to predict whether a company is Alive or Failed based on its financial and operational features.

---

🏁 Competition Overview

The challenge involved developing a classification model that predicts a company’s status_label (Alive / Failed).
Data consisted of company-level records, each belonging to a specific Division and MajorGroup.
The evaluation metric was Macro F1-Score, emphasizing balanced performance on both classes.

---

🚀 Approach Summary

1. Data Understanding & EDA

• Checked for missing values, data imbalance, and outliers.
• Verified no company overlap between train and test (leakage prevention).
• Analyzed Division and MajorGroup hierarchies (noting class imbalance).
• Ensured that every company’s status_label is consistent (all rows share the same label).

2. Feature Engineering

• Generated ratio-based and interaction features.
• Aggregated statistics at the company level.
• Applied label encoding for categorical columns.
• Removed redundant or highly correlated features.

3. Model Building

• Implemented LightGBM, CatBoost, XGBoost, and ExtraTrees models.
• Used GroupKFold Cross-Validation (grouped by company_name) to avoid data leakage.
• Optimized Macro F1-Score via threshold tuning.
• Final predictions created using weighted averaging across model outputs.

4. Post-Processing

• Enforced company-level label consistency:

  If any row of a company was failed, the whole company was marked as failed.

• Calibrated thresholds to maximize F1 on validation folds.

---

📊 Results

Metric	Score
Cross-Validation Macro F1	0.63 ± 0.01
Final Leaderboard Macro F1	0.635972
Final Rank	#14 / Top 15 Finalist

---

💬 Key Learnings

• Importance of group-aware validation to prevent leakage.
• Effectiveness of label consistency enforcement at company level.
• Balancing precision and recall through threshold optimization.
• Robust pipeline design beats leaderboard hacks in the long run.

---

🔮 Future Work

• Advanced hyperparameter optimization (Optuna/Bayesian tuning).
• Seed aggregation and model ensembling for variance reduction.
• Hierarchical encoding of Division → MajorGroup relationships.

---

🧰 Tech Stack

• Python, Pandas, NumPy
• LightGBM, CatBoost, XGBoost, Scikit-Learn
• Matplotlib, Seaborn, tqdm

---

🧑‍💻 Author

Devkanth Ravi
Top-15 Finalist, Predict2Protect 2025
📫 Connect: LinkedIn | Kaggle

---

“A consistent, explainable model always outlasts a lucky leaderboard spike.”