Can AI Predict a Company Will Go Bankrupt Before It Happens?
Banks and investors lose billions when companies collapse without warning. I built an AI early-warning system with 93.67% AUC accuracy using ensemble stacking — here is how it works.
Imagine you lend your friend $10,000. Before you hand it over, you want to know: is this person going to pay me back? You look at their salary, their other debts, whether they pay bills on time. You are doing a basic credit check. Now imagine doing that same check — but for 6,819 companies, using 94 different financial signals, all at once. That is what this project does.
Why This Matters
When a company goes bankrupt, banks lose loan principal, suppliers lose unpaid invoices, and investors lose their stake. An early warning system that flags distress months in advance gives everyone time to act — restructure debt, reduce exposure, or exit gracefully.
The 94 Financial Warning Signs 🚨
We engineered 94 features from raw balance sheet and income statement data. Some are direct ratios, others are derived. Think of them like the vital signs a doctor checks: individually, one bad number might not mean much. Together, a pattern of bad numbers tells a clear story.
Feature Categories Used in the Model
📊 LIQUIDITY RATIOS
Current Ratio, Quick Ratio, Cash Ratio
→ Can the company pay short-term bills?
💳 LEVERAGE RATIOS
Debt-to-Equity, Interest Coverage
→ How much debt vs. assets?
📈 PROFITABILITY RATIOS
ROA, ROE, Profit Margin
→ Is the company actually making money?
⚡ EFFICIENCY RATIOS
Asset Turnover, Inventory Days
→ How well does it use what it has?
🔄 TREND FEATURES
Year-over-year changes in all the above
→ Is the situation getting better or worse?Why One Model Is Not Enough: Ensemble Stacking 🏗️
A single expert opinion is good. Five expert opinions are better. Ensemble stacking works the same way. We train four different models on the same data, then use a fifth model — the 'meta-learner' — to combine their predictions into one final answer. Each model has different strengths: XGBoost handles outliers well, LightGBM is fast and catches subtle patterns, CatBoost excels at categorical features, and Logistic Regression provides a linear baseline.
Feature Engineering (94 signals from raw financials)
Raw financial statements have hundreds of line items. We compute ratios, year-over-year changes, and interaction terms to create features that are predictive of distress.
Base Models (4 algorithms trained in parallel)
XGBoost, LightGBM, CatBoost, and Logistic Regression are each trained using 5-fold cross-validation. Their out-of-fold predictions become inputs to the next level.
Meta-Learner (combines all 4 predictions)
A final Logistic Regression model learns the optimal way to combine the four base model outputs. It learns which model to trust more in different scenarios.
SHAP Explanations (why this prediction?)
SHAP (SHapley Additive exPlanations) shows exactly which features pushed the prediction toward bankruptcy or safety — essential for regulatory transparency and analyst trust.
What is AUC-ROC? (Simply Explained)
AUC = 0.5 means the model is as useful as a coin flip. AUC = 1.0 means it is perfect. Our AUC of 0.9367 means: if you randomly pick one bankrupt company and one healthy company, our model ranks the bankrupt one as riskier 93.67% of the time. That is a very strong result for real-world financial data.
0.9367
AUC-ROC score
6,819
Companies in dataset
94
Engineered features
+4pp
Gain from stacking
An important lesson from this project: the ensemble (AUC 0.923) did not outperform XGBoost alone (AUC 0.9367). With severe class imbalance at 30:1, a single well-tuned gradient booster often beats a stacked ensemble. The ensemble adds value in precision/recall trade-offs, not raw AUC. The broader takeaway: always compare ensemble vs best single model — ensemble wins are not guaranteed.