abhinavdread/msme-payment-outcome-predictor-lightgbm

MSME Payment Outcome Predictor (LightGBM)

Overview

This model predicts the probabilistic outcome of MSME payment disputes:

• Win
• Settlement
• Escalation to MSEFC

The model outputs calibrated probabilities for each outcome.

Model Architecture

• Algorithm: LightGBM (Gradient Boosted Decision Trees)
• Calibration: Isotonic Regression (CalibratedClassifierCV)
• Preprocessing:
  • OneHotEncoding (categorical features)
  • Numeric features passthrough
  • Class balancing enabled

Input Features

Feature	Type
claim_amount	float
delay_days	float
buyer_type	categorical (govt/private)
contract_present	binary
industry_sector	categorical
claim_imputed	binary
delay_imputed	binary

Output Format

{
  "predicted_label": "win",
  "probabilities": {
    "win": 0.59,
    "settlement": 0.05,
    "escalation": 0.35
  }
}

Performance Metrics

• Primary metric: AUC-ROC (macro) ≈ 0.72
• Balanced Accuracy ≈ 0.63
• F1 Macro ≈ 0.61

Intended Use

• Legal risk scoring
• MSME advisory tools
• Research prototype
• Decision support systems

Limitations

• Based on structured extracted data only
• Does not include full legal document text
• Not intended for judicial automation