Layoff Risk Prediction Model (XGBoost)
Binary classifier estimating the probability that a public company will execute a mass layoff (WARN Act notice) within a 90-day forward window, using fused market, fundamental, sentiment, and macro features.
- Model type: XGBoost (
XGBClassifier),n_estimators=300,max_depth=6,learning_rate=0.05,subsample=0.8,colsample_bytree=0.8 - Task: Binary classification, $Y \in {0, 1}$ — layoff notice within 90 days
- Framework: scikit-learn API / XGBoost, tracked via MLflow
- Source: github.com/hongming111/layoff-risk-prediction-pipeline
Training data
13 features fused from five data domains, resampled/forward-filled onto a shared daily grid:
| Domain | Features |
|---|---|
| Market | close, vol_7d, vol_14d, vol_21d |
| Fundamentals | debt_to_equity, current_ratio, profit_margin |
| Sentiment | sentiment_score, sentiment_score_ma7d, mention_velocity |
| Macro (BLS) | unemployment_rate_total, layoff_rate_total, layoff_rate_tech |
Ground truth labels come from state WARN Act notices (warn-scraper),
entity-resolved to stock tickers via fuzzy company-name matching.
Evaluation
5-fold stratified cross-validation on the training snapshot (497 positive labels):
| Metric | Value |
|---|---|
| ROC-AUC | 0.999 |
| Precision | 0.978 |
| Recall | 0.938 |
| Average Precision | 0.993 |
| Accuracy | 0.992 |
Known limitations
- Suspiciously high CV AUC. These numbers likely reflect the small
current training set and ticker universe rather than a fully validated
production signal —
close(raw price level) is included as a feature, which can make separability easier than it would be in a true forward-looking deployment. Before trusting this for anything beyond a portfolio demo, it needs out-of-time (not just stratified k-fold) validation and a leakage audit. - Small, non-representative ticker universe. Currently evaluated on a handful of large, well-known public companies — not a broad or industry-representative sample.
- Not financial, HR, or legal advice. This is a research/portfolio project. Scores should not be used to make real decisions about any company or its employees.
Intended use
Educational/portfolio demonstration of a multi-modal MLOps pipeline (entity resolution, temporal feature alignment, drift monitoring, and closed-loop prediction-vs-actual evaluation). See the linked GitHub repo for the full architecture, Airflow DAGs, and evaluation loop.