zriri9a/financial-stress-early-warning-nlp-pipeline

Early-warning index for systemic risk & financial stress in Morocco: market-based CISS + CamemBERT news-sentiment index + RNN forecasting. # Financial-stress early-warning index (CISS + French NLP sentiment) An early-warning system for **systemic risk and financial stress in Morocco**. It combines a quantitative composite stress index (CISS) built from market data with a **qualitative sentiment index** extracted from French-language economic news via CamemBERT, then forecasts stress ahead of time with recurrent neural networks. Built as an engineering end-of-studies project, in collaboration with a financial-market regulator in Morocco (name withheld). **Author:** @zriri9a (co-authored) **Period:** 2025-01 → 2025-06 **Status:** shipped (defended) — research prototype, not a production service **Stack:** Python, pandas, numpy, statsmodels, scikit-learn (+ SHAP), TensorFlow/Keras (LSTM/SimpleRNN), Hugging Face Transformers (CamemBERT), `arch` (GARCH), hmmlearn / Markov regime-switching, matplotlib/seaborn ## What's interesting here - **Two complementary stress signals, compared head-to-head.** A market-based CISS and a news-sentiment index are built on the same timeline, and the notebooks show the sentiment index turning *before* the market index at COVID-19 (2020) and the 2022 energy shock — the whole point of the project. - **A GDP-regressed weighting scheme for CISS.** Instead of equal-weighting the four market sub-indices (money, bond, equity, FX), sub-index weights come from a regression on quarterly GDP (`02_ciss_regression_gdp_weighting.ipynb`). The "CISS Reg" reacts more sharply and meaningfully to real macro shocks than the equal-weighted version. - **French financial NLP that actually works.** CamemBERT sentiment scoring of economic articles, plus a lexicon fallback (FEEL + a Kaggle FR lexicon), with an alert threshold on the sentiment score. - **A full forecasting ladder.** Classical (ARIMA / linear / Markov-switching) → tree/ML (RandomForest + SHAP, GARCH volatility) → deep learning (LSTM/RNN). The RNN wins on the non-linear dynamics. ## How to run it The data files are **not** included …