chile-fintech-systemic-risk

Polyglot architecture for systemic risk analysis of the Chilean financial market (IPSA, Central Bank rates, credit risk). Each language was chosen for the task it solves best, not for portfolio completeness.

Code and method Full README

This page shows the project's results. The methodology, the design decisions and the limitations are documented in the repository's README.

`/etl` — what already runs

Chile equity price and realized volatility, animated — The animated version draws both series at the real data's pace (subsampled to ~45 frames from the 2,511 trading days) with a floating label tracking the current value on each line — a quick way to spot where volatility spikes, while the static chart below remains the reference for detailed reading.

Chile equity price and realized volatility — The animated version draws both series at the real data's pace (subsampled to ~45 frames from the 2,511 trading days) with a floating label tracking the current value on each line — a quick way to spot where volatility spikes, while the static chart below remains the reference for detailed reading.

`/ml_predictions` — what already runs

SHAP feature importance — `dti` and `num_prior_delinquencies` account for most of the mean |SHAP value| — exactly the two variables that dominate the logistic function used to generate the synthetic defaults, so this chart doubles as a sanity check that the explainability pipeline recovers real signal, not noise. The PD-score distribution shows the (partial, consistent with a 0.62 AUC) separation between applicants who actually defaulted and those who didn't.

PD score distribution — `dti` and `num_prior_delinquencies` account for most of the mean |SHAP value| — exactly the two variables that dominate the logistic function used to generate the synthetic defaults, so this chart doubles as a sanity check that the explainability pipeline recovers real signal, not noise. The PD-score distribution shows the (partial, consistent with a 0.62 AUC) separation between applicants who actually defaulted and those who didn't.

LSTM vs. majority-class baseline — The LSTM bar sits below the baseline bar — not decorative, it's the visual evidence for the honest-negative finding: always predicting the more frequent class beats the trained model here.

`/quant_analytics` — what already runs

Credit and volatility clusters in feature space

`/core_engine` — what already runs

Simulated PnL distribution with VaR/ES — Histogram of a 20,000-path subsample of 1-day PnL paths (the VaR/ES metrics themselves are still computed over the full million). The vertical lines mark where the 99% VaR and 99% Expected Shortfall fall on that distribution — the left tail is exactly what both metrics are measuring.