Portfolio Value at Risk (VaR) Engine

อ่านภาษาไทย

Computes portfolio VaR using three methods -- Parametric (Variance-Covariance), Historical, and Monte Carlo -- finds the portfolio weights that minimize risk or minimize tail loss via constrained optimization, and runs a rolling 1-year backtest (Kupiec POF test) to evaluate how each model performs over time.

Structure

var_risk_engine/
├── src/
│   ├── config.py            # all tunable constants (tickers, dates, weights, horizon, etc.)
│   ├── data.py               # load_close_prices, compute_returns
│   ├── portfolio_stats.py    # scale_mu, scale_cov (horizon scaling)
│   ├── var_models.py         # performance_portfolio, historical_var, monte_carlo_var
│   ├── optimize.py           # find_min_risk_portfolio, find_min_loss_portfolio
│   ├── backtest.py           # rolling windows, calculate_var_metrics, kupiec_test
│   └── report.py             # show_portfolio, show_var_comparison
├── notebooks/
│   └── var_risk_engine.ipynb # narrative notebook that imports from src/
├── requirements.txt
├── README.md                 # this file (English)
└── README.th.md              # Thai version

Calculation logic lives in src/; the notebook is the narrative/display layer. This split means the logic can be unit-tested or reused in a script independently of the write-up, and the notebook stays readable instead of mixing implementation details with explanation.

Setup

python -m venv .venv
source .venv/bin/activate       # Windows: .venv\Scripts\activate
pip install -r requirements.txt
jupyter notebook notebooks/var_risk_engine.ipynb

Known limitation -- read before trusting the backtest numbers

src/backtest.py has a documented look-ahead bias: the rolling window used to estimate VaR for a given day currently includes that same day's return, instead of stopping the day before it. This means the model has partially "seen" the outcome it's being scored against in the Kupiec test, which likely makes all three models look more accurate than they actually are.

This has not been fixed in the current version — see the docstring at the top of src/backtest.py for the exact indexing mechanism and the one-line fix that resolves it. Treat Section 11-12 results in the notebook as illustrative of the workflow, not as validated model performance, until this is addressed.

A second, lower-severity caveat: with N_ROLLING_WINDOWS = 252 and 95% confidence, the expected number of VaR exceptions is only ~12-13. The Kupiec test has limited statistical power at this sample size — a "Pass" result is weak evidence of model adequacy, not proof.

Method summary

Method	Assumption	Approach
Parametric	Returns are normally distributed	return + z * risk, z from the normal quantile
Historical	No distributional assumption	Empirical quantile of realized portfolio returns
Monte Carlo	GBM with correlated shocks	Cholesky-correlated simulation, quantile of simulated portfolio

VaR is stored as a negative number throughout (more negative = worse loss) — this affects how find_min_loss_portfolio is read: it maximizes VaR (pushes it toward 0), which is the same as minimizing the magnitude of the tail loss.