Foreign Exchange Risk Analytics
Advanced Value at Risk computation plus backtesting.
Multi-method estimation, method comparison, and walk-forward backtesting with rolling recalibration.
About
Cloud-based platform for quantitative risk analysis of financial portfolios. Configure confidence level, time horizon, and observation period, then choose from multiple estimation methods ranging from classical statistical approaches to machine learning. Selected methods can be compared side-by-side or validated through walk-forward backtesting with a comprehensive battery of statistical tests.
Historical VaR
Non-parametric approach using the empirical return distribution. No distributional assumptions required. 1-day VaR is estimated directly from the observed percentile of historical log returns. Recommended when return distributions exhibit fat tails or skewness.
Parametric VaR
Variance-covariance method assuming normally distributed returns. VaR is derived analytically from the return standard deviation and the inverse normal quantile. Computationally efficient with closed-form Expected Shortfall.
EWMA
Exponentially Weighted Moving Average volatility model. Assigns exponentially decaying weights to past observations, capturing volatility clustering more effectively than equal-weight estimation. Configurable decay factor (lambda).
GARCH(1,1)
Generalized Autoregressive Conditional Heteroskedasticity model. Explicitly models time-varying volatility through autoregressive dynamics. Supports normal, Student-t, and skewed-t error distributions for flexible tail behavior.
Cornish-Fisher
Adjusts the normal quantile for skewness and excess kurtosis. Captures fat-tail effects without requiring a full distributional model beyond the first four moments of the return distribution.
Method Comparison
Simultaneous execution of the selected VaR methods on the same data window. Identifies the most and least conservative estimate and reports the spread across methodologies.
LSTM Neural Network
This model estimates Value at Risk directly using LSTM quantile regression and pinball loss. It is trained specifically for the confidence level and VaR horizon selected by the user. Daily returns remain the underlying inputs, while each time step contains five features: the sequence mean and the first four Chebyshev-polynomial transformations of centered, robustly scaled returns. The input sequence length is set to twice the selected horizon, subject to a minimum of 50 and a maximum of 252 trading days: it is 50 days for horizons from 1 to 25 days, twice the horizon for horizons from 26 to 126 days, and 252 days for horizons from 127 to 252 days. This provides at least 50 days of context, expands the context with the forecast horizon, and caps it at approximately one trading year. The network consists of one 8-unit LSTM layer, a 16-unit ReLU layer with Layer Normalization and dropout, and one quantile output. Multi-day VaR is predicted directly from the future cumulative return over the selected horizon; square-root-of-time scaling is not used.
Walk-Forward Backtest
Rolling-window recalibration backtest with configurable horizon and window size. LSTM is trained once through the Cutoff Date and then reused with updated input sequences. Compares predicted VaR against non-overlapping realized horizon returns over the test period and reports coverage, independence, and loss diagnostics.