Financial Risk Horizons

An Interactive Contrast: Vanilla Conformal vs. Nonstationary SSRN 2688367

Simulation Engine Active

Simulation Parameters

Timeline Steps (\(T\)) 200
Forecast Horizon Start 120
ROA Coefficient of Variation (\(\tau\)) 0.25
ROA Mean Drift (Nonstationarity) 0.0003
Volatility scales with the drifting mean
Coverage Level (\(1 - \alpha\)) 95%
Rolling Window (\(W\)) 21
Equity / Assets (\(EA_t\)) 0.100

Test Set Metrics

Vanilla Conformal Static Interval
Observed Coverage: --
Mean Bound Width: --
Nonstationary Model Adaptive Interval
Observed Coverage: --
Mean Prediction Width: --
Mean \(Z_7^k\): --

Visual Risk Horizon Canvas

Shaded areas depict the predicted uncertainty boundary (\(1 - \alpha\))

Methodological Contrast & Insights

Vanilla Conformal Prediction

Vanilla conformal frameworks assume **exchangeability** of historical calibration samples and future observations. Because of this, the interval boundary is computed as a static threshold calibrated globally from the training split.

Interval Width is Static:
\([ \hat{\mu} - q, \hat{\mu} + q ]\)

Key Vulnerability: When data exhibits drift or structural variance modifications (nonstationarity), VCP's static window will fail, resulting in severe **undercoverage** during high-volatility regimes.

Nonstationary Z-Score Framework

Based on **Mare, Moreira, & Rossi (SSRN 2688367)**, \(Z_7^k\) fits linear trends over overlapping rolling windows, detrends each midpoint realization, and estimates a constant coefficient of variation from the resulting residuals.

Prediction Band is Dynamic:
\(PI_t=f(t)\mp t_{k-2}^{-1}((1+\alpha)/2) \cdot \sqrt{\hat{\sigma}_t^2+\epsilon_t^2}\) \(\epsilon_t=SE[f(t)]\) \(\hat{\sigma}_t=|\bar{\tau}f(t)|\)

Benefit: The forecast mean follows the latest local trend, while heteroskedasticity is represented by \(\bar{\tau}=(1+1/(4N))|s/m|\) and \(\hat{\sigma}_t=|\bar{\tau}f(t)|\).