Introduction
Generalized AutoRegressive Conditional Heteroskedasticity (GARCH) models are fundamental tools in financial econometrics, used extensively to capture "volatility clustering"—the tendency for large price movements to be followed by further large movements. While point estimation via Quasi-Maximum Likelihood (QML) is standard, constructing reliable confidence regions remains a challenge. Traditional methods rely on asymptotic theory, which assumes large sample sizes and specific noise properties (e.g., finite fourth moments). In practice, financial data is often heavy-tailed and samples are finite, causing asymptotic approximations to fail and leading to inaccurate risk assessments.
The ScoPe Methodology
To address these limitations, this research introduces ScoPe (Score Permutation), a distribution-free method for constructing exact confidence regions in finite samples. ScoPe is inspired by Sign-Perturbed Sums (SPS) but adapted for the nonlinear structure of GARCH models.
The method operates on the following principles:
- Score Function Analysis: Instead of the likelihood surface, ScoPe utilizes the gradient of the log-likelihood (the score function).
- Residual Permutation: The model is inverted to extract residuals, which are then randomly permuted to create multiple "alternative" score trajectories.
- Rank-Based Inference: By comparing the original score to the permuted scores, the algorithm assigns a rank.
- Exact Coverage: Because any permutation of independent and identically distributed (i.i.d.) noise is equally likely, the probability that the true parameter falls within a specific rank threshold is mathematically exact, regardless of the underlying noise distribution.
Unlike standard ellipsoidal regions, ScoPe regions can be non-convex or even disconnected, more accurately reflecting the complex geometry of the likelihood surface in finite samples.
Experimental Evaluation
The performance of ScoPe was evaluated using both simulated and empirical datasets:
- Simulations: GARCH(1,1) processes were generated with Gaussian, Logistic, and Laplacian noise to test robustness against heavy-tailed distributions.
- Real-World Data: Daily closing prices from 2014 for the Nasdaq 100, S&P 500, and FTSE 100 (252 observations each) were used to represent typical finite-sample financial scenarios.
Results and Discussion
Comparative analysis against established methods—including Asymptotic Ellipsoids (Asym.Ell.), Residual Bootstrap (Res.Boots.), and Likelihood Ratio Bootstrap (LR.Boots.)—revealed significant advantages for ScoPe.
As shown in the table above, ScoPe maintains superior empirical coverage. In simulations with 100 observations and a target 90% confidence level, ScoPe achieved approximately 91% coverage, whereas asymptotic methods often underperformed, particularly under Logistic noise. Furthermore, ScoPe demonstrated high efficiency, producing confidence regions with smaller relative areas than LR bootstrap methods while providing stronger theoretical guarantees.
On real-world indices, ScoPe's regions were comparable in size to residual bootstrap methods but offered the critical advantage of being distribution-free and valid for non-stationary processes.
Conclusion
ScoPe provides a robust framework for financial risk management by delivering exact confidence regions without requiring large-sample approximations or specific noise distributions. While computationally more intensive than asymptotic formulas and producing more complex geometric regions, its reliability in heavy-tailed, finite-sample environments makes it a valuable tool for modern volatility forecasting.
Summary updated