[Paper Review] A General Framework for Prediction in Time Series Models

Eric Beutner, Alexander Heinemann+1 more|arXiv (Cornell University)|Feb 5, 2019

Financial Risk and Volatility ModelingEconomics, Econometrics and Finance25 references3 citations

TL;DR

This paper proposes a general framework for prediction in time series models that formally verifies high-level regularity conditions for popular models like ARMA, GARCH, and their extensions. It establishes the asymptotic validity of conditional confidence intervals under sample-splitting, bridging theoretical results with practical applications in econometrics and risk management.

ABSTRACT

In this paper we propose a general framework to analyze prediction in time series models and show how a wide class of popular time series models satisfies this framework. We postulate a set of high-level assumptions, and formally verify these assumptions for the aforementioned time series models. Our framework coincides with that of Beutner et al. (2019, arXiv:1710.00643) who establish the validity of conditional confidence intervals for predictions made in this framework. The current paper therefore complements the results in Beutner et al. (2019, arXiv:1710.00643) by providing practically relevant applications of their theory.

Motivation & Objective

• To address the fundamental challenge in time series prediction: balancing conditioning on past data with accounting for parameter uncertainty.
• To formalize a general framework that validates the use of sample-splitting as a realistic alternative to the unrealistic assumption of two independent processes.
• To demonstrate that widely used models—ARMA(1,1), GARCH(1,1), T-GARCH(1,1), and extensions—satisfy the high-level assumptions required for conditional inference.
• To extend the theoretical results of Beutner et al. (2019) by providing concrete, practical applications to real-world time series models.
• To show that conditional risk measures such as Value-at-Risk and Expected Shortfall can be embedded within the same framework.

Proposed method

• The framework defines prediction objects as functions of both the parameter vector and the infinite past of the time series, expressed as $ \psi_{T+1} = \psi(X_T, X_{T-1}, \ldots; \theta_0) $.
• It introduces a sample-splitting estimator by replacing the unknown parameter $ \theta_0 $ with an estimator $ \hat{\theta}(\mathbf{X}_{1:T_E}) $ based on a non-overlapping estimation sample.
• The method relies on verifying a set of high-level assumptions: estimator convergence, differentiability of the prediction function, bounded gradient, and bounded Hessian in a neighborhood of the true parameter.
• Theoretical validity is established via asymptotic equivalence between the sample-split approach and the two-independent-processs assumption, relying on mixing and weak dependence conditions.
• The framework is applied to ARMA(1,1) and GARCH-type models, including E-GARCH, N-GARCH, GJR-GARCH, and Q-GARCH, by verifying the assumptions case by case.
• Conditional risk measures like VaR and ES are mapped into the framework by expressing them as functions of past observations and parameters, enabling inference under the same theoretical structure.

Experimental results

Research questions

• RQ1 Can the theoretical framework of Beutner et al. (2019) for conditional confidence intervals be practically applied to standard time series models?
• RQ2 Do ARMA(1,1) and GARCH(1,1) models satisfy the high-level regularity conditions required for asymptotic validity of sample-splitting-based inference?
• RQ3 Can extensions of GARCH models, such as T-GARCH, E-GARCH, and Q-GARCH, be embedded within the same general framework?
• RQ4 To what extent can conditional risk measures like Value-at-Risk and Expected Shortfall be treated as prediction objects under this framework?
• RQ5 Are there GARCH-type models—such as FI-GARCH or FIE-GARCH—that cannot be accommodated due to long memory or non-mixing properties?

Key findings

• The ARMA(1,1) model with drift satisfies all required high-level assumptions, enabling valid conditional inference for its conditional mean prediction.
• The GARCH(1,1) and T-GARCH(1,1) models satisfy the framework's assumptions, validating the use of sample-splitting for conditional variance and volatility forecasts.
• Extensions such as E-GARCH(1,1), N-GARCH(1,1), GJR-GARCH(1,1), and Q-GARCH(1,1) are formally shown to be embeddable in the framework, with verification of the necessary regularity conditions.
• Conditional Value-at-Risk (VaR) and Expected Shortfall (ES) in the T-GARCH(1,1) model are expressed as functions of past observations and parameters, placing them within the general prediction framework.
• Models with long memory, such as FI-GARCH and FIE-GARCH, cannot be accommodated under the current framework due to failure of standard mixing conditions.
• The paper confirms that sample-splitting asymptotically replicates the behavior of two independent processes, validating its use in practice despite the theoretical challenges of parameter uncertainty.