[Paper Review] Why FARIMA Models are Brittle

Anders Gorst‐Rasmussen, Darryl Veitch+1 more|arXiv (Cornell University)|Mar 28, 2012

Financial Risk and Volatility Modeling18 references3 citations

TL;DR

This paper demonstrates that FARIMA processes are structurally brittle due to their extreme similarity to fractional Gaussian noise (fGn), exhibiting ultra-rapid convergence to fGn under aggregation. This closeness is destroyed by even small additive noise, undermining FARIMA's suitability for modeling real-world long-range dependent signals where robustness and diversity in LRD behavior are essential.

ABSTRACT

The FARIMA models, which have long-range-dependence (LRD), are widely used in many areas. Through deriving a precise characterisation of the spectrum, autocovariance function, and variance time function, we show that this family is very atypical among LRD processes, being extremely close to the fractional Gaussian noise in a precise sense. Furthermore, we show that this closeness property is not robust to additive noise. We argue that the use of FARIMA, and more generally fractionally differenced time series, should be reassessed in some contexts, in particular when convergence rate under rescaling is important and noise is expected.

Motivation & Objective

• To investigate the structural stability and representativeness of FARIMA processes as models for long-range dependent (LRD) time series.
• To assess whether FARIMA processes are typical representatives of the broader class of LRD processes.
• To evaluate the impact of additive noise on the convergence behavior of FARIMA processes toward fGn.
• To challenge the widespread use of FARIMA models in simulations and estimator testing, particularly under small sample sizes and noisy conditions.

Proposed method

• Derives a precise spectral characterization of FARIMA processes and compares it to that of fractional Gaussian noise (fGn) using Fourier analysis.
• Uses the variance time function (VTF) to quantify the convergence of FARIMA processes to fGn under aggregation, showing near-identical behavior.
• Applies asymptotic analysis and bounds on Fourier coefficients to prove that the difference between FARIMA and fGn VTFs decays as $ O(n^{eta - eta'}) $, indicating ultra-fast convergence.
• Employs the double integration operator $ \mathbf{I} $ to transform the autocovariance function (ACVF) into the VTF, enabling clearer analysis of convergence to self-similar limits.
• Introduces and proves lemmas on the decay of $ |T_n^j| $ and bounds on $ f_\alpha(x,y) $, which are critical for establishing convergence rates.
• Provides numerical illustrations to demonstrate brittleness: even minimal additive noise drastically alters convergence behavior.

Experimental results

Research questions

• RQ1 How close is the LRD behavior of FARIMA processes to that of fractional Gaussian noise (fGn) in terms of spectral and variance time function (VTF) structure?
• RQ2 What is the rate of convergence of FARIMA processes to fGn under aggregation, and how does this compare to other LRD processes?
• RQ3 Why is the FARIMA family structurally unstable under additive noise, and how does this affect its use in modeling real-world signals?
• RQ4 To what extent do FARIMA models represent a diverse class of LRD processes, or are they effectively limited to behavior near fGn?
• RQ5 How does the brittleness of FARIMA models impact the validity of results derived from them in Monte Carlo simulations of Hurst parameter estimators?

Key findings

• FARIMA processes are extremely close to fGn in terms of spectrum, autocovariance, and variance time function (VTF), with the difference decaying as $ O(n^{eta - eta'}) $, indicating ultra-rapid convergence under aggregation.
• The convergence of FARIMA to fGn is so fast that it offers no meaningful diversity beyond fGn in terms of LRD behavior, making it atypical among LRD processes.
• Additive noise—even of low intensity—disrupts the near-identical behavior with fGn, changing the convergence rate and rendering FARIMA models structurally unstable or 'brittle'.
• The VTF formulation reveals that FARIMA processes converge to fGn faster than any other known LRD process, highlighting their lack of robustness.
• The results imply that FARIMA models should be reassessed in contexts where convergence rate under rescaling or noise robustness is critical.
• Numerical illustrations confirm that the addition of even mild additive noise significantly alters the convergence dynamics, validating the theoretical brittleness.