Fractional Cointegration: Concepts & Methods

• Fractional cointegration is an extension of cointegration theory that applies to fractionally integrated processes, allowing reduction in persistence via linear combinations.
• It employs advanced estimation methods such as the local Whittle estimator, narrow-band GLS, and eigenanalysis to accurately measure long-memory dynamics.
• Applications in finance, macroeconomics, and functional data analysis demonstrate its utility in elucidating equilibrium adjustments and long-run structural relationships.

Searching arXiv for recent and foundational papers on fractional cointegration. Fractional cointegration is the extension of cointegration theory to fractionally integrated processes. If a $p$-dimensional process $X_t$ is $I(d)$ and there exists a nonzero vector $\beta$ such that $\beta'X_t\in I(d-b)$ with $b>0$, then $X_t$ is cointegrated, with $b$ measuring the reduction in persistence. The classical $I(1)$-$I(0)$ case is the special case $X_t$0; in the fractional setting, both $X_t$1 and $X_t$2 may be non-integer, so cointegrating errors may remain long-memory rather than becoming exactly $X_t$3 (Hualde et al., 2022).

1. Fractional integration and the basic definition

Fractional cointegration rests on fractional integration. For a scalar process $X_t$4, the fractional difference operator is

$X_t$5

with coefficients defined through the binomial expansion. A process is $X_t$6 if $X_t$7 is $X_t$8. In the standard stationary region, $X_t$9, the spectral density behaves near the origin as $I(d)$0, and autocovariances decay hyperbolically as $I(d)$1; $I(d)$2 corresponds to long memory, $I(d)$3 to short memory, and $I(d)$4 to antipersistence (Hualde et al., 2022).

For nonstationary fractional processes, several papers adopt the type II definition, based on zero pre-sample values and a truncated operator $I(d)$5. In this formulation, a process $I(d)$6 is written as

$I(d)$7

which allows a unified treatment of $I(d)$8 and $I(d)$9, and makes the inverse operator well defined for nonstationary cases (Hartl, 2020).

In multivariate form, cointegration means reduction of integration order under linear combination. If the components of $\beta$0 share order $\beta$1, a vector $\beta$2 is cointegrating when $\beta$3 for some $\beta$4. The reduction need not reach stationarity. This broadens the long-run equilibrium interpretation: deviations from equilibrium may be stationary, but they may also be less persistent yet still fractionally integrated. The chapter-length overview of the field emphasizes that this flexibility is one of the central distinctions between fractional and classical cointegration (Hualde et al., 2022).

2. Structural representations and sources of cointegration

A canonical representation is the triangular system

$\beta$5

which makes the cointegrating relation explicit: $\beta$6 is $\beta$7, while the equilibrium error is $\beta$8. More general multivariate systems allow multiple cointegrating subspaces and multiple strengths of order reduction, especially when different blocks of the system load on different memory orders (Hualde et al., 2022).

A major development is the latent-components representation

$\beta$9

where $\beta'X_t\in I(d-b)$0 collects latent fractionally integrated components and $\beta'X_t\in I(d-b)$1 is short-memory. In multivariate fractional components analysis, the components $\beta'X_t\in I(d-b)$2 can have different orders $\beta'X_t\in I(d-b)$3, ordered as $\beta'X_t\in I(d-b)$4. Cointegration then arises because linear combinations of $\beta'X_t\in I(d-b)$5 can eliminate the higher-order latent components, lowering the effective integration order from $\beta'X_t\in I(d-b)$6 to $\beta'X_t\in I(d-b)$7, $\beta'X_t\in I(d-b)$8, and so on; the corresponding orthogonal complements define nested cointegration subspaces (Hartl et al., 2018).

This logic underlies fractional factor models. In high-dimensional macroeconomics, observable series are modeled as sums of common fractionally integrated factors and stationary components. The Dynamic Orthogonal Fractional Components model separates $\beta'X_t\in I(d-b)$9 purely fractionally integrated factors from $b>0$0 stationary autoregressive factors, so that the long-run structure is concentrated in a low-dimensional fractional subspace and cointegration is generated by directions that cancel the shared long-memory factors (Hartl, 2020).

Unobserved-components formulations provide another direct route to fractional cointegration. In a multivariate common-trend model,

$b>0$1

all observable variables inherit the order $b>0$2 of the scalar latent trend $b>0$3, while any linear combination orthogonal to $b>0$4 is $b>0$5. With $b>0$6 observables and one common fractional trend, the model has $b>0$7 fractional cointegration relations by construction (Hartl et al., 2020).

The same principles extend beyond finite-dimensional vectors. For curve time series $b>0$8, a projection $b>0$9 may define a finite-dimensional nonstationary subspace $X_t$0 and an orthogonal stationary subspace $X_t$1. Fractional cointegration then means that $X_t$2 for some projection $X_t$3 and some $X_t$4, even though the dominant directions are $X_t$5 with $X_t$6 (Seo et al., 2022).

3. Estimation and statistical inference

Estimation of fractional integration orders is often semiparametric. Two central procedures are the log-periodogram estimator and the local Whittle estimator, both based on the low-frequency behavior $X_t$7. Under standard bandwidth conditions, the log-periodogram estimator is asymptotically normal with variance $X_t$8, while the local Whittle estimator is asymptotically normal with variance $X_t$9 and is asymptotically more efficient (Hualde et al., 2022).

For cointegrating vectors, ordinary least squares is not generally satisfactory. In stationary fractional cointegration, OLS is inconsistent because regressors and cointegrating errors remain long-memory and correlated. Narrow-band least squares therefore focuses on low frequencies only, using averaged periodograms over a shrinking neighborhood of zero frequency. Weighted narrow-band GLS and local Whittle procedures refine this idea by weighting frequencies according to the memory of the residual process; in the bivariate stationary case, the local Whittle version is asymptotically more efficient than unweighted NBLS (Hualde et al., 2022).

Parametric and quasi-parametric methods become particularly important when the goal is efficient inference on both memory parameters and cointegrating vectors. In strong fractional cointegration, feasible GLS and pseudo-maximum-likelihood estimators yield $b$0-consistent and asymptotically mixed Gaussian estimates of $b$1. In weak fractional cointegration, the same objective is harder because the cointegration gap is small and residual persistence remains strong, but feasible root-$b$2 procedures are available under additional structure (Hualde et al., 2022).

State-space methods provide a complementary route when fractional cointegration is embedded in latent-variable models. Exact type II state-space representations are high-dimensional, so several papers replace fractional filters by low-order ARMA approximations and then estimate by Kalman filtering and maximum likelihood. In multivariate unobserved fractional components models, finite-order ARMA approximations outperform simple autoregressive or moving-average truncation in approximation quality and computational cost, and support EM-based likelihood maximization in potentially nonstationary systems with cointegration (Hartl et al., 2018).

These ideas carry directly into applications. Fractional factor models for macro forecasting use a two-stage procedure: principal components for preliminary factors, followed by state-space estimation with the Kalman filter and an EM algorithm. For the DOFC model, the fractional and stationary factor spaces are separated using the Chen–Hurvich method based on eigenvectors of averaged periodogram matrices at low frequencies, after which dynamic orthogonal components rotations and likelihood-based ARFIMA estimation are applied (Hartl, 2020).

Wavelet methods provide an alternative when nonstationarity is prominent. In the implied-realized volatility setting, the wavelet band least squares estimator

$b$3

estimates the long-run fractional cointegrating coefficient using only low-frequency wavelet scales. Because wavelets localize in time and frequency and can handle $b$4 without explicit differencing, WBLS was proposed as a practical analogue of Fourier-domain narrow-band methods for nonstationary long-memory volatility series (Barunik et al., 2012).

4. Rank determination, testing, and identification

Testing in fractional cointegration addresses both the integration orders and the dimension of the cointegrating space. Parametric score-type approaches generalize the Johansen-style rank problem by testing whether the rank of the long-run matrix is $b$5, given a specified memory order. Semiparametric alternatives instead exploit the low-frequency rank of the spectral density matrix, allowing cointegration rank to be identified through the eigenvalues of suitably normalized spectral estimates (Hualde et al., 2022).

A particularly direct method is eigenanalysis of a non-negative definite matrix built from sample autocovariances,

$b$6

with eigenvectors corresponding to the smallest eigenvalues estimating the cointegration space. The methodology was originally developed for unknown and possibly heterogeneous integer orders, and then extended to the fractional case. In the fractional setting, consistency of the estimated cointegration space and of a rank estimator based on transformed eigencomponents’ autocorrelations is established under fixed dimension $b$7 (Zhang et al., 2015).

Variance-ratio procedures offer another nonparametric route. Nielsen’s test constructs generalized eigenvalues from covariance operators before and after applying a fractional summation operator; scaled “small” generalized eigenvalues correspond to nonstationary directions, while large ones correspond to stationary directions. This idea has been extended from finite-dimensional time series to curve-valued processes, where it determines the dimensions of the nonstationary and stationary subspaces and then supports local Whittle estimation of the long-memory parameters in each subspace (Seo et al., 2022).

Empirical work on real convergence in Spain combines two testing strategies. One is Robinson’s nonparametric test statistic $b$8, applied sequentially following Hualde’s algorithm to infer cointegration rank from stationary growth rates. The other is Nielsen’s variance-ratio test applied to nonstationary log levels. In that application, both approaches found no evidence of fractional cointegration among the 17 regional GDP series; for Nielsen’s test with $b$9, $I(1)$0, $I(1)$1, and $I(1)$2, the statistic $I(1)$3 was below the 5% critical value $I(1)$4 (Kamal et al., 2023).

The FCVAR model provides a likelihood-based testing framework within a full multivariate system. In

$I(1)$5

the parameters $I(1)$6 and $I(1)$7 govern the integration order of the system and the cointegration gap, respectively. For $I(1)$8, likelihood-ratio tests for rank and restrictions are asymptotically $I(1)$9; for $I(0)$0, the asymptotic limit of rank tests is a non-pivotal functional of fractional Brownian motion, and critical values are obtained by plug-in simulation (Hualde et al., 2022).

5. Empirical domains

Financial volatility is one of the clearest empirical domains. Implied volatility and realized volatility are both modeled as long-memory $I(0)$1 processes, and the relation

$I(0)$2

is interpreted as fractional cointegration when the variance risk premium $I(0)$3 is less persistent than the individual series. Using S&P 500 and DAX option data, low-frequency wavelet coherence showed that dependence comes solely from the lower frequencies of the spectra, motivating WBLS estimation on scales $I(0)$4. In the long run, volatility inferred from corridor implied volatility provided an unbiased forecast of realized volatility, whereas model-free implied volatility remained biased (Barunik et al., 2012).

Macroeconomic forecasting provides a high-dimensional illustration. Using 112 monthly US macro series from FRED-MD over 1960–2016, fractional factor models were compared with AR, principal-component, PCAR, and factor-augmented error-correction benchmarks. Over 112 series and horizons $I(0)$5, the six fractional factor variants delivered the best MSPE in 987 cases, compared with 357 for the benchmark models. DOFC-KF, which explicitly separates fractional long-run and stationary short-run factors, most frequently delivered the best forecasts at short and medium horizons; DFFD-KF became more dominant at longer horizons (Hartl, 2020).

Fractional cointegration also appears in financial covariance systems. In realized covariance matrices of six US stocks transformed into a 21-dimensional series, the preferred Dynamic Orthogonal Fractional Components specification had $I(0)$6 nonstationary fractional components with $I(0)$7, $I(0)$8 stationary long-memory components with $I(0)$9, and $X_t$00 short-memory AR(1) components. Since all 21 observables load on the two nonstationary factors, the implied cointegration rank is $X_t$01, with relations reducing persistence from about $X_t$02 to at most $X_t$03 (Hartl et al., 2018).

Regional growth applications show the opposite outcome. In Spanish real GDP per capita across 17 autonomous communities, exact local Whittle estimates on levels put $X_t$04 between 1 and 1.4, while local Whittle estimates on growth rates lay between about 0.34 and 0.50. Yet neither the Robinson–Hualde strategy on growth rates nor Nielsen’s variance-ratio test on levels found any fractional cointegration, ruling out stochastic convergence in that sample (Kamal et al., 2023).

Functional data applications extend the concept beyond vectors. For Swedish male and female log-mortality curves, local Whittle estimates were $X_t$05 and $X_t$06, and the variance-ratio procedure estimated $X_t$07 nonstationary functional trends for both genders. For Canadian zero-coupon yield curves, local Whittle estimation again indicated $X_t$08, while the variance-ratio test suggested $X_t$09 nonstationary stochastic trends (Seo et al., 2022).

Recent work broadens the field still further. Geometric functionals of long-memory sphere-cross-time random fields, such as excursion area and boundary length, can be fractionally cointegrated across threshold levels, so that appropriate linear combinations have shorter memory than each component. In continuous-time mortality modeling, long-range dependent national mortality and insurer-specific mortality are tied together through a cointegrated Volterra system, and the paper argues that cointegration is what brings the national long-range dependence into the insurer’s model when direct mixed-fractional-Brownian specifications suffer an identification problem (Caponera et al., 14 Jul 2025, Chiu et al., 12 Mar 2025).

6. Interpretation, controversies, and directions of development

A recurring theme is that fractional cointegration is not merely a relaxation of the $X_t$10-$X_t$11 paradigm but a different view of equilibrium adjustment. When $X_t$12, deviations from equilibrium are stationary but still long-memory; when $X_t$13, equilibrium errors remain nonstationary but are less persistent than the original variables. This suggests a spectrum of mean reversion rather than a binary distinction between transitory and permanent deviations (Hualde et al., 2022).

Several methodological cautions recur in the literature. Structural breaks, level shifts, and regime changes can mimic long memory, making it difficult to distinguish genuine fractional dependence from nonstationary deterministic contamination. The overview chapter emphasizes that long memory versus structural breaks may be largely a matter of labeling in some settings, and robust procedures are therefore essential before interpreting slow decay as evidence of fractional integration or cointegration (Hualde et al., 2022).

Mis-specification of persistence is especially consequential in latent-component models. In unobserved-components trend-cycle decompositions, imposing an $X_t$14 trend when the true order exceeds one produces an upward-biased signal-to-noise ratio, excessively volatile trends, and cycles that are too weak. The fractional trend-cycle literature argues that this distortion would carry over to multivariate systems and therefore to inference on common trends and cointegrating relations (Hartl et al., 2020).

High-dimensional macro factor models reinforce a related point: over-differencing can destroy long-run information, including cointegration relations. Fractional differencing tuned to each series’ estimated order may retain more long-run information than integer differencing, while explicit factor representations of common fractional trends make the cointegration structure interpretable in a low-dimensional space (Hartl, 2020).

Recent developments indicate that the scope of the subject continues to expand. Functional cointegration, geometric-functional cointegration, and continuous-time Volterra systems all preserve the same essential idea: dominant long-memory directions generate observable persistence, and cointegration corresponds to linear operations that remove some of those directions and reduce the order of dependence. This suggests that fractional cointegration is best understood not as a single model class but as a general principle governing equilibrium structure in systems with long memory (Seo et al., 2022).