NARFIMA: Neural ARFIMA Forecasting

• NARFIMA is a hybrid forecasting framework that integrates ARFIMA's long-memory fractional differencing with a neural network to capture nonlinear residual dynamics.
• It blends linear ARFIMAx modeling with exogenous inputs and a dedicated neural module to enhance forecasting accuracy across multiple time horizons.
• Empirical results demonstrate that NARFIMA often outperforms traditional ARFIMA and other benchmarks in forecasting tasks such as exchange rate predictions.

Searching arXiv for NARFIMA and closely related ARFIMA/SARFIMA papers to ground the article. NARFIMA denotes Neural AutoRegressive Fractionally Integrated Moving Average, a hybrid forecasting framework that combines the long-memory representation of ARFIMA with the nonlinear approximation power of a neural network, while flexibly incorporating exogenous causal variables (Chakraborty et al., 8 Sep 2025). Within the broader fractionally integrated modeling family, NARFIMA inherits the central ARFIMA mechanism—the fractional differencing operator $(1-B)^d$—used to represent slowly decaying dependence, and extends it by modeling nonlinear residual structure left after an ARFIMA-type fit [(Contreras-Reyes et al., 2012); (Chakraborty et al., 8 Sep 2025)]. In this sense, NARFIMA belongs to the ARFIMA family of long-memory models, but differs from standard linear formulations by explicitly targeting nonlinearity and covariate-driven effects.

1. Definition and position within the fractionally integrated family

The classical ARFIMA$(p,d,q)$ process is defined by

$\Phi(B)y_t=\Theta(B)(1-B)^{-d}\varepsilon_t,$

where $\Phi(B)$ and $\Theta(B)$ are the AR and MA polynomials, $B$ is the backward shift operator, and $(1-B)^{-d}$ is the fractional differencing or fractional integration operator (Contreras-Reyes et al., 2012). The key role of $d$ is to allow fractional differencing, which produces slowly decaying autocorrelations and long-range dependence when $d>0$ (Contreras-Reyes et al., 2012). The same long-memory mechanism appears across related formulations, including nonseasonal ARFIMA, seasonal SARFIMA, and spatial sp-ARFIMA, though the interpretation of the fractional parameter depends on the domain and operator being fractionally powered [(Reisen et al., 2012); (Otto et al., 2023)].

NARFIMA extends this family in two ways. First, it is formulated in an ARFIMAx style, allowing exogenous variables. Second, it adds a neural network residual component to capture nonlinear structure left after ARFIMA fitting (Chakraborty et al., 8 Sep 2025). The resulting logic is a two-stage procedure: fit a linear long-memory model with exogenous predictors, compute residuals, and then feed lagged series values, lagged residuals, and covariates into a neural network to model nonlinearities (Chakraborty et al., 8 Sep 2025). A plausible implication is that NARFIMA should be understood less as a replacement for ARFIMA than as a hybrid architecture built on top of ARFIMA’s fractional dependence structure.

2. Fractional differencing and the long-memory mechanism

The fractional differencing operator is the mathematical core inherited by NARFIMA from ARFIMA. In the ARFIMAx stage of the neural model, it is written as

$$(1 - B)^d = \sum_{v = 0}^{\infty} \frac{\Gamma(v - d)B^v}{\Gamma(-d)\Gamma(v+1)},$$

with $(p,d,q)$0 in the cited NARFIMA application (Chakraborty et al., 8 Sep 2025). In the broader ARFIMA literature, the same operator is expressed via the binomial expansion

$(p,d,q)$1

or, equivalently, through generalized binomial coefficients (Contreras-Reyes et al., 2012). This slow coefficient decay is the mathematical source of long memory (Contreras-Reyes et al., 2012).

The spectral interpretation is equally central. For ARFIMA models, the spectral density is

$(p,d,q)$2

which makes the low-frequency singularity explicit: when $(p,d,q)$3, the spectrum is highly concentrated near frequency $(p,d,q)$4, corresponding to strong low-frequency persistence (Contreras-Reyes et al., 2012). In time-domain terms, the autocovariances decay hyperbolically rather than exponentially [(Reisen et al., 2012); (Contreras-Reyes et al., 2012)]. For general fractionally differenced processes, the low-frequency behavior can also be written as a short-memory spectral factor multiplied by a fractional singularity at the origin, reinforcing that the long-range dependence is driven by the fractional operator itself (Gorst-Rasmussen et al., 2012).

This long-memory interpretation remains the foundation of NARFIMA. The neural extension does not replace fractional integration; rather, it supplements it when residual nonlinear structure remains after ARFIMA-type fitting (Chakraborty et al., 8 Sep 2025).

3. Model architecture and formal specification

The ARFIMAx stage used in NARFIMA is specified as

$(p,d,q)$5

where $(p,d,q)$6 is the backshift operator, $(p,d,q)$7, $(p,d,q)$8 is white noise, $(p,d,q)$9 and $\Phi(B)y_t=\Theta(B)(1-B)^{-d}\varepsilon_t,$0 are AR and MA orders, and $\Phi(B)y_t=\Theta(B)(1-B)^{-d}\varepsilon_t,$1 are coefficients (Chakraborty et al., 8 Sep 2025). The ARFIMAx fit produces forecasts $\Phi(B)y_t=\Theta(B)(1-B)^{-d}\varepsilon_t,$2, from which residuals are computed: $\Phi(B)y_t=\Theta(B)(1-B)^{-d}\varepsilon_t,$3 The cited work emphasizes that these residuals still contain nonlinear structure (Chakraborty et al., 8 Sep 2025).

The neural stage is a single hidden-layer feed-forward network that takes as inputs $\Phi(B)y_t=\Theta(B)(1-B)^{-d}\varepsilon_t,$4 lagged exchange rate values, $\Phi(B)y_t=\Theta(B)(1-B)^{-d}\varepsilon_t,$5 lagged ARFIMAx residuals, and one lag of each of the $\Phi(B)y_t=\Theta(B)(1-B)^{-d}\varepsilon_t,$6 exogenous variables (Chakraborty et al., 8 Sep 2025). Its generic one-step-ahead forecast is

$\Phi(B)y_t=\Theta(B)(1-B)^{-d}\varepsilon_t,$7

With skip connections, the network is written as

$\Phi(B)y_t=\Theta(B)(1-B)^{-d}\varepsilon_t,$8

Here $\Phi(B)y_t=\Theta(B)(1-B)^{-d}\varepsilon_t,$9 is the number of hidden nodes, $\Phi(B)$0 is the activation function, $\Phi(B)$1 are input-to-hidden weights, $\Phi(B)$2 are hidden-to-output weights, $\Phi(B)$3 are skip-connection weights, and $\Phi(B)$4 is a bias (Chakraborty et al., 8 Sep 2025). The hidden layer learns nonlinear interactions, while the skip path preserves linear or autoregressive effects directly (Chakraborty et al., 8 Sep 2025). This suggests that NARFIMA is structurally a “linear + nonlinear” hybrid, with the long-memory linear backbone retained rather than absorbed into a purely neural architecture.

4. Theoretical properties and probabilistic interpretation

The theoretical analysis in the cited NARFIMA work focuses on a simplified NARFIMA$\Phi(B)$5 process,

$\Phi(B)$6

with

$\Phi(B)$7

and its equivalent state-space or Markov form (Chakraborty et al., 8 Sep 2025). The paper proves that the NARFIMA$\Phi(B)$8 process is geometrically ergodic and therefore asymptotically stationary, under assumptions including a stationary ARFIMA residual process with $\Phi(B)$9, a bounded and Lipschitz neural activation, controllability of the linear part, innovations with a strictly positive continuous density and finite variance, and contraction of the skip-connection weights (Chakraborty et al., 8 Sep 2025).

The proof strategy treats the process as a Markov chain on $\Theta(B)$0, establishes forward accessibility, irreducibility, and a Foster–Lyapunov drift condition using

$\Theta(B)$1

and then invokes the standard conclusion that irreducibility plus geometric drift imply geometric ergodicity (Chakraborty et al., 8 Sep 2025). The corollary states that the process is asymptotically stationary (Chakraborty et al., 8 Sep 2025).

These results place NARFIMA in a distinctive position relative to classical ARFIMA. Standard ARFIMA theory emphasizes stationarity, causality, invertibility, spectral density, and asymptotic covariance behavior in linear Gaussian settings (Contreras-Reyes et al., 2012). NARFIMA preserves the fractionally integrated residual foundation but adds nonlinear dynamics under boundedness and contraction conditions (Chakraborty et al., 8 Sep 2025). A plausible implication is that its theoretical guarantees are not generic consequences of fractional differencing alone; they depend critically on how the neural component is constrained.

5. Relation to ARFIMA variants and neighboring model classes

NARFIMA is part of a larger ecosystem of fractionally integrated models, several of which extend ARFIMA in orthogonal directions. Seasonal ARFIMA or SARFIMA introduces a second fractional operator,

$\Theta(B)$2

thereby allowing long memory at both the zero frequency and the seasonal frequency (Reisen et al., 2012). In the cited environmental application, the full model is

$\Theta(B)$3

with GARCH-type innovations used to model time-varying volatility (Reisen et al., 2012). The resulting SARFIMA-GARCH specification differs from standard ARFIMA by adding seasonal fractional integration and conditional heteroscedasticity (Reisen et al., 2012).

Spatial sp-ARFIMA transports the same fractional-integration idea into spatial econometrics by replacing the time backshift with a spatial weights matrix. Its core form,

$\Theta(B)$4

makes $\Theta(B)$5 a control on the range and shape of spatial spillovers rather than temporal long memory in the usual sense (Otto et al., 2023). This shows that “ARFIMA-type” no longer denotes a single temporal model class, but a broader methodology centered on fractional powers of dependence operators.

NARFIMA differs from these extensions in emphasis. SARFIMA-GARCH enriches the linear stochastic structure by seasonality and heteroscedasticity (Reisen et al., 2012). Sp-ARFIMA generalizes the operator from temporal lags to spatial weights (Otto et al., 2023). NARFIMA, by contrast, retains the temporal ARFIMA core but adds a neural residual layer and exogenous inputs (Chakraborty et al., 8 Sep 2025). This suggests that the defining innovation of NARFIMA is not a new fractional operator, but a new way of combining fractional integration with nonlinear function approximation.

6. Empirical applications, performance evidence, and implementation

The principal empirical application described for NARFIMA concerns monthly spot exchange rates for Brazil, Russia, India, and China from January 1997 to October 2023, with exchange rates drawn from FRED and macro covariates from FRED and policyuncertainty.com (Chakraborty et al., 8 Sep 2025). The candidate predictors considered are Global Economic Policy Uncertainty, U.S. Equity Market Volatility, U.S. Monetary Policy Uncertainty, Geopolitical Risk, WTI oil price growth rate, short-term interest rates, short-term interest rate differentials, and CPI inflation and inflation differentials (Chakraborty et al., 8 Sep 2025). A nonlinear Granger causality test is used to select exogenous variables, and the variables found to be consistently causal for BRIC exchange rates are GEPU, US EMV, US MPU, oil price growth, and country-specific short-term IRD (Chakraborty et al., 8 Sep 2025).

The study uses six rolling forecast horizons—1, 3, 6, 12, 24, and 48 months—and evaluates models using MAPE, SMAPE, MAE, MASE, and RMSE (Chakraborty et al., 8 Sep 2025). NARFIMA is compared against 16 benchmark models: Naïve, AR, ARIMAx, ARFIMAx, ETS, SETAR, TBATS, GARCH, BSTSx, ARNNx, DeepAR, NBeatsx, NHiTSx, DLinearx, NLinearx, and TSMixerx (Chakraborty et al., 8 Sep 2025). The reported conclusion is that across most horizons and countries, NARFIMA is the best or second-best model, and often decisively best; it is especially strong for Brazil, dominates at several horizons for Russia, is often best for China, and is less dominant for India, whose series is described as more linear (Chakraborty et al., 8 Sep 2025). Multiple Comparisons with the Best and Murphy diagrams are also reported to support statistical superiority or strong competitiveness across horizons, especially on RMSE, MAE, MAPE, and SMAPE (Chakraborty et al., 8 Sep 2025).

The paper further reports a residual-selection ablation in which ARFIMAx residual feedback is replaced by residuals from ARIMAx, BSTSx, or Naïve, plus a residual-free variant akin to ARNNx; the stated result is that NARFIMA with ARFIMAx residuals performs best (Chakraborty et al., 8 Sep 2025). This directly supports the claim that the fractional long-memory stage is not incidental but central to the hybrid model’s performance.

Implementation is provided in the narfima R package (Chakraborty et al., 8 Sep 2025). The cited two-stage implementation fits ARFIMAx using arfima from the R forecast package and then trains a single-hidden-layer neural network using nnet::nnet (Chakraborty et al., 8 Sep 2025). The hyperparameters $\Theta(B)$6, $\Theta(B)$7, $\Theta(B)$8, and skip are tuned using time-series cross-validation minimizing RMSE, with $\Theta(B)$9 to avoid overfitting; one-step-ahead forecasts come directly from the trained model, while multi-step forecasts are recursive (Chakraborty et al., 8 Sep 2025). The paper also reports that ARFIMAx residual nonlinearity is verified using the Terasvirta test and the BDS test, motivating the neural stage (Chakraborty et al., 8 Sep 2025).

7. Limitations, cautions, and interpretive issues

A major caution about the broader ARFIMA family comes from the analysis of FARIMA brittleness. That work argues that fractionally differenced processes are mathematically atypical among long-range dependent processes because they are extremely close to fractional Gaussian noise in a precise sense (Gorst-Rasmussen et al., 2012). The paper formalizes this through spectral, autocovariance, and variance-time-function results, including

$B$0

and

$B$1

with a bounded remainder in the variance time function (Gorst-Rasmussen et al., 2012). Its central warning is that this closeness property is not robust to additive noise, so conclusions based on unperturbed FARIMA asymptotics may fail when realistic noise is present (Gorst-Rasmussen et al., 2012).

For NARFIMA, this caution is directly relevant because its long-memory backbone is still generated primarily by fractional differencing [(Chakraborty et al., 8 Sep 2025); (Gorst-Rasmussen et al., 2012)]. The brittleness argument does not invalidate NARFIMA, but it does imply that one should not treat fractionally differenced models as generic surrogates for all long-memory behavior (Gorst-Rasmussen et al., 2012). A plausible implication is that the added nonlinear layer in NARFIMA may improve empirical forecasting without automatically resolving the asymptotic sensitivity associated with the fractional differencing core.

A second interpretive issue concerns terminology. The provided sources explicitly define NARFIMA as Neural AutoRegressive Fractionally Integrated Moving Average in the 2025 exchange-rate study (Chakraborty et al., 8 Sep 2025). Earlier sources discuss ARFIMA and related extensions, but do not define “NARFIMA” as a separate model class (Zhou et al., 2013). Accordingly, NARFIMA should be distinguished from the generic idea of nonlinear or augmented ARFIMA-type modeling. In the supplied literature, the term is specifically attached to the hybrid neural–ARFIMA framework with exogenous variables, residual feedback, conformal prediction intervals, and Markov-chain-based stationarity analysis (Chakraborty et al., 8 Sep 2025).

Finally, the empirical literature surrounding ARFIMA variants indicates that different extensions address different departures from the basic linear long-memory model: seasonality and volatility clustering in SARFIMA-GARCH (Reisen et al., 2012), spatial spillover geometry in sp-ARFIMA (Otto et al., 2023), and nonlinear residual structure with exogenous drivers in NARFIMA (Chakraborty et al., 8 Sep 2025). This suggests that NARFIMA is best understood not as the universal next step beyond ARFIMA, but as one specialized response to the joint presence of long memory, nonlinearity, and external covariate effects.