ARFIMA Long-Memory Models

ARIMA's differencing order $d$ is normally a whole number: 0 (already stationary), 1 (difference once to remove a trend/unit root), or occasionally 2. But this forces a false choice on series like realized volatility, whose autocorrelations decay far more slowly than any stationary ARMA model allows, yet clearly don't have a genuine unit root either — volatility doesn't wander off to infinity the way a true random walk does; it always eventually reverts, just very slowly. ARFIMA (Autoregressive Fractionally Integrated Moving Average) resolves this by letting $d$ be a fraction, giving a whole continuum of memory strength between ordinary short-memory ARMA ($d=0$) and a full unit root ($d=1$).

An analogy: a dimmer switch instead of an on/off switch

Ordinary ARIMA treats "does this series need differencing" like a light switch: either fully off (stationary, $d=0$, memory fades fast) or fully on (a unit root, $d=1$, memory never fades — every past shock has a permanent effect). Real long-memory series need a dimmer switch, not a toggle: memory that fades, but far more slowly than $d=0$ allows and without ever reaching the "never fades" extreme of $d=1$. A fractional value like $d=0.3$ gives exactly that intermediate setting — meaningful, lingering influence from the distant past, but an influence that does eventually decay to nothing, unlike a true unit root.

The model, one piece at a time

ARFIMA extends the ARIMA equation by applying the differencing operator to a fractional power:

$$\phi(L)(1-L)^d X_t = \theta(L)\varepsilon_t, \qquad d \in (-0.5, 0.5) \text{ for a stationary long-memory process},$$

where $\phi(L)$ and $\theta(L)$ are the usual AR and MA lag polynomials, $(1-L)$ is the differencing operator, and $d$ is now allowed to be a fraction rather than restricted to an integer. Expanding $(1-L)^d$ via its binomial series produces an infinite-order AR (or, equivalently, MA) representation whose coefficients decay as a slow power law rather than the fast geometric decay of an ordinary ARMA process — this power-law decay is exactly the mathematical signature of long memory. In plain English: fractional differencing partially "differences out" a series' persistence, by an amount tuned continuously via $d$ rather than forced to be either zero or complete — $d=0$ recovers ordinary short-memory ARMA, $d\to1$ approaches a full unit root, and any fraction in between captures long memory that decays, just very slowly. The parameter $d$ is directly related to the Hurst exponent via $d = H - 0.5$ (see The Hurst Exponent and Long Memory), giving two independent ways to estimate essentially the same underlying memory strength.

Worked example 1: converting a Hurst estimate to d

A realized-volatility series was estimated in a companion analysis to have Hurst exponent $H\approx0.67$ via rescaled-range analysis. Converting to the fractional differencing parameter: $d = H - 0.5 = 0.67-0.5=0.17$. Fitting an ARFIMA(0, 0.17, 0) model to the same series directly via maximum likelihood gives a nearly identical estimate, $\hat{d}=0.19$ — the two independent methods (R/S analysis and direct ARFIMA estimation) agree closely, which is itself useful confirmation that the long memory detected is real rather than an artifact of one particular estimation method.

Worked example 2: how slowly d = 0.2 decays versus AR(1)

Compare the autocorrelation decay of an AR(1) process with $\phi=0.5$ against a fractionally integrated process with $d=0.2$. The AR(1) autocorrelation at lag $k$ is $\rho_k = 0.5^k$: at $k=20$, that's $0.5^{20}\approx0.00000095$ — utterly negligible. The ARFIMA process's autocorrelation instead decays as a power law, roughly $\rho_k \sim k^{2d-1} = k^{-0.6}$: at $k=20$, that's $20^{-0.6}\approx0.19$ — still meaningfully positive, nearly a fifth of its starting strength, twenty periods out. Even at $k=100$, the ARFIMA autocorrelation is $100^{-0.6}\approx0.063$ — still non-negligible, while the AR(1)'s has long since vanished to a number with dozens of zeros after the decimal point. This is the practical difference: a long-memory model keeps a small but real amount of predictive information from shocks that a short-memory model has completely forgotten.

What this means in practice

ARFIMA-family models are standard for volatility forecasting (realized volatility, implied volatility term structure) precisely because these series show the classic long-memory signature: strong but slowly decaying autocorrelation. A volatility forecast that uses an ordinary short-memory model will systematically underweight information from further in the past than it should, while one that (incorrectly) assumes a unit root will overstate how permanently a volatility shock persists; ARFIMA sits deliberately between the two.