ARFIMA stands for AutoRegressive Fractionally Integrated Moving Average.
Stata fits ARFIMA models.
ARFIMA concerns long-memory processes. Long-memory processes are stationary processes whose autocorrelation functions decay slowly. The ARFIMA model provides a parsimonious parameterization of long-memory processes that nests the ARMA (autoregressive moving-average) model, which is widely used for short-memory processes. Long-memory processes can be obtained by fractionally integrating short-memory processes. The ARFIMA model does that. The ARFIMA model also generalizes the ARIMA model by allowing for fractional degrees of integration.
Below we analyze yearly data on the widths of the rings of a tree. This one tree survived 5,000 years! In any case, larger widths represent good years for the tree and narrower widths represent harsh years. Below we estimate the parameters of an ARFIMA model with the fractional difference parameter and a constant. With only those parameters, we will account for the long memory of the process.
. webuse campito, clear
(Campito Mnt. tree ring data from 3435BC to 1969AD)
. arfima width, nolog
ARFIMA regression
Sample: -3435 thru 1969 Number of obs = 5,405
Wald chi2(1) = 1864.44
Log likelihood = -18907.279 Prob > chi2 = 0.0000
| width | Coefficient Std. err. z P>|z| [95% conf. interval] | |
| width | ||
| _cons | 44.01432 9.174317 4.80 0.000 26.03299 61.99565 | |
| ARFIMA | ||
| d | .4468887 .0103497 43.18 0.000 .4266038 .4671737 | |
| /sigma2 | 63.92927 1.229754 51.99 0.000 61.519 66.33955 | |
Below we plot the original series and the fractionally differenced component of the predicted series, which reflects the short-memory component of the process:
;)
Explore more time-series features in Stata.
