Keyboard shortcuts

Press ← or → to navigate between chapters

Press S or / to search in the book

Press ? to show this help

Press Esc to hide this help

Statscraft 0.15.0, release v0.15.0

ARIMA

Seasonal ARIMA models with regressors by exact maximum likelihood, with forecasts.

Syntax

ARIMA [VARIABLES=] series [WITH regressors]
      [/MODEL=[(p,d,q)[(sp,sd,sq)[period]]] [CONSTANT | NOCONSTANT] [NOLOG | LG10 | LN]]
      [/P=n | (lags)] [/D=n] [/Q=n | (lags)] [/SP=...] [/SD=n] [/SQ=...]
      [/AR=values] [/MA=values] [/SAR=values] [/SMA=values] [/REG=values] [/CON=value]
      [/MXITER=10] [/MXLAMB=1E9] [/SSQPCT=0.001] [/PAREPS=0.001] [/CINPCT=95]
      [/FORECAST=EXACT | CLS | AUTOINIT] [/APPLY [SPECIFICATIONS | INITIAL | FIT]].
The model is phi(B) PHI(B^s) (w_t - mu) = theta(B) THETA(B^s) a_t, w_t the series less the
regressors times their coefficients, differenced d times and over the period sd times;
the polynomials are 1 - c_1 B - ..., so a positive MA coefficient means a negative
weight on the past error. P, Q, SP and SQ name the lags to fit (n for 1 to n, a list in
parentheses for those lags only) and win over MODEL. LOG is LG10; logs need values above
0. CONSTANT and the confidence level default to TSET. The estimates maximize the exact
Gaussian likelihood (a Kalman filter that also takes missing values inside the series)
by Levenberg-Marquardt steps on numerical derivatives. It stops after MXITER steps, when
no estimate changes by PAREPS (default TSET CNVERGE), when the adjusted sum of squares
changes by less than SSQPCT percent, or when the Marquardt constant passes MXLAMB.
MXITER=0 takes the initial values as final. AR, MA, SAR, SMA, REG and CON give initial
values in the order of the lags. Output: the iteration history (TSET PRINT=BRIEF leaves
it out, DETAILED adds the estimates of each step), the estimates with standard errors,
t and approximate significance, the residual variance and standard error, the
log-likelihood, AIC and SBC, and the covariance and correlation of the estimates. TSET
NEWVAR saves FIT_n, ERR_n, LCL_n, UCL_n and SEP_n: one-step fitted values (EXACT: from
the Kalman filter; CLS: past errors 0 and missing past values at the mean of the
differenced series; AUTOINIT: the first values of the series as their own fits), the
residuals on the log scale, the prediction limits and standard errors. Fits, limits
and standard errors are taken back from logs. With PREDICT the fit continues as
forecasts; with regressors only where they are valid, and the file is not extended.
APPLY without a name repeats the previous ARIMA with the subcommands given now replacing
its own; INITIAL starts from its estimates, FIT uses them without estimation.

Example

ARIMA daily_calls WITH promo /MODEL=(1,0,1)(0,1,1) 7 NOCONSTANT /FORECAST=EXACT.

statscraft help ARIMA prints the same text in the terminal. All commands