ARIMA forecast returns same repeated value for future.

[jgustave]

Hello,
I've noticed in some cases that ARIMA.forecast will return the same value continuously for the future part of the forecast. Seems suspicious to me. (This is using a fresh build off of master, but I've also seen it in 0.4.1, and is fairly common) (Scala 2.11.7)

Here is a scala unit test:

package junk

import com.cloudera.sparkts.models.ARIMA
import org.apache.spark.mllib.linalg.Vectors
import org.junit.Test

class ArimaTest {
  
  @Test
  def testArimaFoo() {

    val data = Seq(32.284, 32.0722, 31.9126, 33.1172, 33.5671,
                   34.0824, 34.4891, 34.4198, 33.4821, 32.836,
                   32.5691, 31.808, 32.7928, 33.142, 33.2493,
                   33.9647, 34.4421, 35.5485, 35.8389, 35.3523,
                   35.8036)
    val dv = Vectors.dense(data.toArray)
    val model = ARIMA.autoFit( dv )

    println("p="+model.p+" q="+model.q+" d="+model.d+" Invertible:"+model.isInvertible()+ " intercept="+ model.hasIntercept + " stationary:" + model.isStationary())
    println("Coeffs"+model.coefficients.mkString(","))
    println("AIC:" + model.approxAIC(dv) )

    println("###")
    println("Values:"+ data.mkString(","))
    println("Forecast:"+model.forecast(dv,5).toArray.mkString(","))
  }
}

---

And this is the stdout:

Nov 23, 2016 5:19:22 PM com.github.fommil.jni.JniLoader liberalLoad
INFO: successfully loaded /var/folders/3c/7jzx8fts1_ngxgv3nqpdxc600000gn/T/jniloader5183904765086523584netlib-native_system-osx-x86_64.jnilib
Warning: MA parameters are not invertible
p=0 q=1 d=1 Invertable:true intercept=false stationary:true
AIC:37.874997884062026
Coeffs0.31643668642971895

Values:32.284,32.0722,31.9126,33.1172,33.5671,34.0824,34.4891,34.4198,33.4821,32.836,32.5691,31.808,32.7928,33.142,33.2493,33.9647,34.4421,35.5485,35.8389,35.3523,35.8036
Forecast:32.284,32.216978709814185,32.04290469983246,32.30304974018622,33.13601224322286,33.72420694060749,34.161380400670026,34.44217864134951,34.137924958389,33.36684586105238,32.7880136862178,32.34344466808555,32.19103247430219,32.7820941639381,33.1793413757428,33.46386262426504,34.04787138824745,34.76588707136306,35.537935105768284,35.688265028506464,35.8036,35.994074307825585,35.994074307825585,35.994074307825585,35.994074307825585,35.994074307825585

[jgustave]

I suppose this is because P is 0?

[ekote]

@jgustave, if p=0 then there is no autoregression operation when you use ARIMA. Your data are just first once differentiated (d=1) and then MovingAverage part is predicting value. So yes, your predicted data will be the same unless your p >= 1.