ARIMA - odd behavior, time period type?

[cwkoops]

See attached ARIMA workflow that has both the "batched" model factory tools and the regular ARIMA and TS forecast tools.  They both behave the same, which is good...but some of the behavior is strange.  

The data contains 2 sets to make a prediction on the DLT column: 80002_80073_SEA and 80002_80093_SEA.  The group of data 80002_80093_SEA has odd behavior.  The last data point is 63, and the TS prediction for the next data point ends up being 63....even though most of the previous data points are significantly larger than 63.  If I manually make the last data point 10, the prediction ends up being 10.  If I make the last data point something slightly larger, like 74, the prediction ends up being 89 (or at least something that makes sense).  So it appears that if the last data point is small enough, that last data point IS the prediction.  

Now, the data points don't occur once per day, once per month, etc.  Sometimes there are multiple data points on the same day or maybe not another one for another few weeks.  Essentially I just want it to predict the "next" one.  As a result, I have the "Time Period Type" or the "Target Field Frequency" as "Other".  I can't find documentation on if this makes the most sense or not, though I will say that when I change the time period type to "daily" the strange prediction on the "80093" group makes more sense, but the output for the "80073" group doesn't change...which I suppose is good because it made sense to begin with.  

Can someone help me understand:

1) What is driving the strange behavior on the '80093' group based on the last data point, and

2) from a configuration standpoint, what should I use for a 'target field frequency'/ 'time period type'?  To me, "other" makes the most sense (as it varies), though that seems to be contributing to the behavior in #1.  

[cwkoops]

Did some digging into this:  In instances like this one, p,d,q ends up being 0, 1, 0 (when the configuration was set to "other", in this case), so the prediction is essentially the previous data point.