Results Looks great... except for the last Example of the Scoring Set
I am quite frustrated with the results I am getting. My DataSet has a binary label with a single example unknown. The last one in a Daily Time Series, corresponding to the current date.
I run this model Daily to get one prediction for each day. So the current day is really the only prediction I care about.
However the current day prediction is always bad in the sense that is shows always a fixed class, let's call it class ZERO.
When the next day comes and I update the dataset to include one extra example, the prediction of yesterday becomes correct, and the now current one, a bogus class ZERO (it could actually be a class ZERO but I cannot trust it because it ALWAYS shows class ZERO).
Now, the weird part is that if I split the DataSet in two and train the model on just half the data, it still is able to predict the other half...except for the last, current day example, where it always shows class ZERO.
Even unlabeling half the Data does not change this as the model predicts everything correctly except for the last example.
I also tried adding a dummy example at the end but even this dummy one gets correctly predicted and the one I am interested in does not!
What could be happening here?