Using machine learning or other methods for a fitting problem

Question

I have a time series of data of when 'data events' are going to happen.

e.g.

1.2, 4.6, 10.0, 17.3, 23.2, 24.3, 30.6, etc.

I am trying to make predictions as to when the next 'data event' will happen. Fitting the timings against the 'data event' number doesn't produce a good fit. I was wondering if a neural network based method or any other method exists that is suitable for such analysis. Or is this a problem with the data being too chaotic and just not being predictable? I am aware that most methods get much worse/unuseable when extrapolating outside of the dataset but am hoping there is a method.

Answer 1

Are your data events totally random, or do they follow a certain logic ? If they do, you can totally do things. You have to transform each time series into variables to represent it. This way, when you encounter a new serie, you compute the same variables which allows you to predict what you want.

With your example, you can make many things : Let's say you have 1.2, 4.6, 10.0, 17.3, 23.2, 24.3 as input time series, and want to predict the next event. You can say next event is in 30.6 - 24.3 = 6.3 units of time. So you have a target variable of 6.3 for this series.

With values 1.2, 4.6, 10.0, 17.3, 23.2, 24.3, you can calculate many things :

• General variables : mean time between 2 events in general, frequence, etc

• Local variables : It would be easier if you had more values in one sample, but you can also do local variables : time between the 2 last events, mean of time between the 3 last events, number of events in the last 20 units of time, etc

With such variables, you have a good view of what's happening globaly and locally. This creates variables. You're now, from a time-series problem, brought to a simple regression problem, where you have to guess, from a few variables, a target value.