Special points in a timeseries

[richie]

Hi!

Is there a way now to use MEPX to predict special points in a timeseries, for example maxima and minima, anomalies, etc...? If yes, what is the best practice? If not, is there a workaround?

Thanks,

Richie

[Mihai Oltean]

Hi,

Please explain the problem better.

For time series predictions MEP predicts the entire series, not only the minima or maxima points.

best regards,

--
You received this message because you are subscribed to the Google Groups "Multi Expression Programming" group.
To unsubscribe from this group and stop receiving emails from it, send an email to mepx+uns...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/mepx/3ca619f4-28de-4428-9da8-9f4fe7ab7f5dn%40googlegroups.com.

[richie]

OK, let's say I have a chaotic timeseries, for example a Lorenz attractor series which can be generated here https://sprott.physics.wisc.edu/cdg/.
I'm interested in forecasting local maxima or minima, and not how good it predicts other points of the timeseries.

I thought to use symbolic regression or multi-class classification, and to use training data constituted by, for example, derivatives (1st, 2nd,...) of the timeseries as input, and, as target, -1 for minima, 1 for maxima, and 0 otherwise. However, upon training, since -1 and 1 values are sparse compared to 0 values, the result is poor, probably because each outcome is weighted equally. So even if it outputs all zeros, the error is low as there are few -1 and 1.

Best regards,

R.