I'm trying to evaluate a model for a time series, given many time series (plural). For example, i'm using the forecast package and in particular the ets function to forecast based on a time series.

My data was not continuously gathered, so I have around 50 sessions of 1-2 hours each, where each session was recorded on a different day.

How do I evaluate the parameters of a time-series model using multiple experiment sessions data? concatenating the time series is obviously not a good idea because the last samples of session k-1 have no affect on the first samples at session k.

This is a special case of an irregular time series, but I don't think it should be treated as one.

Here is an example code:

# original time series, one per recording session:
ts1 <- ts(rnorm(n = 10, mean = 1, sd = 1),start = as.POSIXct(1433679895,origin="1970-01-01"),frequency = 1)
ts2 <- ts(rnorm(n = 10, mean = 1.7, sd = 1.8),start = as.POSIXct(1433766295,origin="1970-01-01"),frequency = 1)
ts3 <- ts(rnorm(n = 10, mean = 1.5, sd = 1.3),start = as.POSIXct(1433852695,origin="1970-01-01"),frequency = 1)

# concatenate all time series to an its (irregular time series) object,     just as a way to represent the combined ts
dates <- as.POSIXct(c(time(ts1),time(ts2),time(ts3)),origin="1970-01-01")
ts.all <- its(x = c(ts1,ts2,ts3), dates)

ets.model <- ets(ts.all,model='ZNN',alpha = 0.3)

So the model assumes that this is a regular time series, even though it is not. Is there a way to iteratively evaluate the parameters of the model given multiple sessions of data?

This is actually a general question regarding time series analysis in chunks. This problem can happen with any analysis and any package.



Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.