Time series analysis and data transformation

[fog37]

TL;DR Summary

time series analysis and transformations

Hello,
Many time-series forecasting models (AR, ARMA, ARIMA, SARIMA, etc.) require the time series data to be stationarity.

But often, due to seasonality, trend, etc. we start with an observed time-series that is not stationary. So we apply transformations to the data so it becomes stationary. Essentially, we get a new, stationary time series which we use to create the model (AR, ARMA, etc.). But the transformed data is very different from the original data...Isn't the model supposed to work with data like the original data, i.e. isn't the goal to build a model that describes and can make forecasting on data that looks like the original data, not like the transformed data?

Thanks!

 

[Dale]

TL;DR Summary: time series analysis and transformations

Isn't the model supposed to work with data like the original data, i.e. isn't the goal to build a model that describes and can make forecasting on data that looks like the original data, not like the transformed data?

As long as there is an inverse transform then you can get back to the original scale.

The usual problem with computing the statistics on the transformed data is that the residuals usually have different properties. Assumptions on the residual distribution hold on the transformed scale, and when inverse transformed they may be quite different.

 

[FactChecker]

There are many levels and definitions of "stationary". See Stationary process. A lot of people would not consider an ARIMA or SARIMA to be stationary in the most simple sense.

 

[fog37]

As long as there is an inverse transform then you can get back to the original scale.

The usual problem with computing the statistics on the transformed data is that the residuals usually have different properties. Assumptions on the residual distribution hold on the transformed scale, and when inverse transformed they may be quite different.

Ok, I guess the key word is "inverse transformation". We convert the original signal into a new signal, create a model for the new signal, make predictions, and finally apply an inverse transformation to predictions which would now make sense for the original data...
It is the same things as when we convert a time-domain signal ##f(t)## into its frequency version ##F(\omega)##, solve the problem in the frequency domain, get a frequency domain solution, and convert that solution back to the time domain...

 

[Dale]

Yes, that is a good example

 

[fog37]

One realization I just had is that time-series models like ##AR, MA, ARMA, etc.## seem to just be discrete time ODEs, i.e. difference equations...But these linear models are generally used to make predictions/extrapolations of unknown values of ##y_t## without reaching a final solution, ##y=f(t)##, correct? Why not?

For example, a fitted AR(1) model is something like this: $$y_t = a y_{t-1}$$ which can be converted to the ODE model $$y_t = \frac {a} {a-1} y'$$

Why not solve for ##y_t## instead of keeping it as ##y_t = a y_{t-1}##?

 

[FactChecker]

Why not solve for ##y_t## instead of keeping it as ##y_t = a y_{t-1}##?

Because the direct solution for ##y_t## includes the cumulative random terms of all the preceding time steps. That can have a huge random variance. On the other hand, if you know the value of ##y_{t-1}##, why not use it and the random variance from that is just from one time step and is relatively small.

 

[fog37]

I was thinking the following in regards to transformations, inverse transformations, ARMA, ARIMA and SARIMA.

ARMA is meant to model time-series that are weakly stationary (constant mean, variance, autocorrelation). To train an ARMA model, the training signal ##y(t)## must therefore be stationary. If it is not, we need to apply transformations to make it so and apply inverse transformations at the very end.

With ARIMA, we avoid manually doing the stationarizing step since the ##I(d)## part of ARIMA automatically make our input signal with trend and seasonality stationary, if it is not, by taking the difference transform...
But I guess differencing does not take remove the seasonal component from ##y(t)##? Does that means that we would need to remove seasonality manually before using ARIMA?

The best solution seems to then use SARIMA which does not care if the training signal has trend and/or seasonality because it takes care of it internally: we don't need to manually apply any transformations to the raw time series ##y(t)## and inverse transformations to the prediction outputs of the SARIMA model....

Any mistake in my understanding? I would definitely choose SARIMA, more convenient, since we can skip all those preprocessing transformations to make ##y(t)## stationary and inverse transformations after the forecasting...

 

[FactChecker]

Any mistake in my understanding? I would definitely choose SARIMA, more convenient, since we can skip all those preprocessing transformations to make ##y(t)## stationary and inverse transformations after the forecasting...

That is a natural thought. But you should avoid anything that would be like "throwing everything at the wall to see what sticks". A time series analysis tool-set might allow you to automate finding a SARIMA solution that only includes terms that are statistically significant. But you should have some subject-matter reason to include the trend and seasonal terms. A good tool-set should allow you to prevent the inclusion of terms that do not make sense .