Questions about Non-Stationary Time Series Data

Hi,

Currently I’m working on a set of time-series data. My goal is to build an ARMA or ARIMA model; but after testing, the data proved to be non-stationary .

Non-stationary data, as a rule, are unpredictable and cannot be modeled or forecasted.

Can I adjust the time-series data to make it stationary? If so, what preprocessing method should I use?

1 Like

Hi Cheng,

I think first you can apply ADF test to see if the time series can be fitted by an autoregressive model with drift of trend. If ADF test indicates it unstationary, you can try taking the difference or logrithm and see if the time series after transformation can pass the ADF test. You may do the transformations more than once. At last, if the time series is still unstationary, it might indicate that there are some non-linear patterns in the data and you may switch to other models which can capture non-linear relationships like RNN and LSTM. Hope it helps!

1 Like

Thank you. That is very helpful!

Hi Cheng,

Like Muyi said, if you do an autocorrelation function for your log(time series) and it is high/close to one, it might be helpful to difference the function. You can then check to see if the difference was helpful or not by looking at the ACF and PACF tests. If you find a significant level of negative autocorrelation at Lag 1, there is evidence of over-differencing, and this process can be used to help select your ARIMA model.

If this doesn’t work then you might want to switch to a non-linear model.