Why Remove Trend and Seasonality in Time Series Forecasting?

Question

I am struggling to understand why we need to remove trend and seasonality components from non-stationary time series data when performing time series forecasting in Python. Won't removing these components affect the accuracy of the forecasted data, since we actually want to retain trend and seasonality in the final output? Although I realize that we must first make the data stationary to apply models like ARIMA and SARIMAX, I am unsure whether these models will still be able to predict future trend and seasonality accurately. I have researched this extensively on Google and on YouTube but have not found a satisfactory answer yet.

Answer 1

The trend is the long-term change we could observe in time series, which indicates a time dependency, and every time series has a trend and a seasonal trend is non-stationary.
Therefore, It is recommended to make it stationary (time-independent) to remove any hidden trends that will act as noise and are not easily observed by the naked eye.

However, I quote this source which says:

Some cases can be confusing — a time series with cyclic behaviour (but with no trend or seasonality) is stationary. This is because the cycles are not of a fixed length, so before we observe the series we cannot be sure where the peaks and troughs of the cycles will be.

the trend might inflate the correlation between time series as well (statisticians would say 'it will give spurious results'), we might look at the graph and do a correlation test and we might find a strong positive correlation. However, once we de-trend, another story shows up.
Here is another link, which shows how a correlation test between two-time series might fail if we didn't de-trend.

ARIMA could handle some cases of non-stationary time series, however, it is recommended to make them stationary in case you are looking for straightforward ARIMA implementation.

To make time series stationary check how to do de-trending and de-seasonalizing (whether by differencing or regression model could help).

I hope this will answer your question, feel free to drop comments to chat about it further.