Long-term trend prediction of time series data

Question

I have a time series dataset project (single variable time series) on market share changes of a particular product in a region (values are recorded every day from 2018 to 2022) where I need to predict whether the share values will show an upward or downward trend long-term (1 year). I will be using Python for the project.

I have been searching for various methods such as traditional ARIMA, random forest, and also LSTM but I'm not sure which would be best for forecasting long-term trends. I know that predicting exact values for long term isn't feasible due to performance issues, but my objective here is to just forecast a 'directional' long-term trend. So my task is basically trend analysis using time series methods.

Is there a good method to forecast trends among the ones mentioned above? Or is there possibly a better method I don't know? I would appreciate any insight into the issue.

Answer 1

I guess your time series have some kind of seasonality, since it is a product which is probably sold more often during some periods of the year or month?

A way to tell more about the trend is using models that decompose the signal accordingly. One popular option is Facebook's Prophet. It gives you change point detection and trend uncertainty out of the box. But it really depends on the amount and quality of data you have. I would use the output of Prophet more as a metric to get a good estimation of the trend uncertainty and approximate direction.

Another way would be to analyse the changepoints by yourself with the ruptures library. You can choose from various approaches to divide your time series into pieces and calculate individual trends. Depending on the distribution of trend slopes and piece lengths you could estimate a future trend uncertainty interval on your own. For example an interval between the mean and median of the trend slopes weighted by its length?

Maybe not the answer you were hoping for, but hopefully a direction to follow ...