I am working on traffic prediction per city for a medium term 3-6months ahead based on time-series daily input 6-month data. The traffic data has trends as well as seasons. Wondering if anyone could suggest/recommend models for the prediction. I am starting with Holt-Winters but seems failing at long term. Also, thinking of LSTM but would it suit for data with trends?
Your hints/comments would be much appreciated.
If you are trying to forecast a horizon that is as long as your history (or half as long) you are going to find that the accuracy isn’t good. The best “forecast” may be a flat line with the average or whatever the value was the previous year (if you know it). Within the ets family (Holt-winters being a form) you could use a trend-only version.
You should aim to have a history with several seasonal cycles. If your time series varies by week of the year (as I imagine traffic would) then you want to have several years. You won’t be able to forecast with weekly seasonality unless you have at least one year because you need at least one observation for every period within the season.
