Adding hour data to Date feature in time series prediction

Question

I'm a noobie with regards to working with time series data. I have a time series data which has date column in YYYY-MM-DD and a hour column containing 0 to 23 representing hour of the day and the value containing a value between 70 to 100 which is what we need to predict.

Models I tried: I checked for Stationarity of the data and I split the dataset based on time(oldest 80% train and newest 20% as a test) and tried two models: AR and ARIMA. AIC score for ARIMA was around 2500.0

My questions are: Earlier I discarded the hour feature and built models with only Date feature. So Will Adding hour data improve my model? If I'm adding hours, Should I keep it as a separate column or should I include the hour in the YYYY-MM-DD date feature?

Thanks in advance

Answer 1

I am a bit confused by the setup of this question but I'll try my best.

First, if your time series has YYYY-MM-DD, and then you also have an hour column with ranges 0:23, then what you have is a hourly time series. That is, the natural index of the time series is by the hour, not by any other frequency of time, and therefore, you would have no need to include either of these two columns in your "dataset".

If the hour of the day is a strong "predictor", then this is just another way of saying that your time series exhibits hourly seasonality and therefore, including hour of the day may help if you see that the residuals of the time series are not consistent with white noise (an assumption of ARIMA models) and perhaps have a superimposed pattern on the ACF plot. A seasonal ARIMA model might be of interest to you, or perhaps an ARIMAX model with an exogenous hour of the day variable. Furthermore, certain particular days might have noticeable and consistent effects, or even months. There may also be one time effects, etc. which might be useful to include as exogenous variables.