It will depend on the nature of your data and any expected periodicity. If these are daily observations of human behavior, there might be a strong weekly component. In the unlikely event you're looking at moon phases, then 28 days will be more appropriate. If it's a natural phenomenon, there might be an annual periodicity or none at all.
If you're not sure, the same Hyndman text discusses how you could look at autocorrelation to understand any apparent periodicity: 2.8 Autocorrelation | Forecasting: Principles and Practice (3rd ed)