Cross Validated is a question and answer site for people interested in statistics, machine learning, data analysis, data mining, and data visualization. It only takes a minute to sign up.
Sign up to join this community
I seem not to find this in any textbooks. So I post these questions.
Weekly forecasting is quite inadequate due to the deterministic effect of holidays and other events. Weekly data can be severely skewed by when the holiday occurs and activity before and after the holiday. Daily data analysis can provide not only good daily forecasts but good weekly forecasts. In my opinion most software errs by exclusively using autoprojective schemes when combining deterministic and autoprojective schemes should be used. Care must be taken not to assume the form of the deterministic structure and to allow the data to speak to the identification. The problem with using Fourier procedures is that it enforces a presumed structure which leads to forecasts looking like the fit but not necessarily like the data. This is easily observable by performing diagnostic checks of model residuals from these kinds of assumed procedures. If the error aren't random, the model/parameters should be questioned. If the software you are using doesn't verify the assumption of random errors to ensure that this requirement is in place by both ACF (stochastic effects) and Intervention Detection (deterministic effects) then you might be wary. We have used Fourier approaches and have been stunned by the non-independence of the error terms (residuals)
On the other hand we have found that incorporating day-of-the-week effects , changes in day-of-the-week effects, month-of-the-year effects, long-weekend effects, lead and lag structures around holidays, specific- week-of-the-month effects, specific day-of-the-month effects, changes in either levels or trends over time CAN be very useful in predicting daily, weekly and monthly aggregates.
Seldom are there any repetitious effects for a particular week in the year although we have seen some! If you are insistent on building a weekly model identify which weeks of the year, if any, have a repetitive pattern. When encountering a year with 53 weeks some clients have eliminated one week where the response is low.
Better for what? If you want weekly forecasts, use weekly data. If you want monthly forecasts, use monthly data.
Of course. Weekly data is often seasonal. If the software isn't modelling seasonality that is obviously there, then you are either using the wrong model or using the software incorrectly.
No. There isn't an integer number of weeks per month, so there is no way to reliably split the weeks that overlap month boundaries. You could apportion the week according to the number of days which fall within each month, but that doesn't take account of day-of-week effects, and with weekly data there is no way of estimating day-of-week effects.
Badly. See my blog post on this (http://robjhyndman.com/hyndsight/longseasonality/). There are several problems:
A better approach is to handle the seasonality using Fourier terms (as explained in my blog post). That can be done within an exponential smoothing framework (the TBATS model) or within an ARIMA framework (a regression with ARMA errors). I've compared these models in another blog post: http://robjhyndman.com/hyndsight/forecasting-weekly-data/
