Time Series and Different Time Scale Lag ContributionsHello JMP-Com,
I got this question from a customer who wanted to use either ARIMA models or time series forecasting (TSF) for their application.
Column 1 is a time stamp
Column 2- Column N has the data points
The Question: Can the ARIMA platform or TSF account for contributions from lag points that are in a different time scale?
For example; say I have monthly sales data and I know the sales in April are impactful to the sales in June. So I know that a lag of 2 months is important for my model. But is there a weekly/daily/hourly lag that can be impactful to today as well? So I know a 2 month lag is important, but can I see if 3 weeks ago lag is also important?
Months and weeks are on different time scales, and the customer has data given in minutes. But they want to know which time scale would be the most important for their time series model: yearly, quarterly, monthly, weekly. AND if its possible for a time series model to use lag points from different time scales.
Thank you,
lovy
Re: Time Series and Different Time Scale Lag ContributionsIn ARIMA and Exponential Smoothing (TSF), there is only one time scale.
I can think of a couple of ways to accommodate multiple scales:
- Work with the finer time scale. If you have monthly and weekly information, work with the weekly information. Monthly information is aggregated from weekly, or weekly is distributed out using the monthly. How to aggregate, how to distribute, those are the next questions. Also what the "lags" will be using the finer time scale, e.g. which "weekly" time point in the series represents the "monthly" information you need to refer to.
- Work with Transfer Function model, which can incorporate information other than the series itself. E.g. fit a "monthly" series, while add "weekly" information as an additional "input". (But which week to use? So the first suggestion of working with finer scale is still relevant.)
Re: Time Series and Different Time Scale Lag Contributionsjust throwing in a crazy idea to find the most appropriate time lag and not sure this will work out.
You may want to create multiple columns from your time column to build different time-lagged time columns. Then use Structural equation models to find the most representative lag for your response data.
It might be shooting with elephants on sparrows but might work 
