1
$\begingroup$

I'm going through Google's Unnoficial Data Science blog, specifically Our quest for robust time series forecasting at scale.

Their approach to forecasting includes making weekly total forecasts, and then distributing those totals to the individual days of each week. I'm wondering what is the recommended way to do this.

Specifically, if I have a time series of daily/hourly data, aggregate it as weekly totals, and then make forecasts for the following weeks, how can I distribute those totals to daily/hourly values?

One approach I've considered is somehow applying Rob Hyndman's advice from Forecasting: Principles and Practice on Ensuring forecasts stay within limits, but it feels like I'm overcomplicating the problem.

Edit: I've found several references for using optimal forecast reconciliation to adjust the forecasts for both the aggregated data and the individual values. Still trying to determine whether it can be used to only reconcile the daily/hourly values, based on the weekly totals.

Improve this question
$\endgroup$

2 Answers 2

Reset to default
0
$\begingroup$

Some solutions :-

  1. Forecast on weekly data and then use certain hard-coded rules to attribute weekly forecasts to hourly / daily basis.
  2. Train yet another AI model using RNN / LSTM - Vector to sequence translator - which learns how to spread weekly total to hourly / daily spread by training on historical data which maps weekly total to hourly / daily spread - this essentially will LEARN the mean mapping from weekly total to hourly / daily spread.
  3. Train the model at hourly or daily time series level (do not total to weekly level) - maybe this causes scale issue as too many points at hourly or daily level then first two solutions above can be tried.
Improve this answer
$\endgroup$
1
  • $\begingroup$ Thanks, here are my comments - 1. I'm looking for the "recommended" approach to take here, I don't think hard-coding rules qualifies (I have thought however about using past hour averages as percentages out of the total amount); 2. Sounds like overkill, and it won't guarantee that the hourly forecasts add up to the totals; 3. Doesn't answer the question $\endgroup$
    – Vlad
    Commented Jan 29, 2022 at 17:30
0
$\begingroup$

One more idea is to do recurrent time series modelling on hourly / daily data itself (do not sum upto weekly totals), using a more sophisticated models like LSTM or even better state of the art Time Series models with Transformers with ATTENTION and Skip Connection architecture which can capture very very long sequences successfully allowing scalable time series forecasting.

Improve this answer
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.