3

We have predicted and actual (daily) data for past 3 years. We use 90 days of data for prediction. Generally our predictions are very accurate, but we receive unusual traffic for few days/weeks ( like thanksgiving-for few days, Christmas - around 2 weeks) and our predictions go off. How should i go about detecting and correcting this pattern using time series data that we have?

Thanks and Regards!

user2451089
  • 31
  • 2
  • It's not really clear how your model works, but it looks like you should control for additional features (f.ex month, week, etc). Maybe [this](http://www.ssc.wisc.edu/~bhansen/390/390Lecture14.pdf) could be a starting point – IcannotFixThis Jul 01 '15 at 11:38
  • If you know the times around which your predictions fail to be accurate, could you manually intervene and correct the predictions by adjusting your model during those times? – Vivek Subramanian Jul 01 '15 at 11:54

1 Answers1

3

With just 3 years of data it can be difficult to detect annual seasonality with pure statistical means, unless it's very clear, but in this case you don't need statistics anyways.

In your case it's best to use dummy variables, which you place in certain days. You'll use the domain knowledge to figure out which days to use dummies for.

Aksakal
  • 55,939
  • 5
  • 90
  • 176