Prophet: Suggestions on tightening lower and upper forecast range

Created on 18 Jan 2019  路  2Comments  路  Source: facebook/prophet

Hello All,

I am interested in ways to tighten the forecast range:

image

As you can see the accuracy level is pretty good. Although the sway in the forecast is large. I am using change points value of 250 and scale of 0.05. Will picking appropriate transformation in Tukey Ladder of Powers be the solution? Or there are other ways / config params to control the sway?

Thanks,
Dhanashree

Most helpful comment

By tighten the range, do you mean tighten the interval between yhat_lower and yhat_upper?

If so, then there are two things you can do. The first is to use the parameter "interval_width" to reduce the width of the uncertainty interval, as described here: https://facebook.github.io/prophet/docs/uncertainty_intervals.html
By default it is 80%, which means if there were no model misspecification you'd expect the actual values to lie within the region 80% of the time. You can reduce that, and it will give tighter intervals.

You can also reduce the number of past change points and/or the changepoint prior scale. This allows less flexibility in the past, and so the uncertainty intervals will reflect less flexibility in the future (i.e., will be narrower). (Assuming you don't significantly reduce the model fit error). 250 seems like a rather large number, and I'd expect it could be reduced without impacting predictive performance.

All 2 comments

By tighten the range, do you mean tighten the interval between yhat_lower and yhat_upper?

If so, then there are two things you can do. The first is to use the parameter "interval_width" to reduce the width of the uncertainty interval, as described here: https://facebook.github.io/prophet/docs/uncertainty_intervals.html
By default it is 80%, which means if there were no model misspecification you'd expect the actual values to lie within the region 80% of the time. You can reduce that, and it will give tighter intervals.

You can also reduce the number of past change points and/or the changepoint prior scale. This allows less flexibility in the past, and so the uncertainty intervals will reflect less flexibility in the future (i.e., will be narrower). (Assuming you don't significantly reduce the model fit error). 250 seems like a rather large number, and I'd expect it could be reduced without impacting predictive performance.

Yes, thanks for all the advice, that is exactly what I was looking for Ben. Have a nice day!

Was this page helpful?
0 / 5 - 0 ratings

Related issues

GretaShi3084 picture GretaShi3084  路  3Comments

ahash52 picture ahash52  路  3Comments

datafool picture datafool  路  3Comments

kimhale picture kimhale  路  3Comments

sriramkraju picture sriramkraju  路  3Comments