Altair RISE

A program to recognize and reward our most engaged community members

Nominate Yourself Now!

Why HoltWinters Forecast returns null?

DylanZZZ

Simple Process

Image: https://us.v-cdn.net/6030995/uploads/editor/e6/hmo0ldr7aehq.png

Only one attribute with no missing

Image: https://us.v-cdn.net/6030995/uploads/editor/3w/l3q155mym858.png

Param of Holt-Winters

Image: https://us.v-cdn.net/6030995/uploads/editor/wj/zwe9myyallqg.png

Param of Apply Forecast

Image: https://us.v-cdn.net/6030995/uploads/editor/ya/f4a980l6livm.png

However, it gives nan prediction

Image: https://us.v-cdn.net/6030995/uploads/editor/jb/so50du3s6abo.png

What happens??

Find more posts tagged with

AI Studio

Time Series

Forecasting

Accepted answers

All comments

DylanZZZ

I have figured out. That's because there are zeros in data. So, how to train model with zero value?

YYH

Hi @DylanZZZ,

Glad to see you figured out by your self. The trick to handle zeros could be, re-scale data and add a small "error" to every example. For instance, using the formula

b0*10+0.01 ------>> b0'

to transform your original input b0, and make all zeros become 0.01 or even smaller (anything > 0 works).

You will forecast the new re-scaled value, and inverse it back using (b0' - 0.01)/10

Hope it helps.

YY

David_A

Hi @DylanZZZ,

I think the problem in your case is a combination of a very short seasonality value (are you sure about this, it would mean you have a pattern of length two in your data) and the zero values. In that case you will get segments in the model training, where very likely a division by zero or something similar happens.

Best,
David