search

WillianFuks / tfcausalimpact

Python Causal Impact Implementation Based on Google's R Package. Built using TensorFlow Probability.
Apache License 2.0
620 stars 73 forks source link

Can I use only one observed variable and date, then nothing else? #36

Open keyuchen21 opened 3 years ago

keyuchen21 commented 3 years ago

Can I use only one observed variable and date, then nothing else?

WillianFuks commented 3 years ago

Hi @km5ar ,

As far as for the theoretical concept behind the package it's certainly possible, there's nothing wrong.

Still, for most data we work with, this approach by itself doesn't generally lead to good results. One way you can test it is to divide your training data into two sections where first will be your new training data and the second the test. You can then run this simple model and observe what results you get for the impact in your new test data.

If it's good then you have more chances of getting good results on your original data; if it's not good enough (you could find for instance that the model asserts an impact being detected with statistical validity on data where you know nothing has changed) then in this scenario you could either add more components to the model (such as autoregression feature) or you'll be better off by finding covariates that help on explaining observed data through the linear regression, as offered by this package by default.

Let me know if this answers your question.

Best,

Will