I'm an amateur in statistics. I was reading a paper in which it is mentioned that

A model gives high out-of-sample likelihood.

What does that mean? What is exactly "High-out-of-sample likelihood"? What would be the simple explaination of it?


For other folks who might have an answer, the relevant quote from the paper is as follows:

The objective is to learn a good model $f(w_{t} ,\dots ,w_{t−n+1}) = \hat{P}(w_t|w_{t−1})$, in the sense that it gives high out-of-sample likelihood

After skimming the relevant section of the paper you linked, I think you could rephrase the author's comment (in plain English, avoiding stats jargon as much as possible) as follows:

"A good model is one that is consistent with the data we actually observe"

Two important concepts to understand here.

  • Training data vs. real-world (out-of-sample) data:

    • Training data is used to fit the model initially. That's different from the the 'real-world' data you really need the model to handle. The author seems to be referring to this real-world data when he/she uses the term out-of-sample data.
  • Likelihood:

    • When fitting a model, you want to know "how plausible is this model?". To fit the model, you maximize this "plausibility", which is represented by the likelihood of seeing the data, given the model.
    • This is a tricky concept - it's not the likelihood of the model being "right", but is still a measure of how well the data "matches" the model.

So the author is saying that we want to learn a "good model" in the sense that it is the most plausible model for the real world data we encounter.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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