First off, it's important to understand that Kanban relies on a backlog built of stories that are all of similar size and complexity. Yes, that means you need to estimate. But those estimates come in the form of "is this roughly as complex as other work the team has done?" So it's a lot less onerous to assess.

Then you can measure throughput (number of stories per unit of time) and cycle time (amount of time a story takes from start to finish).

Using those measures you can forecast, with monte carlo simulation as a common method. Here's a book on the topic:

https://www.amazon.ca/Forecasting-Simulating-Software-Develo...

There's plenty of other resources about this if you bust out your google-fu.