Weird Development Ages on Sample Dataset when using pandas 2.0.2

[FlawlessNa]

Hello,

First of all thanks for this great package! It is extremely useful!

I recently started using the chainladder package and first noticed an issue when combined with pandas 2.0.2.
I'm not entirely sure whether the issue arises from pandas itself, or from the triangle creation within chainladder. This is why I decided to ask in the discussion section first.

Note: I've ran the script under both Python 3.8 and 3.11 and got the exact same results.

Here is the script I am running to outline the issue (I've tried with a random triangle sampled from "clrd" and observed the same problem as well):

If I run this script with pandas 1.5.3 (similar version used in the chainladder documentation), then everything goes smoothly, there are no problems with the output triangle, as below:

Now if I run this script with pandas 2.0.2, notice what happens to the Development Ages (and similarly, the data.valuation dates end up being problematic as well)

This is effectively a problem because it messes up the cl.Development() model fit when combined with various parameters (n_periods, drop_low/high, etc).

I've tried digging to find the root of the discrepancy, and it looks like it first arises in TriangleBase._development_lag which is called when using cl.load_sample.
I'm not sure of the best solution but I thought somebody might be able to help!

Thanks!!

[jbogaardt]

Thanks, I haven't yet run our test suite against pandas>=2.0. Someone also just posted issue #438 referencing the same. Sounds like there are breaking changes that need to be addressed. This issue gives me a good place to start looking.