numerai / numerai-tools Goto Github PK

Hi, With the current package settings numerai-tools does not work with scikit-learn 1.4 onwards. Can we relax the scikit-learn version so it works. I don't think that newer 1.x scikit-learn versions would contain breaking changes to the numerai-tools code.

I have a PR ready that solves this #22

Pandas is currently at 2.2.1 and the requirement makes numerai-tools conflict with for example parquet-tools, which is fairly useful.
Requiring 'old' package versions is always bothersome, so there should be a fairly good reason for such a requirement imho.

I got this error on my datasets:

157 assert np.array_equal(df.index.sort_values(), df.index), "Index is not sorted"
158 result = np.sign(df) * np.abs(df) ** p
--> 159 assert ((result.std() == 0) | (result.corrwith(df) >= 0.9)).all()
160 return result

AssertionError:

After some research I find the correlation is not match the assert

What should I do in this case ?

numerai-tools has a quite old version of scipy (~1.2.1) set as a dependency. Most other libraries nowadays depend on more recent versions of scipy. This makes it hard to use numerai-tools with a dependency manager like Poetry.

Could the dependencies be updated to more recent versions in setup.py? Especially the old scipy version is a bottleneck.

Example error from Poetry:

Because no versions of numerai-tools match >0.0.12,<0.0.13 and numerai-tools (0.0.12) depends on scipy (>=1.2.1,<1.3.0), numerai-tools (>=0.0.12,<0.0.13) requires scipy (>=1.2.1,<1.3.0).
So, because project depends on both scipy (^1.10.0) and numerai-tools (^0.0.12), version solving failed.