inf treated as a number, which is very weird

As noted here: https://stackoverflow.com/questions/48371747/how-to-modify-a-liner-regression-in-python-3-6

This formula causes patsy to raise an error:

patsy.ModelDesc.from_formula("inf ~ x")

the problem is that in patsy.parse_formula._read_python_expr, patsy tries to figure out whether an arbitrary Python expression is a numeric literal, and the way it does this is by calling int(...) and float(...) on the expression, and seeing if they work.

In this case, float("inf") does work, so patsy decides that the Python expression inf is a numeric literal. Whoops.

The same thing probably happens if you try to use nan as a variable name in a formula.

I guess a more reliable way of checking for numeric literals would be to check the tokenize output: if an expression is a single token, and that token has type tokenize.NUMBER, then it's a numeric literal.

Start in patsy.parse_formula._read_python_expr and reproduce the failure with patsy.ModelDesc.from_formula("inf ~ x"). Inspect the tokenize output for the expression and compare it with the current int() and float() checks. Done means inf and nan can be used as formula variable names without being classified as numeric literals.