DOC: 7.2 release notes

README.md

@@ -186,9 +186,9 @@ is built by the same workflow on every push to main.
 
 ## Requirements
 
-**linearmodels 7.1 is the last release that supports Python 3.10 and NumPy 1.x
-(NumPy < 2).** Later releases will raise the minimum supported versions of
-Python, NumPy and the other dependencies.
+**linearmodels 7.x supports Python 3.10 and NumPy 1.x (NumPy < 2).**
+linearmodels 8.0 will raise the minimum supported versions of Python, NumPy and
+the other dependencies.
 
 ### Running
 

doc/source/changes/7.0.rst

@@ -1,3 +1,65 @@
+Version 7.2
+-----------
+
+.. TODO(REL): Add the release date.
+
+Version 7.2 adds wheels for Windows on ARM64 (CPython 3.13 to 3.15). PyPI accepts the
+files of a release only for 14 days after its first upload, so these wheels could not
+be added to 7.1. 7.2 also contains the fixes below.
+
+.. TODO(R2): Add #863 if it is merged for 7.2.
+
+.. rubric:: Changes that alter results
+
+- A missing value of a categorical or string column is a missing value of each of its
+  dummy variables in ``IVData`` and ``PanelData``, so that the observation is dropped
+  with a ``MissingValueWarning``. 7.1 converted it to a row of zeros, which kept the
+  observation, in the first category when ``drop_first`` is True (:issue:`858`,
+  :issue:`867`).
+- The model F statistic of the IV estimators leaves out an intercept that dummy
+  variables imply, such as dummies that sum to one in a model without a constant. 7.1
+  also tested the implied intercept, with one more degree of freedom (:issue:`780`,
+  :issue:`859`, :issue:`866`).
+
+.. rubric:: Backward-incompatible changes
+
+- ``predict`` of the panel models raises a ``ValueError`` for an out-of-sample
+  prediction (``exog`` or ``data`` is given) with ``effects=True`` or
+  ``idiosyncratic=True``, which are only available in sample, or with ``fitted=False``
+  and no other output. 7.1 ignored these options and returned the fitted values
+  (:issue:`768`, :issue:`849`).
+
+.. rubric:: Bug fixes
+
+- The IV estimators do not store weighted copies of the data when all of the weights
+  are 1, which is the case without weights or with constant weights. This reduces the
+  memory that a model uses by about a third (:issue:`607`, :issue:`862`,
+  :issue:`865`).
+
+.. rubric:: Documentation and maintenance
+
+- The tests that failed on FreeBSD no longer depend on rounding or on ``pytest -v``,
+  and the README explains how to run the tests without the other pytest plugins of an
+  environment (:issue:`853`, :issue:`854`).
+- The CI builds and tests a wheel on Windows on ARM64 (:issue:`868`).
+
+.. rubric:: Compatibility
+
+- Wheels are built for Windows on ARM64 for CPython 3.13 to 3.15, the versions for which
+  NumPy and statsmodels have Windows ARM64 wheels. The other wheels are built as in 7.1.
+- Python 3.10 and NumPy 1.x are still supported, although the 7.1 notes said that 7.1
+  would be the last release to support them.
+
+.. rubric:: Deprecations and future changes
+
+- linearmodels 8.0 will raise the minimum versions of Python and NumPy, and of the other
+  dependencies.
+
+.. rubric:: Contributors
+
+- Thanks to inchang-ing (:issue:`849`), jackiectl2 (:issue:`858`), CodingSelim
+  (:issue:`859`) and DresdenGman (:issue:`862`).
+
 Version 7.1
 -----------