It would be nice to specify a single df filter that accepts:
[a] list of panel criteria
[b] list of row criteria (indexes like range or every other, values like > or nan),
[c] lists of column criteria (column names, column values like nan or <).
Names for these would be explicit (e.g. ‘column_filter’, not ‘iloc’).
More attention to setting dtypes.
When passing a dictionary of astype()… it doesnt always work, and i have to force individual column dtype one at a time to get it working.
Tighter integration with parquet because it preserves dtypes well except for dates. Pyarrow and other engines are not great.
Every type of ingestion should have parameters for handling null conditions for when column names aren’t set.
numpy 3D array of arrays of arrays are wicked easy to work with. Panels should be too. Pandas is supposed to mean panel data, but imo it failed there. E.g. create panels from a list of dataframes or python lists or numpy arrays.
don’t know about you guys but i have never had to perform broadcasting-like operations.
tight integration with jupyter and plotly (upgrade from matplotlib).
support for in-memory graph data format: nodes, edges via two paired dfs.
i know there’s no innovation in this list. just talking through the basics.