•
NeuralForecast v3.2.0 (2026-07-10): categorical exogenous for univariate non-recurrent models via embeddings (cat_exog_list, categorical_cardinalities, cat_emb_dim). Educational after-work note. Not investment advice.
NeuralForecast v3.2.0 (published 2026-07-10) adds categorical exogenous support for univariate non-recurrent models via learned embeddings — phase 1 of a broader categorical-exog roadmap. Key knobs: cat_exog_list, categorical_cardinalities, cat_emb_dim. See the v3.2.0 release and PR #1552. After-work note for Python / data / ML / forecasting notebooks — not investment advice.
The only course CTA is Introduction to Python Programming.
Many after-work panels carry region / product / day-of-week style categoricals. Before 3.2 you often one-hot or scale them like continuous exogs. 3.2.0 lets supported univariate non-recurrent models embed columns listed in cat_exog_list (drawn from hist/futr/stat exog lists), with categorical_cardinalities mapping each column to its distinct count and cat_emb_dim choosing a size strategy (fastai / sqrt / half or an int).
| Item | Detail |
|---|---|
| Version | neuralforecast 3.2.0 (2026-07-10) |
| Headline | Categorical exog via embeddings (univariate non-recurrent; phase 1) |
| Knobs | cat_exog_list, categorical_cardinalities, cat_emb_dim |
| PR | #1552 |
python -m venv .venv-nf32 && source .venv-nf32/bin/activate
pip install "neuralforecast==3.2.0"
python -c "import neuralforecast as nf; print(nf.__version__)"
Skim the release notes for which models gained the knobs, then try one synthetic categorical column on a toy univariate series (educational only). Stop after a fit on synthetic data — do not claim a live trading edge.
This Hub post is only NeuralForecast 3.2 categorical exog. It does not bundle Prophet / Chronos / TimesFM / sktime / Darts-MLflow Hub posts — those stay separate.
Pin neuralforecast==3.2.0, read #1552, embed one toy categorical. If Python packaging is the blocker, Introduction to Python Programming.