foretools is the companion toolbox that sits next to foreblocks.
Use foreblocks when you are building and training forecasting models. Use foretools when you need support utilities around that workflow: synthetic data, black-box search, exploratory diagnostics, decomposition, or feature engineering.
| Tool | When to use it | Docs |
|---|---|---|
foretools/tsgen |
create synthetic series with known structure and ground-truth components | Time Series Generator |
foretools/bohb |
run budgeted hyperparameter optimization with Hyperband + TPE | BOHB Search |
foretools/emd_like |
decompose signals into oscillatory modes with VMD, EMD-family methods, hierarchical VMD, and multivariate support | VMD Decomposition |
foretools/fengineer |
automated feature engineering with transforms, interactions, MI selection, and RFECV | Feature Engineering |
foretools/tsaug |
data augmentation — jitter, scaling, time-warp, window-slice, and AutoDA search | AutoDA Augmentation |
foreminer is an exploratory-analysis toolkit for understanding your time series before modelling.
Key capabilities:
- Changepoint detection — locate structural breaks in long series
- Cluster analysis — group series or windows by similarity
- Dimensionality diagnostics — PCA and UMAP projections of window embeddings
- Group-level summaries — aggregate statistics and seasonal decomposition across cohorts
- Stationarity checks — ADF and KPSS tests with automated reporting
Quick import path:
from foretools.foreminer import ForeMiner
miner = ForeMiner(series) # series: [T, D] numpy array
report = miner.run() # returns a dict of diagnostic frames
miner.plot_changepoints()
miner.plot_clusters(n_clusters=4)foreminer is primarily notebook-oriented. It does not expose a stable training-time API and is best used in exploratory phases before committing to a preprocessing and model pipeline.
Data augmentation utilities. See AutoDA Augmentation for the full guide.
foreblocksis the main model and training API.foretoolsis a set of practical companion modules. Some are notebook-oriented and some are reusable library code.foretoolsimports are deeper and less consolidated thanforeblocks, so the safest entry points are the specific modules documented here.
- Time Series Generator if you need synthetic datasets or decomposition examples.
- BOHB Search if you need hyperparameter optimization outside the
foreblocks.dartsneural architecture search stack. - VMD Decomposition if you need decomposition, denoising, or mode extraction workflows.
- Feature Engineering if you need automated feature construction, mutual information selection, or RFECV-based pruning.
- Repository Map if you want the broader code layout.