UniBM

UniBM is a Python package for dependence-aware block-maxima inference in heavy-tailed time series.

It exposes two complementary inferential targets:

• severity via the extreme value index (EVI) and design-life levels
• persistence via the extremal index (EI)

The installable package lives under src/unibm. Repository-level benchmark, application, report, and static-site workflows are orchestrated through the root justfile.

UniBM is developed and maintained by Tuoyuan Cheng under the project supervision of Kan Chen. It is distributed under the MIT License.

Package surface

The public package is organized around four entrypoints:

• unibm for the core EVI, pooled EI, and design-life workflows
• unibm.evi for the severity-side workflow
• unibm.ei for the persistence-side workflow
• unibm.cdf for the public empirical CDF helper

UniBM exposes six core functions directly from unibm. The four additional root imports are not in PyPI 0.1.0: use unibm.ei for prepare_ei_bundle, bootstrap_bm_ei_path, and estimate_pooled_bm_ei, and unibm.evi for estimate_design_life_level_interval on that release. All existing subpackage import paths remain supported. When upgrading from 0.2.0, review the 0.3.0 API migration notes for removed arguments and stricter validation.

Quick start

Source checkout: 0.3.1 (unreleased). Latest published release: 0.3.0.

Install the PyPI release with Python 3.11 or later:

python -m pip install unibm==0.3.0

This installs unibm, including its estimators, interval helpers, design-life levels, and plotting helpers. Research scripts, datasets, and just full belong to the GitHub checkout and are not included in the wheel or source distribution.

For repository development:

git clone https://github.com/TY-Cheng/UniBM.git
cd UniBM
just check

The just recipes use zsh. On Windows without zsh, use the direct uv commands in Development setup.

No .env or external project is required. Defaults are data/, out/reports/, and uv's normal project environment. .env.example documents optional report, environment, and native-extension settings. Top-level just tasks load .env and sync the development environment automatically.

For ad hoc uv commands that should use .env, use just --command, for example just --command uv run pytest -q tests/test_unibm_cdf.py. A plain uv sync or uv run does not automatically load .env before choosing its project environment; without an exported override it uses .venv/.

Acceleration

The source checkout includes NumPy/SciPy optimizations and optional Cython kernels for EVI mode KDE, bootstrap quantile rank searches, and long-series EI bootstrap rolling minima. Source builds attempt to compile these kernels; without a C compiler, the same APIs use NumPy. Set UNIBM_NO_EXTENSIONS=1 before building for a pure Python distribution, or before starting Python to disable an installed extension.

EVI estimation and EVI/EI bootstrap accept n_threads=None for automatic thread selection, or a positive integer cap. Use n_threads=1 inside an outer process pool; repository workflows already assign this inner budget to their workers. NumPy/SciPy BLAS thread settings remain separate. Design-life point estimates and intervals reuse the EVI fit without another bootstrap.

The additional bootstrap and window-selection optimizations are in the unreleased 0.3.1 source checkout. See Getting Started for thread and memory behavior, and Native acceleration for build instructions.

Results and reports

Calculation results stay in the code repository:

• out/benchmark/: benchmark CSVs and sensitivity summaries
• out/benchmark/cache/: reusable simulation caches
• out/applications/: application CSVs and JSON

Final PDF figures and LaTeX tables go to Figure/ and Table/ inside a single report destination. UNIBM_REPORT_DIR unset or blank means out/reports/. An explicit value selects that destination directly, without an additional local copy. Relative paths are resolved against the code repository root:

UNIBM_REPORT_DIR="/path/to/your/report-project" just reports

An invalid destination raises an error; it never silently falls back. Report output directories and files must not redirect writes through symlinks. Table filenames describe their contents, for example application_summary.tex; direct JoH updates use these names and their matching LaTeX labels. Benchmark figures and tables use paired benchmark_evi_* / benchmark_ei_* names; raw benchmark CSVs use evi_* / ei_*. Application method comparisons use application_evi_methods.csv / application_ei_methods.csv. Combined application summaries retain the general application_summary name. report_subset_manifest.json indexes expected paths, labels, producers, and placements for the curated four-case report subset. It is not a complete inventory or a verification of generated files. manifest_code_commit and manifest_code_worktree_dirty describe the code checkout when the index was written, not the version used to calculate the listed results.

The cleanup step in just full removes only explicitly named workflow outputs for the selected benchmark sample size and report destination. Those names are reserved for generated files. It preserves caches, other sample-size runs, research notes, historical report folders, and unrecognized files. It never deletes an entire output directory. Web snapshots under docs/assets/ are refreshed by the report workflows separately.

Documentation

The site presents a statistical Benchmark followed by nine records across six case pages: streamflow, NFIP claims, Houston, Phoenix, GOES, and SPY/QQQ. Each page defines its raw or normalized observation scale. GOES reports OLS EVI point diagnostics only; other cases include EVI and EI.

To regenerate just the documented case figures and numerical records without external report export or network access:

PYTHONPATH=scripts uv run python -m application.docs_cases

GHCN, USGS, and NFIP use archived repository inputs. GOES/finance require the local snapshots described on their pages. Add --available to retain frozen optional assets when these inputs are absent; corrupt inputs still fail. The ordinary application workflow uses this optional-input policy. Site builds need only the tracked static assets, not provider access. Use just --command for these uv commands when relying on environment overrides from .env.

Package documentation is available at:

• https://ty-cheng.github.io/UniBM/

Useful local docs command:

just --command uv run mkdocs serve

This builds the static site under site/ and launches the local preview server.

Main repo entrypoints

The stable top-level entrypoints are:

• just check
• just check-full
• just data
• just refresh-data
• just benchmark
• just application
• just reports
• just full

just check runs tests affected by local changes in parallel, then checks all formatting and lint rules. just check-full runs the complete parallel test suite with the coverage gate. just data validates the tracked canonical inputs and prepares the four report cases without network access. just refresh-data is the networked recipe for canonical inputs; the optional GOES downloader is documented separately. The recipe refreshes fixed-cutoff provider snapshots and leaves their Git diff for review. just reports reuses valid benchmark summaries, computes missing ones, and reruns application fits to refresh reports and web snapshots. just full checks the project, cleans named outputs, and rebuilds the benchmark, application, report, and static-site outputs offline. These research workflows can be expensive; they are not required to use or install the package.

Minimal package example

import numpy as np
from unibm import estimate_design_life_level, estimate_evi_quantile

sample = np.random.default_rng(7).pareto(2.0, 4096) + 1.0
fit = estimate_evi_quantile(
    sample,
    regression="FGLS",
    quantile=0.5,
    sliding=True,
    bootstrap_reps="adaptive",
    random_state=7,
)
design_life = estimate_design_life_level(fit, years=np.array([10.0, 50.0]))

See the docs site for API details, returned objects, and worked examples.

The example uses the default adaptive policy with checkpoints 128, 256, 512, 768, and 1024. An explicit integer such as bootstrap_reps=480 instead fixes R. Inspect bootstrap_reps_used and bootstrap_precision_met: reaching the cap does not imply precision was met. Adaptive R controls numerical Monte Carlo error, not statistical CI width or coverage. FGLS uses fixed shrinkage 0.73 for EVI and 0.37 for EI.