OptiBin

Efficient sample allocation and error estimation for binary classification evaluation.

This project demonstrates:

• Optimal allocation with stratum caps + iterative reallocation
• Expansion estimator of error proportions with finite population correction
• Wald confidence intervals for error rates
• Automated visualisation via Plotly + Puppeteer (static PNGs)

---

Example: Confidence Intervals

The plot below shows 95% Wald confidence intervals for three strata with different error prevalences (and therefore variances):

---

Getting started

Clone and install:

git clone https://github.com/grthomson/optibin.git
cd optibin
npm install

Run the test suite:

npm test

Rebuild + export demo figures:

npm run build
node dist/examples/export_cis.js

Outputs will be written to the images/ folder.

Project Structure

optibin/
  src/         # core allocation/estimation code
  tests/       # vitest unit tests
  examples/    # demo scripts that generate plots
  images/      # exported static figures for README/docs

Notes

OptiBin focuses on binary-outcome (proportion) evaluation with a clean-room TypeScript implementation. This was largely a self-learning project; the following alternative tools are likely to provide a better sollution that OptiBin.

• optimall (R) Established R package for optimal allocation workflows (including Neyman-style allocation). If you work in R, it's a good cross-check for allocations and design decisions.

• ssepy (Python) Useful for sample size and proportion calculations.

These tools serve similar planning/estimation aims; OptiBin's value is offering a minimal, typed TS implementation that's easy to plug into web dashboards or CI pipelines, plus headless PNG exports for docs.

CI

The GitHub Actions workflow runs on every push and pull request.
It installs dependencies, type-checks the TypeScript, runs the Vitest unit tests, and builds the project on both Linux and Windows using Node.js 20.