Risk-limiting Financial Audits via Weighted Sampling without Replacement

#Confidence Sequences for Weighted Sampling without Replacement

Installation

Requires Python 3.8.3. Install dependencies with pip install -r requirements.txt

Running experiments

Figures generated from running experiments will be in the figures/ directory.

• bash scripts/run_exps.sh runs all the experiments in the main body of the paper.
• bash scripts/run_cs_comp.sh runs all the experiments in Appendix D that compares different types of CSes.

These figures can be rendered in a LaTeX --- below is an example template.

\documentclass[10pt]{article}
\usepackage{tikz}
\usepackage{pgfplots}
\pgfplotsset{compat=newest}
\pgfplotsset{scaled y ticks=false}
\usepgfplotslibrary{groupplots}
\usepgfplotslibrary{dateplot}
\tikzstyle{every node}=[font=\small]
\pgfplotsset{
    yticklabel style={/pgf/number format/fixed},
}
% For fixing legend of histograms
\pgfplotsset{compat=1.11,
 /pgfplots/ybar legend/.style={
 /pgfplots/legend image code/.code={
 \draw[##1,/tikz/.cd,yshift=-0.25em]
 (0cm,0cm) rectangle (3pt,0.8em);},
 },
}

\begin{document}
\input{<path to tex figure>}
\end{document}

Code structure

All of the code is the src/ directory.

The notebook src/example.ipynb explains the basic steps in setting up an experiment, and constructing confidence sequences.

Confidence sequence code is in hoeffding.py, bernstein.py, (for Hoeffding and empirical-Bernstein) and the betting CS code is in weightedCSsequential.py, along with the helper functions.

utils.py contains code for generating the transaction/misstatement values, as well as other useful functions.

Reproducing the figures

ExperimentBase.py and Experiment{1,2,3}.py contain the code for running experiments and simulation setups that occur. Experiment4.py contains the code for comparing the performance of three methods (propM, propM+CV, and uniform) on a 'semi-real-world' dataset of house prices.