Different flavours of CUSUM for change point detection.
-
Updated
Nov 3, 2026 - Python
Different flavours of CUSUM for change point detection.
Monte Carlo audit of repeated testing, alpha spending, power, and stopping time
Final model for the ADIA Lab Structural Break Challenge 2026: 23rd of 729 teams, solo, 66.01 points where guessing scores 50. Code, weights, tests and write-up.
Adversarial game-integrity system that detects rigged casino games and advantage players using sequential hypothesis testing (SPRT/CUSUM) and anomaly detection.
Official implementation of On the Tightness and Computational Tractability of Higher-Dimensional Confidence Sequences
A/B testing engine with always-valid sequential inference (mSPRT) — peek continuously without inflating false positives. CUPED variance reduction, SRM detection, guardrails. FastAPI + React.
Pre-registration for forward tests: hash-pinned specs, one-look enforcement, held-out guard, anytime-valid confidence sequences, and a trial ledger.
From-scratch A/B experimentation library: power, fixed-horizon inference, CUPED, sequential testing. Validated by simulation, then run against the Upworthy and ASOS experiment archives, ending in analyst memos with ship/no-ship calls.
Simulation framework measuring how often A/B decision rules are wrong. Daily peeking: 5% -> 21.5% false positive rate.
Prompt versioning and A/B testing that refuses to declare a winner from a peeked p-value: sequential boundaries calibrated for the number of looks you actually take.
An original firewall built on real math — provably consistent rules and a tunable, proven bound on false positives.
Statistically optimal circuit breaker for stochastic systems. 4-state CUSUM-based FSM with provably minimax detection delay (Moustakides 1986). Zero dependencies.
Confidence sequences and sequential tests for stopping a model evaluation early without inflating the error rate
Reproducible benchmark for sequential change-point detection (fixed threshold, CUSUM, windowed GLR, W-formula) applied to thermal fault detection in silicon microring resonators. Validated on public Monte Carlo simulation and a real oscilloscope transient.
Real-time structural-break detection in time series: whitening by the series' own history, CUSUM/GLR/Shiryaev-Roberts on the whitened stream, LightGBM rankers and causal TCNs. ~Rank 160 of 1,716 in the ADIA Lab / CrunchDAO challenge, with the full 158-experiment research record run with an LLM assistant.
Acceptance sampling for data quality: how many records to inspect, what error count to tolerate, and what the sample says about the batch.
Code for the paper: "Profit over Proxies: A Scalable Bayesian Decision Framework for Optimizing Multi-Variant Online Experiments".
Bayesian A/B testing toolkit: Beta-Binomial posteriors, expected-loss decisions, sequential early stopping — applied to real campaign data.
Research on post-alarm re-baselining and stopping-time selection in sequential drift monitoring, with Lean formalization and certified numerical evidence.
Sharp review-count and cost bounds with exact certified-uncertainty policies for AI-assisted financial audits
To associate your repository with the sequential-analysis topic, visit your repo's landing page and select "manage topics."