My experiments for the ADIA Lab Structural Break Challenge on CrunchDAO (May – September 2025).
Each sample is a univariate time series (roughly 1,000–5,000 points) with a known boundary point that splits it into two periods. The task is to predict the probability that a structural break occurred at that boundary, i.e. that the process generating the data changed. Submissions are scored with ROC AUC.
It's a code competition: a submission is a Python file or notebook exposing train() and infer() functions that run on the CrunchDAO platform.
This repo is a record of exploration rather than a polished solution. The approaches fall into three groups:
Statistical tests and hand-crafted features + gradient boosting
- Classical break tests computed on the two periods: t-test, Kolmogorov–Smirnov, Levene, Chow, CUSUM / CUSUM of squares, MOSUM, and others
- Time-series feature libraries:
tsfresh,pycatch22,nolds, wavelets, change-point detection withruptures - Models: LightGBM, XGBoost, CatBoost, and ensembles of them; feature-subspace ensembles; AutoGluon
Deep learning
- Siamese network comparing the pre- and post-boundary segments
- CNN + BiLSTM classifier
- Attention / Transformer-based models, plus attention-derived features fed into boosting models
neuralforecastmodels
Probabilistic
- A Bayesian model of the change between periods
notebooks/
LightGBM_*.ipynb, XGBoost_*.ipynb Boosting models built on the CrunchDAO quickstarter
statistical_tests.ipynb Statistical break-test features
features.ipynb, feature_model.ipynb Feature engineering + boosting
feature_subspaces_model.ipynb Ensembles trained on feature subspaces
GROUNDUP_Model.ipynb Combined pipeline: hand-crafted + attention features
AutoML.ipynb AutoGluon / CatBoost
Siamese_Network_*.ipynb, siamese_network.ipynb
CNN_BiLSTM.ipynb
attention_model.ipynb, attention_models.ipynb
NeuralForecast.ipynb
bayesian_model.ipynb
exploration/ Early baselines and package exploration
src/ Script versions of several models (feature model, Bayesian,
Siamese, anomaly detection) and a submission entry point
- Install dependencies (not every notebook needs all of them):
pip install -r requirements.txt
- The competition data isn't included. Download it with the CrunchDAO CLI, using the token from your CrunchDAO account:
Wherever a notebook contains
crunch setup-notebook structural-break <YOUR_CRUNCH_TOKEN>
<YOUR_CRUNCH_TOKEN>, replace it with your own token. Most notebooks were written in Google Colab, so some paths point to Google Drive and need adjusting for a local run.