The dataset for this competition comprises tens of thousands of synthetic univariate time series, each containing approximately 1,000 to 5,000 values with a designated boundary point. For each training time series, a label (True for break, False for no break) indicates whether a structural break occurred at this boundary point.

The time series in this competition are designed to represent various real-world scenarios where structural breaks may occur, with different levels of difficulty in detection. This includes scenarios similar to those found in financial markets, climate data, industrial sensor readings, and biomedical signals, among others. The challenge is to develop algorithms that can generalize across these scenarios and accurately detect structural breaks in new, unseen data.

Data Format

For the training data, the X variable will be provided as a pandas.DataFrame with a MultiIndex structure. Here's an example of the format:


                    value  period

  id     time

  0      0       0.001858      0

         1      -0.001664      0

         2      -0.004386      0

         3       0.000699      0

         4      -0.002433      0

  ...              ...        ...

  10000  1890   -0.005903      1

         1891    0.007295      1

         1892    0.003527      1

         1893    0.007218      1

         1894    0.000034      1

The DataFrame has the following structure:

A MultiIndex with two levels:
- id: Identifies the unique time series (each time series has a unique ID)
- time: The timestep within each time series
The columns include:
- value: The actual time series value at that timestep;
- period: A binary indicator where 0 represents the period before the boundary point, and 1 represents the period after the boundary point. The structural break occurs at the change in value, but it may take some time (values) to become detectable/apparent.

The y variable is a boolean pandas.Series, with id as index, indicating whether a structural break
occurred at the boundary point for that time series (True if there was a break, False otherwise).

The test data will follow the same format. Your code will need to process these time series and generate predictions of the likelihood of a structural break for each unique time series ID.

Data Size

The training set consists of 10,000 datasets and is the same as last year's.

There will also be another 10,000 new datasets for each of the public and private test sets.

The testing data provided for local usage only consists of 100 datasets. Participants must consider that their code will run on 100 times more datasets, with a maximum limit of 15 hours of computing time.

A determinism check will re-run the infer function on 30% of the data (3,000 datasets) to ensure your model is deterministic. The results must be equal, with a tolerance of 1e-8.

Comparison against the Top 10

The participants with the highest scores in the original challenge will be used as fixed reference points. These models will only be used as a baseline for benchmarking purposes and will not be eligible for prizes.

Leaderboard Rank	Cruncher(s)	ROC AUC Score*
1st place	Humberto Brandão Mário Augusto Filho Rafael Alencar João Pedro Peinado	90.14%
2nd place	Farukcan Saglam	89.85%
3rd place	Lucas Morin	89.59%
4th place	Julian Mukaj	89.32%
5th place	Guoqin Gu Mutian Hong	89.28%
6th place	Leo Huu Dung Nguyen	89.17%
7th place	Tuah Jihan	88.34%
8th place	Abhishek Gupta	87.78%
9th place	Arina Streltsova	87.72%
10th place	Rakesh Jarupula	86.72%

*Previous edition scores.

A Spearman correlation test will be used to assess similarity with the Top 10 models. Models deemed too similar, with an out of sample correlation above 95 percent, will not be eligible for rewards.

You can watch the Talks and Award Ceremony here.