Structural Break Real-Time — is any legal online-horizon / length metadata available at serve?

Competition ADIA Lab
1

Hi,

For Structural Break Real-Time, x_online is a streaming generator and I cannot call len(x_online) (and the W23 fix correctly closed full-series leaks in the cloud).

Docs say the online segment is typically between 10 and 1,000 observations, and the historical segment is fully available at the start.

Question: at inference time, is any of the following legally available to the submission (as an argument, file, metadata field, or similar), or is the online horizon intentionally unknown until the stream ends?

  1. Exact online length NN for the current series
  2. A coarse horizon / length bucket (e.g. short / medium / long)
  3. Any other per-series length or timing metadata beyond the historical values and the streaming online points

Context: with true online NN, simple phase features like t/Nt/N materially change ranking under TS-AUC; without it, any shared length prior is identical across series at a fixed tt and does not help cross-sectionally. I want to confirm whether recovering NN (or a legal surrogate) is part of the intended challenge, or whether length metadata exists that I should be using.

Thanks!
Nuno

2

Hello Nuno,

The length of the x_online object is not accessible by design.

In the real world, you don’t necessarily know when a stream of data ends.

Attempting to gain an advantage by exploiting a vulnerability will result in severe punishment, as it will be considered cheating. However, we highly appreciate it when you responsibly disclose it to us like what Brandao did for the original leak.

3

Thank you for your answer Enzo.