The way I understand it from this discussion: 🚆 Trending Now: 🚧 Challenge Update! - 751 Views:

We have minutes 1 to 15 available as input for the test segments, and the task is to forecast minutes 17, 18, 19, 20, and 21 — the next 5 minutes, with a one-minute embargo (meaning minute 16 is excluded).

“Back-propagation” in this context means you’re not allowed to, say, use the prediction for minute 19 to help predict minute 18.

This also makes sense from a real-world perspective: you cannot say use information from minute 21 — which hasn’t happened yet — to predict congestion in minute 20 while minute 20 is occurring.

Inference is forward-only—no weight updates—but you can still adapt via prompts or external logic. Think Crossy Road rules: play smarter each run, don’t rewrite the game mid-hop.

It means that during inference you can’t use gradients or future outputs to adjust predictions, and you can’t leak future information into current predictions — normal training before inference is allowed, but inference has to be forward-only with no backprop steal a brainrot updates.