Questions / Quant research, Time-series reasoning

A model can read backwards without seeing the future

Challenge, about 8 minutes. Question R020.

At time T, a bidirectional sequence model reads only the already-observed window [T−99,T] to forecast T+1. Can this setup be free of future leakage?

  1. ANo; reversing historical inputs makes future target labels available.
  2. BYes; both passes can use only the observed historical window.
  3. CNo; the backward pass necessarily reads observations beyond time T.
  4. DYes; the backward pass reverses time, which cancels any look-ahead.

Answer it in the app It's free. You get hints, a second try if you miss, and a follow-up question.

Show the answer and worked solution

Answer: B. Yes; both passes can use only the observed historical window.

  1. At time T, every input timestamp from T−99 through T is already observed.
  2. A backward pass may start at T and traverse that same history. This architecture alone adds no unavailable observation; the feature and target pipeline still needs checking.

Interview shortcut. Bidirectional is fine if every input timestamp is already observed.

Takeaway. Audit the information set, not the direction name.

Why the other choices fail

  • No; reversing historical inputs makes future target labels available. Reordering already observed inputs does not create future labels. Leakage depends on input construction and availability.
  • No; the backward pass necessarily reads observations beyond time T. Backward processing describes the order within the supplied sequence, not access to timestamps outside that sequence.
  • Yes; the backward pass reverses time, which cancels any look-ahead. Reversing the order of inputs does not cancel look-ahead. This setup is safe because its window ends at T.

In the app this has a follow-up where one thing about the setup changes. Try the follow-up.

More time-series reasoning questions