TimeDelay Loop Start

Start of a sliding-window feedback loop for iteratively generating time-series rows. Together with the TimeDelay Loop End node it builds autoregressive workflows in which each new row is computed from the previous N time steps — typical use cases are multi-step forecasting and synthetic time-series generation (e.g. recursively producing a Fibonacci sequence from a two-row seed).

The seed table at the input port must contain at least N rows, where N is the configured size of the time set. On the first iteration the node emits a single wide row holding the last N rows of the seed table side by side: every input column c is unrolled into N lagged columns c(t - N-1), c(t - N-2), ... , c(t - 0) , ordered from oldest to newest. The body of the loop is expected to compute the next time step from these lagged features and pass it through the TimeDelay Loop End node, which records the new row in the loop's accumulated output and feeds it back to this node. On the next iteration the window slides forward by one step — the oldest time step is dropped and the just-generated row becomes the new t = 0 . The loop runs for the configured number of iterations, producing that many generated rows in total.

Input ports

1. Type: Table
   Initialisation table
   Seed time-series table. Must contain at least as many rows as the configured size of the time set; the last N rows are used to initialize the sliding window. The columns of this table also define the schema that the row fed back through the TimeDelay Loop End node has to match.

Output ports

1. Type: Table
   Data row
   A single row holding the current sliding window: every input column c is unrolled into N lagged columns c(t - k) for k = N-1 , ... , 0 , ordered from oldest to newest. On the first iteration these cells come from the last N rows of the seed table; on later iterations the window has slid forward by one step, with the oldest time step dropped and the row produced by the loop body in the previous iteration appended as the new t = 0 .