Sliding-Window Supervision for Time Series

Ordinary supervised learning needs a table of independent rows, each with its own inputs and a label. A time series is just one long sequence of numbers with no such rows to begin with — so how do you turn a single scroll of daily prices into anything a standard model, or even a neural network, can train on?

The analogy: flashcards from a scroll

Imagine unrolling a long scroll of text and making flashcards: each card shows the last few lines as the "question" and the very next line as the "answer," then you slide one line down the scroll and make the next card. A sliding window does exactly this to a time series — cut out a fixed-length window ending at time $t$ as the input, pair it with the value at $t+h$ (some steps ahead) as the target, then slide forward by one step and repeat.

The mechanics

Given a series $x_1, \dots, x_T$, a window length $L$, and a forecast horizon $h$, each training example is:

$$X_t = (x_{t-L+1}, \dots, x_t), \qquad y_t = x_{t+h}$$

In words: the input is the last $L$ observed values ending at time $t$; the label is the value $h$ steps into the future from that same point. Sliding $t$ forward one step at a time (or by some larger stride) across the whole series produces many overlapping training pairs from a single sequence.

Worked example 1: building the pairs

An 8-day price series: $10, 11, 13, 12, 14, 15, 14, 16$. With window $L=3$, horizon $h=1$, stride $1$: pair 1 uses $X=(10,11,13)$, $y=12$; pair 2 uses $X=(11,13,12)$, $y=14$; pair 3 uses $X=(13,12,14)$, $y=15$; pair 4 uses $X=(12,14,15)$, $y=14$; pair 5 uses $X=(14,15,14)$, $y=16$. Five overlapping training examples from one 8-value series.

Worked example 2: increasing the stride

Same series, same $L=3$, $h=1$, but stride $2$: pair 1 uses $X=(10,11,13)$, $y=12$ (starting at $t=3$); pair 2 uses $X=(13,12,14)$, $y=15$ (starting at $t=5$, two steps later). Only 2 pairs instead of 5 — fewer, but far less overlapping, since each new window shares less data with the previous one. Larger stride trades training-set size for less redundant, more independent-looking examples.

What this means in practice

This is literally how raw price or return history becomes a training set for a return-prediction model. Window length $L$ is a real hyperparameter: too short and the model lacks context; too long and old, possibly stale information dilutes recent signal. Because of the overlap, a train/test split must respect time order — never shuffle windows randomly the way you would independent tabular rows — and typically needs a gap ("embargo") between the end of the training windows and the start of the test windows.