Sliding Windows for Forecasting

Turn a series into a supervised learning dataset with sliding windows.

To train a regressor or a neural network to forecast, you need pairs of inputs and targets. Each input is the last lookback values, and each target is the horizon values that come right after them. Slide the window forward one step at a time:

X[j] = series[j : j + lookback]
y[j] = series[j + lookback : j + lookback + horizon]

There are n - lookback - horizon + 1 windows for a series of length n.

Function signature

def make_windows(series, lookback, horizon):

Example

Input

make_windows(np.array([1.0, 2.0, 3.0, 4.0, 5.0, 6.0]), lookback=3, horizon=1)

Output

X = array([[1., 2., 3.],
           [2., 3., 4.],
           [3., 4., 5.]])
y = array([[4.],
           [5.],
           [6.]])

Reasoning

The first window sees 1, 2, 3 and must predict 4. Each later window moves one step. The last full window sees 3, 4, 5 and predicts 6. There are 6 - 3 - 1 + 1 = 3 windows.

Constraints and notes

• Return (X, y) with shapes (m, lookback) and (m, horizon).
• If the series is too short for one window, return arrays with 0 rows and the right number of columns.
• The windows must never let a target value appear in its own input. That is the kind of leak that makes a forecaster look good in testing and fail in use.