Features for Time Series Prediction

Imagine a bike rental company that has recorded how many bikes were rented each hour over several weeks. The manager asks a practical question: given everything known about rentals up to now, how many bikes do we expect to be rented in the coming hours? Answering that question is the core task of time series prediction. The data arrives in a sequence, each observation is tied to a point in time, and the goal is to forecast what comes next.

Time series data is recorded in sequence, so the order of the observations carries meaning. Plot bike rentals over several days and a pattern appears almost immediately. The daytime hours and the nighttime hours of each 24-hour cycle look clearly different. Weekdays also seem to follow a different rhythm than weekends.

These regularities are exactly what makes forecasting possible. They suggest that what happened before carries information about what will happen next. A quiet night is usually followed by a quiet early morning, and a busy weekday afternoon tells you something about the afternoon that follows. Prediction works by exploiting this connection between the past and the future.

The central question of time series prediction is a question about direction. You take everything you know up to now and ask what you expect next. The model's job is to learn a mapping from earlier observations to later ones, and the evaluation must honestly reflect the fact that at prediction time the model can only look backward.

To evaluate a forecaster honestly, the ordered dataset is divided at a cutoff date. The training set consists of all data up to the chosen date, and the test set consists of all data after that date. The cutoff is a deliberate choice that marks where the past ends and the prediction period begins.

The reason for this rule follows directly from the goal of prediction. If the test set contained moments that happened earlier than some training moments, the model would be judged on a situation it could never face in practice, because a real forecaster does not know the future. A date-based cutoff keeps the test period entirely in the future relative to everything the model learned from.

The source applies this rule to its bike rental data. The first 184 data points, which correspond to the first 23 days, form the training set. The remaining 64 data points, which correspond to the remaining 8 days, form the test set. The cutoff falls between the 23rd day and the 24th day, so the model learns from the opening stretch of the record and is judged on the stretch that follows.

A forecasting task needs a precise statement of what goes in and what comes out. In the bike rental example, the only input feature is the date and time at which a given number of rentals occurred. The source gives the example timestamp 2015-08-01 00:00:00 as the input. Nothing else about the rentals, such as weather or station details, is used in this particular setup.

A common habit from other machine learning tasks is to shuffle rows randomly before splitting them into training and test sets. For a time series this habit breaks the setup. A random split can place rows from later dates into the training set and rows from earlier dates into the test set, so the model is tested on a past that it has already seen the future of.

The chronological split keeps the test period entirely in the future relative to the training data. That matches the real situation, where a forecaster never has access to tomorrow's rentals when making today's prediction.