P3-5.3 Why Can We Not Immediately Call Raw Time Series a Learning Input¶

Section ID: P3-5.3 Version: v2026.07.20

When readers see a raw time series, many of them think like this: There are many values and they are ordered, so can't we just pass this straight in as the learning input? But we should pause here once. The mere fact that a raw time series exists does not yet mean we can say a ready-to-use learning input has been prepared.

The raw time series may contain a lot of information, but what should count as one sample, how much length should be grouped into one input, and what we want to predict may still be undecided. A learning input is not another name for raw data. It is an input structure whose boundaries are fixed to fit a problem. So the first distinction to hold onto in this section is that a raw time series exists and a learning input is ready are not the same sentence.

The shortest first split looks like this.

So a raw time series exists and a learning input is ready do not mean the same thing.

Why Just Put It In As-Is Sounds Fast¶

The idea of feeding the raw time series in as-is sounds fast for three reasons.

1. When there are many rows and columns, it already feels like a rich dataset.
2. When we hear that deep learning learns representations, it can feel as if input design can be skipped.
3. The more complex the time-series structure is, the more it seems that the model will learn it by itself better than a human summary would.

But all three are illusions that appear when defining the input structure and learning a representation are read as if they were the same thing.

What Else Is Needed to Turn Raw Time Series into a Learning Input¶

To turn a raw time series into an actual input structure, at least the following decisions must already exist.

For example, the same raw time series can become completely different input structures like these.

So a raw time series does not naturally point to one model. Depending on the problem we set up, it can be rebuilt into very different input structures. Even with the same record, whether one-action vector, segment sequence, or recent aggregate input comes first is determined by the question.

Why Summary Tables and Intermediate Representations Are Still Needed¶

A common misunderstanding here is the thought that if we are going to look at time-series models later anyway, summary tables or segment representations matter less. But summary tables and intermediate representations are not just temporary objects. They remain important for at least the following three reasons.

So the existence of a raw time series does not make the summary table unnecessary. Some questions are read more directly from the summary table first, and some may later move on again to a longer input structure.

Looking Again Through One Scene¶

Suppose an action contains 300 time-point records.

1. In Part 3, we first decide whether these 300 points form one full action.
2. We divide them into early, middle, and late segments and build a summary table.
3. If needed, we preserve more structure through an intermediate representation such as UP, FLAT, DOWN.
4. Only after that can we choose whether to view that one full action as one input, as several segment sequences, or as a recent-range aggregate.

So the correct order is not a raw time series exists -> it is immediately a learning input, but a raw time series exists -> decide what input structure to convert it into.

A Small Diagram¶

The core sequence in this section is that the story does not end at a raw time series exists. Only after sample boundaries, segment/length rules, and target labels are fixed can we finally choose an input structure.

flowchart TD
    R[Raw time series exists]
    S[Decide sample boundary]
    G[Decide segment and length rule]
    L[Decide target label]
    I[Choose input structure]

    R --> S --> G --> L --> I

What this section should hold onto is not the model type, but the order that says before calling a raw time series an input, we should first decide sample boundaries, segment criteria, and target structure. So saying that the raw time series is not yet a learning input means less the data is insufficient and more the boundaries and purpose of the input structure have not yet been specified.

Sources and Further Reading¶

• Google for Developers, Machine Learning Glossary: labeled example. Because an example is a structure where features and labels are defined together, it supports the claim that before calling a raw time series an input, the boundary of one sample and the result column should first be fixed. https://developers.google.com/machine-learning/glossary / Accessed: 2026-07-20
• Google for Developers, Machine Learning Glossary: label leakage. Because it explains a design flaw where a feature becomes a proxy for the label, it reinforces the point that unless the input structure and target structure are fixed first, part of the raw time series can be passed through incorrectly as the input. https://developers.google.com/machine-learning/glossary / Accessed: 2026-07-20
• W3C, PROV-Overview. Because the provenance framework says identifying an object, derivation, and reproducibility should be supported, it reinforces the higher-level frame that input length and segment rules should also remain reproducible as a designed structure. https://www.w3.org/TR/prov-overview/ / Accessed: 2026-07-20