Numerical Timeseries Data

From Readings to a Sequence

Imagine a weather station recording conditions, waiting for the same amount of time, and recording them again. A single reading tells you about one moment. The complete ordered collection tells you how the measurements develop over time. That ordered collection is numerical timeseries data.

The word numerical emphasizes that the observations are measurements, such as temperature, pressure, or humidity. The word sequence emphasizes that their order matters. The word regular means that observations are collected at fixed intervals rather than at unrelated or unspecified times.

The Weather Dataset

The dataset comes from a weather station at the Max Planck Institute for Biogeochemistry in Jena, Germany. It contains 14 weather-related quantities recorded every 10 minutes. The original data reaches back to 2003, while this forecasting example uses data from 2009 through 2016.

Each time point can be viewed as a group of aligned measurements: the weather quantities recorded at that moment. Air temperature is therefore not detached from the rest of the record. It is one quantity associated with the same time as atmospheric pressure, humidity, wind direction, and the other weather measurements.

History and Forecast Target

The forecasting task separates the data into two roles. A few days of recent weather data form the historical input supplied to the model. The air temperature 24 hours in the future is the forecast target that the model is asked to estimate.

Why Order Matters

A weather record can be considered in relation to the records before and after it because the station produces observations in an ordered history. The measurements are not just isolated records with no connection. Each new observation follows an earlier observation after a fixed interval, and the accumulated history is used to frame a future prediction.

This is why sequence data is not limited to words in a sentence. Words have an order in a sentence, and weather measurements have an order in time. In both cases, the position of an item within the sequence helps describe what it means in relation to the other items.

Mistakes in Reading the Task

• Treating the dataset as an unordered collection of weather records

  Timeseries data is defined by measurements arranged in sequence at regular time intervals. The ordering is part of the information.

  Fix: Read each measurement as belonging to a particular position in the ordered history.

• Calling every weather measurement the forecast target

  The forecasting setup specifically asks for air temperature 24 hours in the future. The recent measurements serve as input evidence.

  Fix: Name the role of each part: recent weather data is the historical input, and future air temperature is the target.

• Confusing one time point with one isolated number

  Each time point can be viewed as a group of aligned weather quantities recorded at that moment.

  Fix: Associate the timestamp with air temperature, atmospheric pressure, humidity, wind direction, and the other recorded quantities.

• Assuming sequence data means only text

  A weather station also produces a sequence by recording measurements at fixed intervals over time.

  Fix: Look for ordered observations at regular intervals, whether the observations are words or numerical weather measurements.

Applying the Definition

Key Takeaways

1. Numerical timeseries data consists of measurements arranged in sequence at regular time intervals.
2. The weather dataset contains 14 weather-related quantities recorded every 10 minutes at a station in Jena, Germany.
3. The forecasting example uses weather data from 2009 through 2016, although the original data reaches back to 2003.
4. A few days of recent weather measurements are the historical input, while air temperature 24 hours in the future is the forecast target.
5. Weather measurements are sequence data because the station produces an ordered history in which each observation follows another after a fixed interval.