AI for Time Series: Model Types (Part 1)

From the series: "The Evolution of Artificial Intelligence for Sequential Data"

In our first blog post in this series, we took a look at the nature of sequential data and the special characteristics of time series. Starting with this article, we will now focus on various approaches that can be used to model such data.

What does "modeling" mean? What distinguishes an AI model?

The goal of applying artificial intelligence in the analytical field is always to capture the information contained in the data in the best possible way, i.e., the relationships between the individual data points. An optimal model would recognize all relationships in the training data, i.e., the data from which the model "learns" (or more accurately, with which it is "trained"). Provided that the training data covers all possible cases, the trained model can also find these relationships in new data, as long as it is recorded under the same conditions. On this basis, it can decide, for example, whether the quality of a workpiece is drifting or whether the service life of a machine part is coming to an end. Since the correlations in data can be completely different in nature, there is no universal rule for constructing an optimal model. Over time, many different types or classes of models have emerged. In this series, we will limit ourselves to model classes that are suitable for time series and other sequential data.

Autoregressive sequence models

Statistically motivated ARIMA and SARIMA models are an established method for predicting time series data. They are essentially based on a correlation analysis of the data, i.e., they determine the statistical relationship between current and previous data points in the respective time series. After training, you obtain a model consisting of the coefficients or weights for the values from different time periods (this is the autoregressive "AR" part), as well as additional weights that are intended to compensate for any deviations that occur (the moving average "MA" part).

Since time series often exhibit long-term trends and seasonal fluctuations, the original ARMA model was modified so that these influences can be partially taken into account (integrated "I," seasonal "S"). However, this requires manual intervention in the data and model parameters.

In the example of the time series shown, an (S)ARIMA model would look back at the last n = 20 measured values at each point in time in order to calculate a prediction of the next value. Which of the 20 values are actually used depends on the weighting of the respective time intervals determined by the training. It is intuitive to assume that the fixed weighting of certain time intervals will not always lead to an optimal prediction. The advantage of this model type is the traceability of the predictions and the comparatively simple training.

In summary:

• (S)ARIMA models map the statistical correlation between data points from different time periods in the past and the present.
• They are easy for humans to interpret.
• Fixed model parameters lead to suboptimal predictions over time, while dynamic parameterization requires manual effort.
• Overall, (S)ARIMA models are not as flexible as deep learning methods, which we will discuss in upcoming articles.