Types of Prediction

Prediction can be classified according to various criteria. The basic criteria are:

  • the data we have available for training and prediction
  • what we want to predict - a value or a trend

Predicting value or trend

When we want to obtain the exact future value (or several future values) of a variable, we are predicting a value. Another possibility is to predict the trend of a variable, that is, whether the value will go up or down without considering the size of the change - in that case, we are predicting a trend. Trend prediction is in fact a classification into two (or three) classes: up, down, or no significant change. Predicting a nearby value is generally easier than predicting a trend. Besides the trend, we may also want to predict related quantities such as changes in a moving average.

Data for prediction

For time-series prediction, we usually have values of a variable sampled at equidistant intervals. We can then try to predict the future development of the value based only on historical values and time. In this case, the historical time series should be sufficiently long and dense.

We may also have additional information about the time series, such as its derivative. This information can then be used to improve prediction accuracy. Important information can also be added using so-called interventional variables (intervention indicators), which represent information about the time series or about the period for which we are making a prediction. For example, when predicting energy consumption, knowing whether we are predicting Monday or Saturday can improve the result dramatically - this information does not follow from the time series explicitly and must be added. It is often very helpful to use intervention indicators when creating a model for prediction.

We can also have information about other related variables, preferably also in the form of time series. From the history of related variables we can infer information about other variables. The relationship can be expressed in various ways. One example is a static (or slowly changing) sum of two variables. The dependency does not have to be expressed explicitly - for example, stock prices within one sector may be related, but the dependency can be hard to express computationally. This kind of information is selected in the hope that it will correlate with the predicted value, although we cannot be sure in advance. The field of data mining can help with selecting appropriate information and interpreting it.

ยท