Working with predictions for time series models
After deploying a time series model, you can use it to create predictions on new data. To create time series predictions, prepare an apply dataset containing the required data and structure. For full details, see Preparing an apply dataset.
Prediction mechanism
The process for time series model predictions has some differences from classification and regression models.
Classification and regression models predict future values for an entire column that is non-existent or unknown in the apply dataset. For example, suppose you have a binary classification model that predicts future values of a Churned column. The apply dataset will not typically contain a Churned column, and if it is present, it is ignored—instead, the values for this column are generated by the model as predictions.
On the other hand, with a time series model, predictions are created as rows rather than columns. The predictions are still generated for a target column, but they are records corresponding to specific future time values. The structure of the columns does not change between the training and apply datasets.
Also, an apply dataset for a classification or regression model should contain only new data that the model has not seen yet, for which you want to generate predictions. On the other hand, the apply dataset for a time series model needs to contain some historical data corresponding to the apply window for the model. These requirements are further specified in Working with predictions for time series models.
Available tools for creating an apply dataset
The following tools are available in Qlik Cloud to help you build and dynamically generate your datasets:
-
Qlik scripting, available in the Data load editor in apps, as well as the standalone script editor accessible from the Analytics activity center.
-
Table recipe, available as a standalone tool from the Analytics activity center. Table recipe is designed for creating single-table datasets, such as those used in machine learning.
-
Data flow, available as a standalone tool from the Analytics activity center.
-
Data manager in analytics apps.
Commonly needed operations
Changing feature types
It is possible that when you prepare your apply dataset, you will need to transform columns to modify their feature types.
The most common example is if you have a numerical column that you have configured as a group in your model. When you select a feature as a group within your experiment, its feature type is automatically changed to categorical if not already, so that it can be handled as a group. When you add your apply dataset for predictions, the column data still likely contains numeric feature data, so you will need to change the column to string data so it will be identified as a categorical feature. This transformation can easily be done using Qlik scripting — using the text() function — or with other Qlik data preparation tools such as table recipe and data flow.
For an example using Qlik scripting, see Preparing the dataset.
Limitations
-
Data drift monitoring is not available for time series models.
-
SHAP datasets cannot be generated during predictions with time series models.
-
The Qlik Predict analytics connector does not support time series models.
For limitations related to time series experiments, see Limitations and considerations.