Getting started with Qlik Predict: forecasting time series

Predicting the future: who wouldn't want that? In the first three parts of this blog series, we explained step by step how to build predictive models with Qlik Predict (formerly AutoML). We started with feature engineering, showing how to enrich your dataset with smart features. Then we dove into model selection, discussing how to choose the best algorithm for your problem. In part three, we explained how to operationalize a trained model and put it to work in dashboards and processes. So far, we have worked with three problem classes: binary classification, multiclass classification and regression. But since October 2025, there has been a new player on the field: multivariate time series forecasting. In this blog, we walk you through this new functionality and show how you can use trends and patterns to look ahead.

What is a time series?

Before we dive into the technology, it is important to understand what we mean by a time series. A time series is a sequence of observations ordered in time. Think, for example, of the number of sales per day in a shop. At first glance, that seems simple: every day you count how many products go over the counter. But look a little further and you will see that those figures do not stand on their own. The number of sales on a given day is influenced by all kinds of factors: the weather, the season, public holidays and, of course, campaigns and promotions.

On a sunny day, more ice creams are sold, while rainy days make umbrellas popular. Around Christmas, demand for luxury products rises, while January is dominated by the sales rush. We call these kinds of patterns seasonality.

We also often see trends: a product that is slowly becoming more popular, or one that is selling less and less. Sometimes there is a cycle, such as recurring peaks at the weekend. This sequential nature and these patterns allow us not only to look back and explain sales figures, but also to predict what lies ahead. The better we understand these dynamics, the better we can anticipate what is coming.

Forecasting single time series in Qlik

Forecasting time series has been possible in Qlik Sense for some time, via the line chart object in Qlik Cloud. To create a time series forecast, your chart must meet three conditions:

• You have exactly one measure and one dimension
• The dimension value is numeric or a time interval
• The dimension axis is configured as continuous

Suppose you have a chart showing revenue per day. As soon as you set the dimension to a continuous time axis and add one expression (for example, revenue), you can enable forecasting.

Once you turn on the forecast, you can choose between two forecasting methods: SSA (Singular Spectrum Analysis) or OLS (Ordinary Least Squares). The simpler of the two is OLS. In short, OLS draws a straight line through all the points in your chart to predict the next points. This method is especially suitable for forecasting datasets that follow a relatively linear course. You can use OLS as soon as your chart contains 2 points, but it does not really give a very good forecast if your data has, for example, seasonal patterns.

A somewhat more intelligent forecast is SSA. The SSA algorithm splits the line in the chart into short segments. For each of these segments, it determines 3 components: the trend, various cycles and noise. Based on those 3 components, SSA constructs a new line that makes the best possible use of these patterns to predict the next values. This method is therefore much better at handling seasonal patterns. The downside is that you need more points. You can enable SSA with at least 6 points, but the more points you have, the better the forecast will be. SSA also lets you display a so-called confidence interval. If you turn this option on, you will see a band around your line. If you set the confidence interval to 0.95, the future value will fall between the upper and lower bounds of the band in 95% of cases.

Forecasting multivariate time series with Qlik Predict

But what if you want more than a simple trend line? Suppose you not only want to know how much you will sell tomorrow, but also want to take into account factors such as promotions, store locations and product groups. Then you need a multivariate time series model. Building such a multivariate time series model is now possible with Qlik Predict. The process is similar to what you are used to with other problem classes, but with a few specific steps.

You start by choosing your dataset and selecting the target: the column you want to predict, for example ‘Number of sales’. Next, you specify which column contains the date index, so Qlik knows what the unit of time is.

Then you define groups, such as store locations or product categories. These help the model learn patterns per segment. Choosing a good set of groups not only makes the predictive model more intelligent, it also means that later on you can easily use selections to create partial forecasts for specific store locations.

Next, you add future features: variables that influence future values, such as planned promotions or public holidays. Think of our example: if you know there is a discount campaign next week, you want the model to take that information into account.

For the time series, specify which field contains the date
and which groups and predictive fields the dataset contains

Then you set the forecast horizon: how many steps ahead do you want to predict? And the forecast gap size: how many days after the last observation should the forecast start? Finally, you choose the algorithms you want to train. Just as with classification and regression, you can run multiple models and compare them later.

Once you have made these settings, the process runs just as you are used to in Qlik Predict. By clicking “Run experiment”, you give Qlik the go-ahead to train predictive models with all the selected algorithms.

The forecast window and the forecast gap indicate from
which date and how many days ahead you want to forecast

Choosing the right model

When your experiment is finished, you get an overview of the performance of each model. Qlik naturally offers several metrics for this, of which MAE (Mean Absolute Error) and MASE (Mean Absolute Scaled Error) are the most important.

MAE is simple: for each point, Qlik calculates the average deviation between the forecast and reality. The lower, the better. The drawback of MAE is that this deviation is sensitive to scale. An error of 10 units may be negligible if you sell 10,000 products, but huge if you only sell 50.

MASE solves this. To calculate the MASE, a fictitious second forecast is made, a so-called “naïve forecast”. This naïve forecast takes the value of the previous step for each next step, in other words “just predict the same number as yesterday”. The MASE is then calculated by dividing the MAE by the MAE of the naïve forecast.

Besides the accuracy of the model, the speed of the forecast is also relevant. If you have trained a binary classification model, you can generate a new prediction based on a single observation, but with time series you always need to provide a series of observations for a good forecast. That means a lot more computation is needed for each new forecast. If you want to build an interactive dashboard with, for example, a “what-if” analysis, it may be smart to choose a slightly faster but less accurate model over a slower but slightly more accurate one.

Based on the settings, Qlik tells you which model can generate the best forecast