8 min read
Getting Started with Qlik Predict: Forecasting Time Series
Who wouldn't want to predict the future? In the first three parts of this blog series, we explained step by step how you can build predictive models with Qlik Predict (formerly AutoML). We started with feature engineering, where we showed how you can 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 deploy it in dashboards and processes. Up to now, we've worked with three problem classes: binary classification, multi-class classification, and regression. But since October 2025, there's a new player on the field: multivariate time series forecasting. In this blog, we walk you through this new capability and show how you can use trends and patterns to look ahead.
What Is a Time Series
Before we dive into the technical details, it's important to understand what we mean by a time series. A time series is a series of observations ordered in time. Think, for example, of the number of sales per day in a store. At first glance, that seems simple: you count every day how many products go over the counter. But look a bit further, and you'll see that these figures don't stand on their own. The number of sales on a given day is influenced by all kinds of factors: the weather, the season, holidays, and of course promotions and sales.
On a sunny day, more ice cream is sold, while rainy days make umbrellas popular instead. Around Christmas, demand for luxury products rises, while January brings the hustle of the sales season. These kinds of patterns are what we call seasonality.
We also often see trends: a product that's slowly becoming more popular, or is being sold less and less. Sometimes there's a cycle, such as recurring peaks on weekends. This sequencing and these patterns allow us not only to look back and explain past sales numbers, but also to predict forward. The more we understand this dynamic, the better we can anticipate what's coming.
Forecasting Single Time Series in Qlik
Forecasting time series has already been possible in Qlik Sense for a while, via the line chart object in Qlik Cloud. To create a time series forecast, your chart needs to meet three conditions:
- You have exactly one measure and one dimension
- The dimensional value is numeric or a time interval
- The dimensional axis is configured as continuous
Suppose you have a chart with revenue per day. As soon as you set the dimension to a continuous time axis and add one expression (for example, revenue), you can activate forecasting.
As soon as you turn on the forecast, you get a choice between 2 forecasting methods: SSA (Singular Spectrum Analysis) or OLS (Ordinary Least Squares). The simplest of the two is OLS. In short, OLS comes down to drawing 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 pattern. You can already use OLS when your chart contains 2 points, but it doesn't actually give a very good forecast if your data has, for example, seasonal patterns.
A somewhat smarter forecast is the 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 uses these patterns as effectively as possible to predict the next values. This method is therefore much better able to handle seasonal patterns. The downside is that you need more points. You can activate SSA with at least 6 points, but the more points you have, the better the forecast will be. In addition, SSA gives you the option to show a so-called confidence interval. If you turn this option on, you'll see a band around your line. Set the confidence interval to 0.95, and in 95% of cases the future value will fall between the upper and lower bounds of that band.
Forecasting Multivariate Time Series with Qlik Predict
But what if you want more than a simple trend line? Suppose you don't just want to know how much you'll sell tomorrow, but also want to take factors such as promotions, store locations, and product groups into account. 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're used to from 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 indicate which column contains the date index, so Qlik knows what the time unit is.
After that, you define groups, such as store locations or product categories. These help the model learn patterns per segment. By choosing a good set of groups, you not only make the predictive model smarter, it also means you'll be able to easily create sub-forecasts for specific store locations using selections later on.
Next, you add future features: variables that influence future values, such as planned promotions or holidays. Think of our example: if you know there's a discount campaign next week, you want the model to factor in that information.
For the time series, set which field contains the date
and which groups and predictive fields the dataset has
Then you set the forecast horizon: how many steps ahead do you want to forecast? And the forecast gap size: how many days after the last observation does 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've made these settings, the process runs the way you're used to from Qlik Predict. By clicking "Run experiment," you give Qlik the signal to train predictive models with all the chosen 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
Once your experiment is finished, you get an overview of the performance per 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 downside of MAE is that this deviation is sensitive to scale. An error of 10 units might be negligible if you sell 10,000 products, but enormous if you only sell 50.
To solve this, there's the MASE. To calculate the MASE, a fictional second forecast is made, a so-called "naive forecast." This naive forecast takes, for each next step, the value of the step before it, in other words, "just predict the same number as yesterday." The MASE is then calculated by dividing the MAE by the MAE of the naive forecast.
Besides the model's accuracy, the speed of the forecast is also relevant. If you've trained a binary classification model, you can generate a new prediction based on 1 observation, but with time series you'll always need to provide a series of observations for a good forecast. That means a lot more computation is needed for a new forecast. If you want to build an interactive dashboard with, for example, a "what-if" analysis, it can 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
Conclusion
Time series forecasting opens up a world of possibilities. Whether you want to anticipate busy periods in your stores, plan energy consumption, or optimize inventory levels: Qlik Predict makes it easy to model complex patterns and turn them into action. Want to know more? Or would you like to see Qlik Predict in action? Get in touch with us.
Written by Lennaert van den Brink
Cluster Manager/Senior BI Consultant