POOL: AI-Based Forecasting with External Factors
July 29, 2025
Show Transcript
Hello and welcome to “future”. In this video, I’ll show you how you can use artificial intelligence to find suitable external influencing factors for your use cases and use them for your forecasts. What I’m showing you today are ways in which you can work with futureNOW. Our provided frontend allows for a particularly easy entry into data-based forecasting. However, what you learn in this video can also be used if you want to integrate an automated solution into your workflow. For this, the comprehensive toolbox futureEXPERT is available, which is easiest to use with the Python client.
I’m starting from a finished CHECK-IN. I’ve prepared my dataset here. My data contains the sales figures for various products in a small ice cream parlor in Würzburg. I select the product for which I want to create forecasts, click on “Continue with”, and go to “Forecast via NOW”. There, I confirm the forecast object again and am now in the forecast module.
In this window, I can select for how many days forecasts should be created and could also choose a different forecast object here.
That’s all I need to do to start making forecasts. However, I have the option to add indicators or influencing factors as covariates to my configuration, which I want to make use of.
By clicking on the “Plus” icon, I choose where to get the influencing factors from. I can select a dataset from my uploaded data or another time series that I have prepared with CHECK-IN. Today, we’re looking at the possibilities offered by the POOL. The POOL is a wide selection of prepared external influencing factors provided by us, which are useful for many different applications.
Therefore, I click on “POOL”. A window opens, and we see the various sources from which we can obtain the indicators – for example, the German Federal Statistical Office (Destatis), the OECD, but also calendar data, school holidays, or weather information. One can now click on one of these sources and get an overview of the various influencing factors from that source. I can also remove the respective source from the source filter and then see all influencing factors. Or I can enter a specific term into the search field, for example, “temperature” or “holidays”, and we see various indicators that match the respective search query.
In total, we have a selection of 170 different indicators here. I don’t want to go through all of that now; it’s a bit too much work for me, and that’s exactly why there is the ‘AI selection’ feature.
The ‘AI selection’ is a feature that uses a Large Language Model, which we developed to quickly find suitable influencing factors for the use case that my time series reflects, via a natural language description.
I specify here that my data concerns “Ice cream sales in my ice cream parlor in Würzburg”. The AI Selection is able to extract geographical information from the prompt and incorporate it into the scoring of the influencing factors. I could also turn this feature off and specify the geographical data via the settings. But we see here that the tool correctly identified Europe, Germany, Bavaria, Würzburg. In general, the more precise I am with my context description, the easier it is for the AI to find the truly well-fitting factors.
It now searches through all influencing factors from the POOL in the background and creates a score for each one. The score is the AI model’s assessment of how well the respective influencing factor fits the description of my data. Furthermore, it automatically detects how the influencing factor relates to my data – in the form of the direction of influence.
Now I see a pre-selection of the indicators that have the highest potential, sorted by their scores. From this list, I then select a subset. I want to use the school holidays, the temperature (daily maximum value), and then also the public holidays. These are three different influencing factors that, in my opinion, fit the time series very well.
Now I click on “Add indicator”, and it is shown below that they have been successfully applied. Thus, I can create forecasts that are calculated taking the respective covariates into account. The whole process takes a moment, and the results should be available shortly. As soon as they are ready, I can view the forecasts in a plot and download them as a CSV.
The AI Selection helped me to quickly identify data that contextually fits my use case and potentially improves the quality of my forecasts. Give it a try with your own use case! Have fun experimenting, and see you next time.
External data for better forecasts: The POOL
To improve forecast accuracy, it is often crucial to consider external factors. The POOL is a wide range of prepared external data provided by us, useful for many use cases. This includes weather data, public holidays, school vacations, or macroeconomic indicators.
In the FORECAST module, you can easily add this data to your forecast. To do this, click the Plus icon in the Indicators field and select POOL as the data source.
Table of Contents
How to find the relevant influencing factors
You have two ways to find the most relevant data for your forecast:
Option 1: Manually search and filter
This approach is ideal if you already have an idea of which data you need.
- Filter & Search: Use the filters to narrow down the selection.
- Filter by source: Click on a data source (e.g., “Weather data,” “School holidays,” or “Federal Statistical Office”) to see all related indicators.
- Free-text search: Enter a term (e.g., “temperature,” “inflation”) into the search bar to get targeted suggestions.
- Select indicators: Choose one or more suitable indicators from the results list.
- Add: Click Add indicator to include the selection in your forecast configuration.
Option 2: AI-powered recommendations (AI selection)
This approach is perfect if you want a quick, data-driven recommendation for suitable covariates. The AI Selection uses a Large Language Model (LLM) to suggest the most relevant indicators based on a description of your use case.
- Start AI Selection: In the POOL window, click the AI Selection button.
- Describe use case: Enter a short, precise description of your time series into the text field (e.g., “Sales figures for ice cream in my parlor in Würzburg”). The more precise the description, the better the recommendations.
- Provide geographical information: The AI can extract geographical information directly from your description. Alternatively, you can define it manually via the settings (cogwheel icon).
- Review recommendation & add: The AI analyzes all available data and provides a list of suggestions sorted by relevance. Each suggestion includes a score reflecting the estimated relevance for your use case. Select the appropriate indicators and click Add indicator.
Further information & links
Here you can find more resources to get the most out of your data.
MATCHER analysis: The MATCHER helps you with statistical analysis and assessing the predictive power of potential covariates. Learn in our notebook how to automatically find the best covariates for your time series.
Quickstart: Our Quickstart tutorial shows you how to create your first forecast in under 5 minutes.
Blog: Discover interesting use cases and valuable tips in the articles on our blog.
Videos: An overview of all our video tutorials helps you get to know other features.
Questions, problems, or ideas? Feel free to contact us by email at support@future-forecasting.de. For new feature requests, you can also directly create an issue in our GitHub Repository.
With your input, we can make future even better!
We look forward to hearing from you!