Improving Demand Forecasting Models with Event Features
Learn how to integrate PredictHQ's event features into your demand forecasting model.
The inclusion of PredictHQ's event data can significantly enhance the performance of machine learning models, such as those used in demand forecasting. For example, the delivery company Favor reported a reduction of 5-6% in their forecasting error, with others noting decreases of up to 5-10% or more.
This tutorial will guide you through the process of identifying, retrieving and integrating relevant, demand-driving event data as features into a demand forecasting model. Follow along by running the accompanying Jupyter notebooks while referring to the provided links for more technical details. The tutorial concludes with a practical example of a demand forecasting model that incorporates PredictHQ event features.
Use Cases
Demand Forecasting
Relevant Industries
Accommodation, Consumer Packaged Goods, Grocery and Supermarkets, Leisure, Travel and Tourism, Marketing and Advertising, Parking, Restaurants, Retail, Transportation and Delivery and Others
Harnessing Event Signals
Events Driving Demand
Events, such as concerts, expos and public holidays, are known to affect consumer behavior and drive demand. PredictHQ offers event data across more than a dozen categories, featuring a wide range of labels. The powerful data processing pipeline ensures the delivery of high-quality, enriched event data that can be seamlessly incorporated as features into any demand forecasting model.
Integrating Event Features
Built upon extensive event coverage, PredictHQ’s event features aggregate similar events into predefined groups for specific locations at set intervals, such as daily aggregations. These prebuilt, forecast-ready features can be added directly to machine learning models without further preprocessing. Access to an extensive library of features is available through the Features API. We recommend starting with the Important Features identified by the Beam API.
How-To Guide
The sections below guide you through integrating event features into your demand forecasting models. Follow these instructions and run the accompanying Jupyter notebooks to understand how you can adapt this approach to fit your workflow and improve the accuracy of your models.
Overview
Adding event features to a demand forecasting model involves straightforward steps. These include pulling a list of Important Features from the Beam API and retrieving prebuilt, forecast-ready features from the Features API for a store or location.
Most steps are handled by PredictHQ APIs; you just need to provide the following for each store or location:
Historical demand data
Latitude and longitude
Industry
Step 1. Select Relevant Event Features
With countless events taking place globally throughout the year, identifying events that impact demand at your location is crucial. The Beam API automatically provides a list of Important Features based on your historical demand data and location. Alternatively, you can access Beam via Demand Analysis in Control Center and directly copy the Important Features from your browser.
There are two main strategies for determining a list of Important Features for a store or location: Important Features tailored specifically to the store or location, or Important Features based on a group of stores or locations. See below and choose the approach that best suits your operational needs.
For technical details, visit:
For practical implementation:
Step 2. Get Features
The Features API provides access to a library of prebuilt, forecast-ready features ready for direct integration into your machine-learning models. Simply specify the date range, location and list of features, all of which can be sourced from the Beam API.
Responses from the Features API vary based on the type of feature. Most come with a suite of statistics that indicates how the underlying event data is aggregated daily for a location, e.g. sum, max, count. For phq_rank_* features, the response is the daily number of events for each of the five rank bands. We recommend the following aggregations:
| Feature Type | stat | other |
|---|---|---|
| sum |
|
| max |
|
|
| rank-weighted sum |
| sum |
|
| max |
|
Demand Impact Patterns
Some features consider the additional impact from events before and after scheduled dates, offering a more accurate representation of event impacts on demand. This is known as Demand Impact Patterns (DIP), which vary by event category and industry. Features including DIP are denoted with an industry suffix, such as phq_attendance_sports_accommodation or phq_impact_severe_weather_air_quality_retail.
Step 3. ML Model and Future Predictions
Event features provided by the Features API are prebuilt, forecast-ready and ready for immediate use. They can be integrated into your existing dataset by merging based on location ID and date. Incorporating these event features can enhance your model's performance by adding valuable demand-driving event data.
For future predictions, you can access forward-facing data, such as the next two weeks or the upcoming month, by querying the Features API. Work closely with your engineering team to ensure these new features are effectively incorporated into your production pipeline.
For practical implementation:
Conclusion
By following this tutorial, you should now understand how to enhance your demand forecasting models by integrating PredictHQ's event features. As known drivers of demand, incorporating event signals can noticeably improve the accuracy of your forecasts, empowering you to make more informed decisions and strategize more effectively.
Last updated
