2025 Changelog
Last updated
Was this helpful?
Predicted Impact Patterns for the Retail industry have been updated for US public holidays, observances, and school holidays using data-driven analysis of real retail demand data. The updated patterns deliver absolute MAPE improvements of 0.69–1.02 percentage points compared to the previous patterns - a meaningful improvement for customers in grocery, pharmacy, and general merchandise retail.
Machine-readable OpenAPI specifications are now available for all PredictHQ APIs: Events API, Features API, Beam, Saved Locations, Forecasts API, Suggested Radius, and Places. Specs are SwaggerUI-compatible and published to the PredictHQ public GitHub repo. Use them to generate client libraries, power IDE tooling, or integrate with AI-assisted development workflows.
A significant improvement to school holiday data quality: the false negative rate for school holiday deduplication has been reduced from 53% to under 1%. Previously, predicted school holiday events were not correctly matched to active events when start dates differed by more than a few days. The revised model now links predicted and active events within a 3-month window, resulting in cleaner school holiday data across all markets.
The Forecasts API now supports polygon-based Saved Locations as the location input for forecast models. Customers using custom polygon catchment areas can now use the same location definitions consistently across Events API, Features API, Beam, and Forecasts API.
You can now pass a Beam analysis_id directly to the Features API to automatically apply the correct location, radius, rank filters, and relevant categories from that analysis. This removes the need to manually reconstruct analysis parameters when querying the Features API. Group Analyses are also supported. See the Features API docs for the beam.analysis_id parameter.
Predicted Impact Patterns for the Retail industry have been rebuilt for attendance-based events - concerts, sports, conferences, expos, festivals, and performing arts - using data-driven analysis of real retail demand data. The updated patterns deliver a 5.2% mean relative MAPE improvement across retail analyses, with 86% of analyses showing positive improvement.
The Features API now supports filtering by Local Rank (local_rank.*) in addition to PHQ Rank. Local Rank adjusts event significance based on the population density of the surrounding area, making it more appropriate for customers operating across locations with highly variable local contexts — such as a mix of urban and regional stores. This filter is available for phq_attendance, phq_spend, phq_viewership, and phq_impact feature families.
New phq_impact_* feature families are now available in the Features API for Public Holidays, Observances, School Holidays, and Academic events. These categories previously only had phq_rank-based features. The new impact features provide a demand-weighted signal that is more directly usable in forecasting models, reducing the need to manually engineer features from rank values.
The 2025–2026 US school holidays calendar has been added, covering 61,648 school holiday events across more than 11,000 school districts. District-level coverage ensures accurate signal at the location level for customers in retail, restaurants, accommodation, and other sectors where school holidays drive demand.
The Forecasts API is now generally available. It delivers event-driven demand forecasts directly via API, without requiring customers to build and maintain their own forecasting models. Key capabilities:
Create and train forecast models against your own historical demand data
Retrieve daily-level forecasts enriched with PHQ event features
Access phq_explainability outputs to understand which events are driving forecast changes
Compare against a baseline to measure the MAPE improvement attributable to PredictHQ data
See the Forecasts API documentation to get started.
Saved Locations now support GeoJSON polygon geometries (Polygon, MultiPolygon, LineString, MultiLineString) in addition to point-and-radius. Custom polygon areas propagate through Events API, Features API, Beam, and Forecasts API. The external_id field has been added (replacing the deprecated location_code). See the Saved Locations docs for the updated schema.
PredictHQ now includes predicted school holiday events for future dates where official school calendars have not yet been published. 87,925 predicted school holiday events have been added covering future dates through early 2029. These events carry state: predicted and are updated automatically when official schedules are confirmed, extending the school holiday coverage horizon for forecasting models.
Academic events for the 2025–2026 calendar year have been added to PredictHQ's dataset, covering 10,826 events across 958 US institutions. This extends coverage for customers in accommodation, restaurants, retail, and transportation where university and college schedules are a material demand driver.
The attendance model used to calculate PHQ Rank and attendance signals for sports events has been fully retrained. The previous model dated from 2019. The retrained model covers Basketball, Ice Hockey, Baseball, American Football, Rugby, Soccer, and Volleyball, and delivers 16–58% reduction in mean absolute error across sport types. Approximately 950,000 events from 2022 onwards have been republished with updated attendance and rank values.
Events with a confirmed date but unconfirmed time (status: predicted) are now included in Features API results by default. Previously these events were excluded, which resulted in incomplete feature coverage - particularly for events announced well in advance. A new predicted_events.exclude parameter is available if you want to opt out. Approximately 72,000 events are affected.
You can now filter Events API results by beam.analysis_id or beam.group_id. The API automatically applies the location, radius, rank, and category parameters from the specified analysis — removing the need to reconstruct these manually. The Python SDK and tech docs have been updated.
Local Rank is now calculated and applied to every public event in the platform. Previously, Local Rank was missing for some event types. Events where population-adjusted ranking is not meaningful now use PHQ Rank as their Local Rank value. Historical events have been backfilled. This ensures consistent Local Rank availability across all API queries and Beam analyses.
Several usability improvements to Beam: rank and radius fields are now optional when creating an analysis (the API defaults to Suggested Radius and appropriate rank settings for the selected industry). Readiness checks now validate whether best-practice settings for industry, rank, and radius are in use. Analysis search has been improved for analyses with underscore-separated names.
Beam analyses now support Local Rank as the default rank type for new analyses. The rank.* field is now optional - Beam defaults to appropriate rank settings per industry. Suggested Radius has been updated to support all product industries. A toggle between Local Rank and PHQ Rank is available in the WebApp.
The Python SDK now covers the full Beam API surface: analyses CRUD, demand data upload, analysis refresh, events search, correlation, feature importance, value quantification, analysis groups, and dataframe utilities. The beam.* feature family is also available via the Features API SDK module.
Last updated
Was this helpful?
Was this helpful?