Skip to main content

FOR-SV430-XPFv4.3.0-PV930 Release Notes

Author: Matthew Lindon, Created: 2026-07-31

Release Date​

August 3, 2026

Release Overview​

‌The theme of this release is simplifying end-user adoption, reducing use case maintenance complexity, and decreasing the complexity to create new use cases.

Forecast Agent (Beta)​

  • Ask questions about SensibleAI Forecast projects in plain English. Get grounded answers about forecast accuracy, key drivers, model explanations, and job status - no dashboard navigation required. Example questions:

    • Why is our forecast for next quarter so high for North America?
    • What features are driving the forecast for Consumer Electronics in April?
    • What's the growth rate for my region compared to last cycle?

  • Access the Forecast Agent via one of two locations:

    • Click on the AI icon in the top right corner of the environment

      image-20260729-142349.png

    • Click on the “Agent Mode” section from the home page of the FOR solution

Monitor System​

  • The Monitor system has received a large revamp and is now a generally available feature. Use the monitor system to gain control over which forecast model is used for each target in your project. After a prediction run produces forecasts from multiple competing models per target, the Monitor system uses user-define rules to determine the Effective Model - the single model whose forecast flows into your consumption groups, dashboards, and exports.
  • For more information on capabilities and usage instructions, please see Monitor System - Rule Based Model Selection & Alerting

Source Data Features Revamp​

  • New Unified Data Management Dashboard — replaced legacy interface with a modern web dashboard page organized as three connected panels: Data Sources (left), Generators/Data Preview (middle), and Features grid (bottom), streamlining the entire feature data pipeline configuration in one place.
  • Enhanced Datasource Editing — supports bulk editing across datasources with guided create, edit, and delete workflows. Users can configure data connection properties, column mapping, and aggregation settings.
  • Assignable Source Features — source data features are now assignable, allowing users to target specific source features for assignment alongside standard features.
  • Feature Settings Bulk Edit Controls — simplified UI to apply settings across multiple features at once. Supports per-feature overrides for Known In Advance, Data Cleansing Method, Aggregation Method, Hierarchy Groupings, Frequency Override, and more.
  • Forecasting Method Configuration — added explicit Forecasting Method selector (None, Lag, Statistical, ML, Smart) at both the datasource and import levels, controlling how feature data extends to the forecast horizon.
  • Inline Data Preview & Validation — built-in Data Preview tab shows the first 100 rows of raw data from a selected source with connection metadata and detected frequency.
  • Auto-Refresh on Background Jobs — SDF and Feature Package Manager grids automatically refresh when feature data load jobs complete, eliminating manual page refreshes.

Cube Data Integration​

  • Enables OneStream native Cubes to be used as target data sources and supports Consumption Group integration for exporting project results directly back to Cubes.

Forecast Analysis​

  • New Analytics page added under Utilization > Analysis, providing visibility into forecast performance metrics and model analytics.
  • Embedded web analytics dashboard parameterized by project ID and build info, giving each forecast its own analytics context.
  • Consolidated analytics hub accessible from the main navigation for monitoring model performance and accuracy metrics.

Feature Forecasting

  • Optional Feature Forecasting capability enables future feature values to be automatically estimated when source data does not extend through the forecast horizon, helping preserve predictive signals for downstream forecasts within SensibleAI Forecast. Additional details are available in the Feature Forecasting documentation.
  • Forecasting Method Configuration introduces configurable forecasting methods (None, Lag, Statistical, Machine Learning, and Smart) at the generator level, allowing users to control how future feature values are extended based on the characteristics of each feature.
  • Expanded Feature Generator Support — Feature Forecasting is available across Feature Package Manager, Source Data Features, and Feature Library configuration workflows, providing a consistent forecasting experience regardless of how feature generators are created or managed.
  • Intelligent Method Selection — adds a Smart forecasting option that automatically applies the system-recommended forecasting approach for each supported generator type, simplifying configuration while maintaining flexibility for advanced users.
  • Prediction Horizon Support — automatically extends feature values only when future data is unavailable, ensuring forecasting models continue receiving aligned feature inputs throughout the prediction window without requiring manual data preparation.
  • Improved Explainability Integration — Feature Forecasting is fully integrated with Retro Explanations, allowing users to better understand when differences between forecasted and actual feature values contributed to forecast error.

Assign Page​

  • Safety confirmation dialog when users attempt to override all target assignments with empty selections — warns of the destructive action before proceeding.
  • Override and Append buttons now activate only when specific rows are selected in the grid, preventing accidental bulk operations on unintended targets.
  • Custom Target wizard now respects grid selections, ensuring operations apply only to explicitly selected rows.
  • Inline warning messages in the wizard preview step when a destructive override with empty selections is detected.
  • Dual-scope validation prevents proceeding with assignments unless either Features or Locations is explicitly selected in the assignment scope.

JSON Enabled Routine Implementation​

  • Introduces a global setting to designate authorized user groups that can implement routines through JSON-based configurations.
  • Includes Target Data Source Validation, Simulator, Consumption Groups, Data Source Updates/Refresh.
  • Adds submission documentation outlining required JSON structure, formatting standards, and implementation guidelines.

Hyperscale Storage​

  • Introduces a new storage strategy that supports increased project creation by allowing users to move projects online or offline, with offline projects remaining inaccessible until restored online.

Hierarchical Column Sets​

  • Enables configuration of relationships across target data dimensions, with support for applying the setting to both new and existing projects through Global Settings.

AI Project Context​

  • Adds a new project-level field available during project creation and updates, enabling users to provide contextual information that enhances AI-driven project analysis, recommendations, and agentic development workflows.

AI Activity Center​

  • Provides a centralized hub for monitoring AI project activity, making it easy to review running tasks, track project progress, and gain visibility into ongoing AI operations across the platform.

Retro Explanations​

This is a new tab in Forecast → Analysis page

  • Allows users to run “Retrospective” predictions if data used in a previous prediction has changed (i.e., source feature data has changed, historical actuals have been updated, or forecasted features now have actual values). This allows users to answer the question: “If we had all of the data we have now when we initially ran this prediction, how much more accurate would we have been?”

    • Changes in predicted values are explained with “Value Changes” for any feature with a different value being used in the retrospective prediction, and “Impact Changes” for features whose value didn’t change but whose impact on the predicted value did change in the retrospective prediction.

  • For every data point in a prediction for which there is now an actual value, the miss is calculated and error is attributed to each feature that was used to generate the forecasted value, allowing users to better understand why a forecasted value missed on any date.

    • “Attributable Error” is calculated for each feature by scaling the effect of the feature on the predicted value by the miss amount. This allows users to see all features the model used and whether they pushed the forecasted value toward or away from the actual value, and by how much, giving a better understanding of why a model missed on any given forecasted value.

    • “Unexplained Error” is any difference between the actual value and the confidence interval band. For example:

      • Actual Value = 500, Forecasted Value = 1000, 95% Confidence Band = 600–1400 → “Unexplained Error” = -100 (500 - 600)

  • AI-generated explanations are available on this page to help users better understand the data they are seeing in natural language.

  • These calculations depend on Confidence Intervals, so new projects will default to having a 95% confidence band for the top 3 models for each target. This is still configurable, if this functionality is desired on existing projects that don't have Confidence Intervals enabled please go through a manual rebuild and enable them.

  • This will only be available on predictions ran after performing the upgrade to 4.3.

  • Known issue with reconciled models not supporting retro explanations yet, to be included in next release.

Metric Scores on Forecast Analysis page​

  • The metric scores for each model on this page had previously come from Pipeline evaluations, now metrics are derived from predictions where there are actuals to compare to and if no actuals are available yet the table will say “Awaiting actuals”.

Bug Fixes​

Data Update Snapshot Local Time​

  • Instead of using UTC as the standard timing, snapshots are now converted to the users local time.

Explanation of Reconciliation Strategies in Hierarchical Forecasting​

  • Full descriptions of each reconciliation strategy right at the implementation source instead of searching through documentation.

Defaulted Confidence Intervals​

  • Confidence intervals will now be enabled at all times.

Dynamic Scaling of Summation Numbers​

  • Easier readability of large numbers throughout the solution.

Simulator in Utilization​

  • Users can now utilize simulator actions in utilization and not just prior to the pipeline.

Dependencies​

Xperiflow v4.3.0

XBR-SV230-XPFv4.3.0-PV930

Known Issues​

  • Filling out multiple drop-down boxes within Source Data Upload, Feature Data Upload, or Data Update dialog input forms too quickly will fill in with defaulted dimension values instead of persisting the previously selected values. Recommendation: do not click through as fast as possible.
  • Within the Pipeline Arena page or the Analysis Forecast page, when navigating from the Waterfall or Periodic Explanations views to the Tug Of War view will result in the Tug Of War plot only taking up half the screen. Recommendation: click a dropdown from the model arena table to fix the width of the tug of war visual.
  • In the windows application, within the Pipeline Arena page, Analysis Forecast page, Artifacts page on theRuns page of STU*,* or when inputting parameters for Routine Runs within STU, users may see tooltips over buttons that do not make sense.
  • Within the Pipeline Arena or the Analysis Forecast page, the PolyElasticNet and Cubist Model variants will generate Feature Impact results but will not generate Feature v. Actual plot data.
  • The solution does not block you from making project model build alterations while that project is being copied. Recommendation: it is advised that you avoid making alterations to the project while it is being copied.
  • When running model builds that leverage Hierarchical Forecasting with reconciliation, it is possible for targets disregards the Allow Negative Targets setting No and actually product forecasts with negative target values.
  • When projects are running pipelines or other costly jobs consuming more than 75% of server resources, page rendering speeds may increase by 25%.
  • On a project restart, previously configured events may be removed. Recommendation: revisit events and assignments after restarting a project.
  • Users are not blocked from running the restart job while a project copy is running. This is not recommended as the projects may end up being in different states. Recommendation: do not run a restart job (or any other job) while a project is being copied.
  • Direct export of SensibleAI Forecast DMF tables that utilize hierarchical forecasting will require manual intervention to populate unused dimension columns with real values.
  • When viewing large paginated tables of the error log within the Forecast Agent, the “next page” button may not be visible or usable. Please use the AI Activity Center to parse through more records as needed.
  • Backtest functionality has not been functional for several releases now, cannot enable backtest in a split configuration. Can be mitigated with manual backtests (i.e. cut training data off one forecast range early → predict and compare to actuals).
  • Reconciled models currently do not support retro explanations, will be addressed in a future release.

Upgrade Guide​

Install Guide - SensibleAI Forecast (FOR)

Compatibility Guide​

Xperiflow Compatibility Guide v4.3.0

Last updated on by Nic Napoleone

Was this page helpful?