Part 3 – Roadmap, Limitations, Considerations & Best Practices

 

🔗 Read the Full Series

• Part 1 – Configuration and Model Setup
• Part 2 – Algorithms and Use Cases
• Part 3 – Roadmap and Best Practices

Introduction

In the prior parts, we built your foundation and understanding of how Advanced Predictions works and how to interpret outputs. In Part 3, we zoom out—exploring where Oracle is headed, the architectural constraints, and how to operationalize Advanced Predictions in real organizations.

 

1. Roadmap: What’s Coming & What to Watch

Oracle’s disclosed roadmap includes:

Explainable Predictions / Feature Importance

• Visual ranking of driver impact
• Contribution charts / SHAP-style decomposition
• Ability for business users to see “which driver moved forecast how much”
• Enhances trust, auditability, and stakeholder buy-in

Intelligent Feature Engineering

• Lag detection: auto-detecting lagged relationships (e.g. promo in Jan influences Apr)
• Window aggregation features: rolling mean, median, variance
• Time-based features: month, quarter, day-of-week, etc.
• Automating these enrichments to accelerate modeling

Dynamic Parent-Level Prediction

• Predict at aggregated levels even when leaf-level data is incomplete
• Useful in organizations that forecast at roll-up levels (e.g. “All Products” or “Total Region”)
• Reduces manual data cleanup when leaf-level data is spotty

 

2. Architectural & Licensing Considerations

Here are some considerations:

• License: Enterprise EPM only
• Embedded ML: OCI Data Science is built in; no extra infra cost
• Enablement: Opt-in feature; must activate in application settings
• UX Theme: Only available in Redwood theme
• Cube types: Works in both BSO and ASO cubes
• App coverage: Available in Modules, Custom Applications, FreeForm, Sales Planning, Strategic Workforce Planning, Predictive Cash Forecasting

Also, note that certain operations like explainability or dynamic parent-level prediction may not yet be fully supported in Smart View.

 

3. Best Practices & Deployment Strategy

Data Strategy

• Maintain consistent period granularity (e.g. all monthly)
• Clean driver data and document handling logic
• Standardize dimension naming across historical and forecast slices for mappings

Modeling Strategy

• Always start with AutoMLx; only diverge when business justification exists
• Limit driver count early; bias toward parsimonious models
• Control lookback window; adding too much historical noise may hurt performance

Versioning & Governance

• Save multiple prediction versions (e.g. AutoMLx, manual, algorithm variants)
• Allow planner toggles on forms to compare output versions
• Use the export reports to document assumptions

Backtesting & Monitoring

• Use rolling-origin or k-fold time-series splits
• Monitor error drift over time
• Re-train or re-configure periodically (quarterly or annually)
• Create dashboards to compare predictions vs actuals over time

Scenario Design

• Write back Base / P10 / P90 slices for comparative planning
• Use event flags to differentiate scenarios (e.g., “with event vs without”)
• Encourage planners to overlay manual adjustments on top of ML predictions

Integration & Automation

• Schedule regular prediction runs (e.g. nightly, weekly)
• Use Groovy rules or scheduled jobs to refresh predictions
• Integrate predictive output into data flows and dashboards

 

4. Limitations & Cautions

• Lack of full explainability (feature importance is on roadmap)
• Risk of overfitting with many drivers and limited history
• Input driver data gaps degrade model quality
• Event mis-labeling can mislead forecasts
• Model latency for large-scale prediction jobs
• Some operations not available via Smart View
• The prediction job status message might be misleading sometimes, when that happens, a SR is suggested.

These limitations are real, but most can be mitigated with strong data discipline, iterative testing, and stakeholder alignment.

 

5. Embedding Into FP&A Process

To move from pilot to production requires:

• Pilot use case (e.g. high-dollar product forecast, workforce planning)
• Stakeholder sponsorship & review (show the initial performance uplift)
• Documentation & process handoff (driver register, event logic)
• Training planners to interpret CI bands and scenario outputs
• Governance schedules to refresh models, validate error drift
• Expand scope to other domains (e.g. cash, workforce, services)

Over time, you evolve from “ML experiment” to “driver-based forecasting core” in your FP&A engine.

 

Conclusion

Over this three-part series, you now have:

• Deep technical knowledge of how Oracle EPM’s Advanced Prediction works
• Step-by-step set-up and best practices
• Insights into internal modeling, interpretation, and roadmap
• Guidance on operationalizing the feature in real organizations

As Oracle continues to build out explainability and dynamic prediction at parent levels, your investment in this capability will pay dividends in forecast accuracy, stakeholder trust, and scalable driver-based insight.