Unified Package Forecast (UPF) Architecture

Note: Specific program names, internal tool references, and other potentially sensitive details have been generalized in this summary for confidentiality.

Introduction: Framework for Volume Forecasting

The Unified Package Forecast (UPF) architecture enhances our forecasting capabilities by generating comprehensive volume forecasts for our delivery network. It implements a hybrid system that combines top-down network planning with granular, bottom-up station-level metrics. This approach aligns with operational processes, starting with package-level guidance and incorporating factors like territory coverage and delivery partner strategies. The result is a more precise, explainable, and adaptable forecasting system that improves operational planning.

Addressing Past Limitations

Previous forecasting models tracked site-level metrics but used a rigid, top-down linear structure. This created misalignment with core business planning processes, particularly regarding transportation capacity plans. Post-processing steps often obscured the forecast's logic, making it difficult to trace deviations to their source. Inconsistencies in definitions and metrics hindered communication across the supply chain ecosystem. UPF was designed with adaptability and modularity to enable continuous improvement and clearer insights.

The UPF Solution: A Hybrid, Modular Approach

UPF integrates top-down guidance with bottom-up data through a structured process. The final Total Volume Attainable (TVA) forecast is influenced by four key components:

1. The Network Signal: Derived from upstream planning systems and supplemented by macro-level share assumptions.

2. Granular Geo-unit Metrics: Detailed calculations at the zip code level for metrics like FTG (Fall-to-Ground), GAV (Gross Addressable Volume), CAV (Core Addressable Volume), attainment percentages, and program/cycle splits.

3. Jurisdiction rules: Define how zip codes are mapped to specific stations based on program-specific delivery plans.

4. Integration layers: Include Alignment, Constraint, and Scenario planning, which reconcile different signals and apply business rules to shape the final outputs.

This structure ensures both high-level strategy and ground-level performance inform the forecasts.

Key Design Principles

The UPF architecture is built on core principles that ensure its effectiveness:

1. Package-Level/EAD-Based Guidance: Starts with package-level guidance based on Estimated Arrival Date, aligning directly with upstream planning outputs and reducing discrepancies from intermediate conversion assumptions.

2. Territory and Carrier Integration: Dynamically adapts to changes in geographic coverage by different delivery programs and evolving delivery partner strategies. Accounts for expansion of first-party (1P) delivery services and manages potential overlaps in service areas.

3. Unified Program Modeling: Provides consistent forecasting across all major delivery programs and supports planning across multiple jurisdiction layers while maintaining granular geographic visibility.

4. End-to-End Explainability: Allows forecast deviations to be attributed and analyzed at the zip-code level by delivery profile, improving maintainability and enabling targeted improvements.

Technical Architecture Highlights

Supporting the business logic is a technical foundation designed for flexibility:

1. Modular and API-Based Design: Employs a modular architecture where different forecasting components are decoupled. This allows independent development and deployment of models for specific metrics without disrupting the entire system.

2. Metadata Management: All inputs, assumptions, and model versions are logged and versioned. This tracking provides transparency and aids in troubleshooting and analysis.

3. Cloud-Based Implementation: Built using AWS and Python, offering better control over calculations, improved visibility into metrics, and built-in validation compared to traditional ETL systems. This platform supports analytics dashboards and automation.

How UPF Works: The Forecasting Process

The UPF process transforms high-level plans into station forecasts through these steps:

1. Network Signal (Top-Down): Starts with total network volume and target volume allocated to internal delivery, sourced from upstream planning systems. Macro share assumptions dissect these figures by geographic coverage. By applying assumptions based on historical data and trends, a geographically dimensioned network plan provides targets for the bottom-up process.

2. Geo-Unit Metrics (Bottom-Up): Focuses on zip-level metrics calculated from an Actuals Master Dataset. The process follows a waterfall approach:

   • FTG (Fall-to-ground): Total volume share landing in each zip code
   • GAV (Gross Addressable Volume): FTG minus volume handled by non-addressable internal programs
   • Specialized program volume: Calculated by applying specific share to GAV
   • CAV (Core Addressable Volume): GAV minus specialized program volume
   • Core TVA: CAV multiplied by 'eligible attainment' percentage based on historical win rates adjusted to include volume lost due to controllable constraints
   • Total forecasted volume: Sum of specialized program volume and Core TVA

3. Jurisdiction and Aggregation: Zip-level metrics are mapped to delivery stations using program-specific jurisdiction plans. Specialized program TVA is aggregated directly to station level. For the standard program, zip-level CAV is aggregated to the station, then a station-level attainment percentage is applied to derive the Station Core TVA. Total station forecast combines standard and specialized program forecasts.

4. Cycle/Window Splitting: Daily station forecasts are broken down into operational cycles/windows using dedicated logic for cycle launches, additional cycles, and capacity utilization targets, balancing historical ratios with operational requirements.

5. Alignment and Integration: Ensures bottom-up forecasts reconcile with top-down targets through a sequential process. Intermediate bottom-up outputs are compared against corresponding top-down inputs, and adjustments are applied stepwise: jurisdiction volume alignment, GAV alignment, CAV alignment, and standard program adjustments. This method provides visibility into misalignments and enables targeted adjustments.

Adoption as Standard Architecture

Following successful development, the UPF model was adopted for the next generation forecasting product. Cross-team alignment between business, product, software, business intelligence, and research science functions supported this transition. The UPF architecture became the standard framework for the organization's forecast model, ensuring a unified approach to volume planning.