Automating Financial Forecasting

Modular Forecasting Approach

Each line item in the profit and loss-statement represents a distinct forecasting challenge, as data structures, dynamics, and influencing factors differ significantly by metric. For this reason, our [at] experts chose a modular approach where customized statistical model would be fitted to individual P&L line items. These models’ outputs would then aggregate into a complete, end-to-end forecast. This approach made it possible to coordinate the work of multiple independent teams, and to reflect domain-specific characteristics while maintaining a consistent overall logic across the forecast.

Central Forecasting Engine

To avoid time loss for duplicating the same forecasting logic across multiple line items, and to ensure that changes and enhancements would be applied consistently, a centralized technical forecasting architecture was established. This architecture included an automated input data processing pipeline, model training in Kubeflow, post-processing in Databricks, and result visualization in Power BI dashboard. The dashboard provided insight on the long-term accuracy of models, and benchmarked them against the previous manual forecast, providing business teams and decision-makers with a reliable basis for good business decisions.