Interpretable AI: How InXiteOut Built a Trustworthy, High-Accuracy OpEx Forecasting Solution 

Client Context  

Our client, a global pharma major, generated 24-month look-ahead forecasts for 15,000+ Operational Expenditure (OpEx) time series across ~500 cost centers in multiple countries and currencies. 

The entire forecasting process was done manually, making it laborious, time-consuming, error-prone, and non-standardized. 

The finance team partnered with InXiteOut to implement a machine-learning solution that could generate these forecasts automatically with a high degree of accuracy.  

The InXiteOut Approach 

Our objective was to build not just an accurate forecasting engine, but an interpretable one that the client's finance team could understand, trust, and use. 

Baseline and assessment 

First, we collaborated with the finance team to understand their existing manual process and the key drivers of OpEx. We mapped all 15,000+ time series and established baseline accuracy by testing standard forecasting models (e.g., SARIMA, ETS) against the historical manual forecasts. 

Regression modelling, interpretability limitation  

Our financial analytics experts first focused on wages, which corresponded to the most significant portion of operational expenditure. Our initial model (a standard Multiple Linear Regression) proved accurate at the aggregated level, meeting the project's primary validation metric. 

However, the model lacked interpretability. When the finance team examined the granular, sub-category forecasts to understand the rationale, they found the model's logic was unexplainable and often violated fundamental business rules. For example: 

• A pre-defined order in average wages across different workgroups (e.g., the manager cohort must earn more than the associate cohort). 
• Minimum gaps between average wages. 
• Lower and upper cost thresholds for each wage group. 

Standard regression methods cannot incorporate these real-world constraints, rendering the granular forecasts untrustworthy and unusable.  

"Constrained AI" solution 

To solve this, we developed a Constrained Regression Model — a custom AI we could train with the client's specific business logic. 

We converted their real-world business dynamics (like the wage gaps and thresholds) into a set of mathematical Inequality Constraints. The model was then formulated as an optimization problem: its goal was to find the most accurate forecast, subject to the non-negotiable business rules provided. 

This new model was just as accurate as the first regression model, but also interpretable. Its forecasts consistently follow the client's real-world business logic, making them explainable, trustworthy, and ready for operational use. 

Read more about our Constrained Regression model in our blog.  

Productionization and integration 

Based on the accuracy and interpretability achieved, the "Constrained AI" solution was adopted for integration into the client's business processes. 

We built an end-to-end automated pipeline that integrated with upstream finance systems for automated data extraction and with downstream systems to push the final forecast outputs. We also developed a suite of Power BI dashboards for ongoing performance tracking, enabling the finance team to monitor forecast accuracy in real time. 

This fully automated the OpEx forecasting process, establishing a scalable, accurate, and interpretable system for the client's financial planning. 

Technology stack used 

• Azure ETL Ecosystem (Data Lake, Data Factory) 
• Dataiku     
• Power BI  

Benefits Delivered  

The automated process delivered transformative results for the client's finance team: 

• 80%+ Reduction in Manual Effort: The automated forecasting process significantly eliminated the laborious, time-consuming manual work, freeing the finance team for value-added analysis. 
• ~10% Reduction in Forecast Error: The new AI model's forecasts were 10% more accurate than the previous manual forecasts. 
• ~1% Aggregate Percentage Error: The solution achieved a high precision aggregated error of ~1% across all 15,000+ time series.