Introduction
Forecasting revenue is relatively simple when dealing with standard products where customers place an order, receive the item immediately, and payment is recorded right away. But things become far more complex when each item is built from scratch, tailored specifically to individual needs after a series of service steps.
This is especially true in sectors like healthcare, where solutions are often crafted uniquely for each person. The process involves several stages such as assessment, design, approvals, adjustments, and final delivery. Revenue can only be recognized at the very end of this journey, once the product is fully delivered.
These custom workflows often take several weeks to complete, and at any point in time, many such orders are actively moving through different phases. Predicting revenue in this environment is challenging, particularly when information is missing or delayed.
In this blog, we walk through how we addressed this challenge using a step-by-step forecasting system powered by machine learning. We developed a model that predicts when payments are likely to occur and how much revenue can be expected each month—even when some parts of the process aren’t fully Captured.
You’ll learn about our data preparation approach, how we combined statistical methods with machine learning, and how we set up a feedback mechanism to update our model when its predictions start to shift ensuring greater accuracy, transparency, and financial confidence in highly customized service environments.
Business Objective
Our objective is to forecast revenue for the next 30 days, starting from any given day, based on the current snapshot of open orders – even if those orders haven’t yet reached the invoicing stage.
This is not a one-time forecast. The model is designed to predict daily, enabling a rolling 30-day outlook.
For example, if we run the model today, it forecasts revenue for the next 30 days from today. If we run it again tomorrow, it shifts to forecast the next 30 days from tomorrow providing dynamic, up-to-date projections based on the latest open orders data.
Key Use Cases:
- Cash Flow Planning: Anticipate incoming revenue with high granularity.
- Resource Allocation: Align staffing and inventory with expected demand.
- Predictive Reporting to Leadership: Provide proactive, data-driven financial insights.
Challenge: Survival Models Fall Short Due to Workflow Gaps
We Used survival analysis and transition regression models to estimate the time between key stages. While theoretically ideal for modelling sequential workflows, these models failed in our case primarily due to following reasons.
- Critical Data Gaps:
The intermediate transitions showed an unusually large time gap.
- Missing Intermediate Stages:
In practice, stages like approvals and administrative checks occur between Evaluation and Service.
However, these were either not recorded or captured inconsistently, leaving significant blind spots.
- Unexplainable Gaps:
The large delays could not be explained using available features, leading to unstable and uninterpretable model behaviour.
Acuvate recognized the limitations of forcing predictive models on incomplete data. This led us to pivot toward a statistical and ML hybrid approach that leveraged patterns in the data we could trust without relying on missing or inconsistent transitions.
Solution: Hybrid Statistical + ML Approach for Robust Forecasting
We pivoted to a pragmatic, data-efficient approach that worked with the available information.
Model 1: Statistical Forecasting (Baseline)
We designed a statistical method to estimate future revenue based on historical patterns of order conversion and payment timelines. This approach involved calculating the average revenue conversion rate from open orders over similar time windows in the past.
This approach was simple and interpretable, but it needed refinement — especially when business volume fluctuated or historical trends didn’t align with current behaviour.
Model 2: ML Compensation Using Time-Based Patterns
To improve the reliability of our statistical forecasts, we developed a compensation factor using machine learning. This factor was trained on historical data to learn the typical gaps between statistical predictions and actual revenue outcomes. By multiplying the statistical forecast with this compensation factor, we produced a final revenue prediction that more closely aligns with actual performance, effectively correcting for consistent biases or underestimations in the base statistical model.
Final Output:
- Final Revenue Forecast = Statistical Revenue × Predicted Compensation Factor.
- This approach significantly reduced error % and closely matched actual revenue.
End-to-End Workflow for Revenue Forecasting with MLOps Control
To ensure accurate, maintainable, and production-ready revenue forecasts, we implemented a fully automated and traceable machine learning workflow.
- Raw Data Ingestion
All relevant historical and current data is ingested from transactional databases. - Exploratory Data Analysis (EDA)
The data is profiled to identify patterns, outliers, feature relevance, and missing values. - Model Training Pipeline
Using the cleaned and curated dataset, we train a two-part hybrid model:- A statistical baseline model that estimates revenue using historical patterns.
- A machine learning compensation model that learns the difference between the statistical prediction and actual outcomes.
- Model Registry Pipeline
Once training is complete, the best-performing model version is registered in the model registry. - Deployment Pipeline
The registered model is automatically deployed to an inference endpoint using Azure Machine Learning. - Live Inference from Endpoint
Each day, the latest snapshot of open orders is passed to the inference endpoint. The model generates a rolling 30-day revenue forecast, considering current order status and expected payment timelines. - Prediction Storage in Azure Data Lake
The forecasted revenue outputs are saved to Azure Data Lake in a structured format. - Business Reporting via Power BI
Power BI dashboards are connected to Azure SQL or directly to the data lake to present forecast insights to stakeholders. - Monitoring and Drift Detection
Azure Monitoring tracks model performance over time, including forecast accuracy and feature drift. If errors exceed predefined thresholds or data patterns shift significantly, alerts are raised to trigger retraining. - Model Retraining and Governance
When drift is detected, the training pipeline is re-executed using the most recent data. The new model is validated, registered, and redeployed – ensuring the forecasting system evolves alongside business growth and changing workflows.
This unified, end-to-end pipeline ensures that revenue forecasting is not only accurate but also maintainable and transparent. The combination of automation, monitoring, and human oversight makes it suitable for high-stakes environments like healthcare, where workflows are complex and real-time financial visibility is essential.
Visualizations
Plot 1: Daily Revenue - Actual vs Predicted
As seen in the plot, the statistical predicted revenue (dotted blue) initially deviates significantly from the actual revenue (orange). However, after applying the ML-based compensation, the ML corrected revenue (dotted green) closely aligns with the actuals demonstrating the effectiveness of our hybrid approach.
Plot 2: Error Distribution
All error values lie within ±3 standard deviations from the mean, indicating a consistent and controlled prediction error spread with no statistical outliers. This suggests that the model’s performance is stable and does not exhibit high variance.
These visuals were key in building stakeholder confidence.
Challenges & Learnings
Challenges
- Sparse or missing data in intermediate stages
- Larger time gaps in data
- Unstructured revenue behaviour by geography
- Orders with non-linear stage paths
Key Learnings
- Classic survival models break under sparse transitions
- A domain-driven statistical base + ML compensation is highly effective
Conclusion
Forecasting revenue isn’t just about predicting numbers, it’s about bridging the gap between finance, operations, and data science.
Our journey began with traditional survival models, which while theoretically sound but struggled with the real-world messiness of incomplete data and non-linear workflows. Rather than forcing a solution, we embraced the complexity. We pivoted to a hybrid approach that blends statistical intuition with machine learning adaptability.
The result is a forecasting system that is not only accurate but also transparent and resilient. Even when critical workflow stages are missing, the model continues to perform with confidence and clarity.
With MLOps automation now embedded, this system isn’t static—it learns and adapts as the business grows, as workflows evolve, and as market conditions shift. From dashboards that provide real-time insights to automated retraining triggered by drift, we’ve built a solution that is as dynamic as the business itself.
At Acuvate, this isn’t just a model, it’s a strategic capability. One that brings foresight into decision-making and ensures that healthcare organizations are not just reacting to the future but actively shaping it.
