Data Science capstone project batch #37

One Forecasting Problem, Three Approaches:
Predicting Transaction Volumes Across a Client Portfolio

How three forecasting methodologies tackled the same heterogeneous portfolio and arrived at remarkably similar results.

Forecasting transaction volumes across thousands of clients presents a fundamental challenge: not all time series behave the same way.

Some clients generate large and relatively predictable volumes, while others show strong seasonality, high volatility, or sparse activity. A single forecasting model applied across the entire portfolio risks either over-engineering the long tail or under-serving the clients that contribute most to total volume.

This project addressed that challenge through three independent forecasting methodologies. Each was developed end-to-end using the same cleaned dataset, covering approximately 5,500 participants, 29 months of historical data, and a 31-month forecasting horizon. The objective was to produce reliable monthly participant-level forecasts while accounting for the portfolio's highly heterogeneous behavior.

Three methodologies, one forecasting problem

The first methodology combined invoice-volume segmentation with behavioral segmentation based on seasonality. High- and medium-volume clients were grouped according to their temporal characteristics, while the main forecasting pipeline combined Prophet and XGBoost.

Prophet captured the broader temporal structure of each segment, including trend and seasonality, while XGBoost modelled participant-level differences. The resulting forecasts were reconciled to ensure that individual predictions remained consistent with segment-level totals.

The second methodology also began with invoice-volume segmentation but used activity volatility as the second segmentation dimension. It combined autoregressive modelling through SARIMA with nonlinear XGBoost-based approaches, focusing on temporal dependence and differences in participant activity.

The third methodology relied primarily on invoice-volume segmentation and used a horizon-adaptive weighted blend of Prophet, ETS, and Theta. Instead of selecting one model, this approach allowed the contribution of different forecasting methods to vary across the forecasting horizon.

For the remaining ~10% of portfolio volume, all three methodologies used a seasonal-naive approach. This reflected a shared conclusion: sophisticated models are not always justified for sparse, low-impact time series.

FIGURE 1: Portfolio concentration: share of participants vs. share of transaction volume

Different strategies, similar results

Despite their different modelling philosophies, all three approaches produced portfolio-level forecasts within a relatively narrow performance range. WAPE values varied between approximately 9% and 13%, while the final forecasts were also similar in overall magnitude.

 

 

FIGURE 2: High-level comparison of the three forecasting methodologies

These results suggest that there was no single universally superior modelling approach. Instead, several well-designed strategies were able to capture the portfolio's underlying structure and arrive at similar aggregate conclusions.

One particularly important lesson was that segmentation can be as important as model selection. The first question is not always “Which algorithm should be used?” but rather “Which clients should be treated similarly?”

The project also demonstrated that complexity should follow business importance. The clients responsible for most transaction volume justified more sophisticated modelling, while simpler methods were sufficient for the long tail.

 

 

FIGURE 3: Final portfolio-level forecasts from the three methodologies

Final takeaway

This project showed that forecasting heterogeneous client portfolios is not simply about finding the “best” algorithm.

Three different approaches – combining segmentation with hybrid modelling, autoregression and nonlinear learning, or horizon-adaptive forecasting ensembles – achieved WAPE values of approximately 9–13% and produced similar portfolio-level forecasts.

The key takeaway is therefore simple:

Understanding the structure of the portfolio may matter just as much as choosing the forecasting model itself.

For real-world forecasting problems, the most effective approach may not start with “Which model is best?”, but with a more fundamental question: “What makes these time series different, and which differences actually matter?”