The Challenges of Time Series Forecasting: Balancing Hierarchies, Features, and Explainability

During my master’s program, I made a conscious decision to steer clear of time series (TS) analysis. I had no strong inclination toward it, and I saw the domain as something I wanted to avoid. However, in my professional career, things took a turn when I was thrust into a situation where I had to implement time series forecasting for a major client. The sudden responsibility demanded a rapid learning curve, and I was required to grasp its nuances faster than I had anticipated.

The Traditional Approach: Univariate Forecasting

Traditional time series analysis primarily focuses on univariate forecasting. This is the simplest form of forecasting where historical data is used to predict future values of a single variable. While this approach is effective for straightforward problems, it doesn't hold up when clients demand more sophisticated insights.

For instance, one of the biggest challenges arises when the client starts asking the question, “Why did the model predict this particular value?” Unlike general machine learning models that are equipped with various tools for explainability, the basic univariate time series models do not offer clear business-related explanations. They are typically limited in their ability to identify and explain the factors influencing the predictions, which becomes a significant shortcoming in a business context.

Expanding the Model: Adding Features

To overcome this, you might try incorporating additional features, essentially transforming the time series problem into a multivariate forecasting problem. With the advancements in machine learning (ML), new approaches allow you to introduce external data/features to improve the predictive power and provide a more comprehensive view.

However, while these techniques may improve accuracy, they don’t fully address the problem of explainability. Even if you leverage techniques like SHAP to provide insights into how each feature influences the forecast, it can still fall short of satisfying the client’s needs. The client not only wants to know the “what” but also the “why” behind the predictions in a manner that directly ties to their business.

The Hierarchical Forecasting Challenge

A further complication arises when the client moves beyond simple time series forecasting and requests hierarchical forecasting. Hierarchical forecasting involves predicting data at different levels of aggregation (e.g., sales forecasts by region, product, or department), which adds a layer of complexity to the process.

In hierarchical forecasting, you can't easily incorporate external features, because the models are typically built to aggregate forecasts at different levels, based on the inherent structure of the data (e.g., regions rolling up to national sales). As a result, the sophisticated explainability methods that worked for multivariate forecasting no longer apply. Explainability is lost.

Managing Multiple Models for Hierarchical Forecasting

While hierarchical forecasting can be implemented by building individual models for each level or entity (e.g., region, product, etc.), this approach presents another significant challenge: model management.

In theory, you could create separate models for each component of the hierarchy to maintain accuracy and predictability. However, managing such a large number of models becomes increasingly complex. Every model would need to be trained, updated, and monitored individually, leading to scalability issues. The resources required to manage this model ecosystem—both in terms of computational power and human oversight—become overwhelming.

Moreover, with so many models, it becomes nearly impossible to ensure consistency and explainability across all levels. Each model might interpret features differently, resulting in potential misalignment in the forecasts at different hierarchical levels.

Merging Hierarchical Forecasting, Features, and Explainability

Now, the real challenge is reconciling these two approaches: how do you perform hierarchical forecasting with added features while maintaining a level of explainability that satisfies business requirements?

The problem stems from the fact that hierarchical and feature-based forecasting rely on fundamentally different approaches, and therefore the type of explainability required for each also varies. For hierarchical models, explainability is often limited to aggregate data interpretations, whereas feature-based models allow for a more granular explanation using techniques like SHAP or LIME.

Integrating these approaches without losing the ability to provide meaningful, business-relevant insights remains an unresolved challenge. How do we retain the benefits of external features while maintaining the structural integrity of hierarchical forecasting models? Moreover, how do we ensure that the model can provide a level of explainability that satisfies the client’s need for transparency at both the aggregate and granular levels?

Conclusion

These are some of the real-world challenges we face today in time series forecasting. Combining hierarchical forecasting with feature engineering, while ensuring model interpretability, is a complex task that goes beyond traditional methods. As the field of time series forecasting evolves, so too will the tools and methodologies required to meet these business demands, but for now, it remains a space filled with unresolved challenges and learning opportunities.