Conformal time series decomposition with component-wise exchangeability

Derck W. E. Prinzhorn, Thijmen Nijdam, Putri A. van der Linden, Alexander Timans

Abstract: Conformal prediction offers a practical framework for distribution-free uncertainty quantification, providing finite-sample coverage guarantees under relatively mild assumptions on data exchangeability. However, these assumptions cease to hold for time series due to their temporally correlated nature. In this work, we present a novel use of conformal prediction for time series forecasting that incorporates time series decomposition. This approach allows us to model different temporal components individually. By applying specific conformal algorithms to each component and then merging the obtained prediction intervals, we customize our methods to account for the different exchangeability regimes underlying each component. Our decomposition-based approach is thoroughly discussed and empirically evaluated on synthetic and real-world data. We find that the method provides promising results on well-structured time series, but can be limited by factors such as the decomposition step for more complex data.

Introduction

Conformal Prediction (CP) has emerged as a premier framework for generating prediction intervals with rigorous mathematical guarantees. In its standard form, CP ensures that a prediction interval will contain the true value with a user-specified probability (e.g., 90%). However, these guarantees typically rely on the assumption of exchangeability—the property that the joint distribution of data points is invariant under permutation.

Time series data inherently violates this assumption. Sequential correlations mean that past observations heavily influence future ones, creating a non-exchangeable environment where standard CP methods often fail to provide valid or efficient intervals. While "non-exchangeable" CP variants have been developed to weight recent data more heavily, they generally treat the entire signal as a single block. This paper, presented at the 2024 COPA conference, proposes a "divide and conquer" strategy: decomposing the time series into distinct components and applying tailored CP methods to each based on their specific statistical properties.

The Conformal Time Series Decomposition Framework

The proposed framework operates on the principle that while a raw time series is non-exchangeable, its constituent parts may follow different exchangeability regimes. The pipeline consists of three primary stages: decomposition, component-wise conformal prediction, and recomposition.

1. Decomposition

Using Seasonal-Trend decomposition based on LOESS (STL), the input signal is split into three additive components:

  • Trend ($T_t$): The long-term progression of the series.
  • Seasonality ($S_t$): The periodic fluctuations (e.g., daily or weekly cycles).
  • Remainder ($R_t$): The residual noise left after removing trend and seasonality.

2. Component-Wise Conformal Prediction

This is the core innovation of the research. The authors assign a specific "exchangeability regime" to each component and apply a corresponding CP algorithm:

  • Remainder (Global Exchangeability): In a well-specified model, the remainder should ideally behave like independent and identically distributed (i.i.d.) noise. Because this component is approximately exchangeable, the authors apply CV+, a robust cross-conformal method that performs well under these conditions.
  • Trend (Lack of Exchangeability): Trends are characterized by high correlation and potential drift. To handle this, the authors utilize methods like EnbPI (Ensemble Batch Prediction Intervals) or ACI (Adaptive Conformal Inference), which are specifically designed for non-exchangeable, evolving data.
  • Seasonality (Local Exchangeability): The authors introduce the concept of "local exchangeability" for seasonal data. While a seasonal signal is not globally exchangeable, a specific point in a cycle (e.g., 8:00 AM on a Monday) is likely exchangeable with the same point in previous cycles. To exploit this, they propose new weighting schemes:
  • BinaryPoint: Only considers the same phase in previous periods.
  • BinaryLocal: Considers the same phase and its immediate neighbors.
  • ExpLocal: Applies an exponential decay weight based on temporal distance within the same phase.

3. Recomposition

Once prediction intervals are generated for each component, they are aggregated (summed) to create a final, comprehensive prediction interval for the original time series.

Experimental Evaluation

The authors evaluated the framework across a variety of datasets with differing levels of structure:

  • Synthetic Data: A sine wave with a linear trend and Gaussian noise.
  • San Diego Energy Consumption: Hourly electricity usage exhibiting high seasonality.
  • Rossmann Store Sales: Daily sales data with complex, multi-layered seasonality.
  • Beijing Air Quality: Hourly PM2.5 readings, which are highly irregular and noisy.

Key Findings

  • Efficiency Gains: On synthetic and highly structured data (like Energy Consumption), the BinaryPoint method produced significantly narrower (more efficient) intervals than standard baselines while maintaining the target coverage.
  • Conservative Guarantees: On real-world datasets, the method often proved "over-conservative," yielding coverage higher than the requested 90% (sometimes reaching 98%). While this ensures safety, it suggests the intervals could be further optimized for tightness.
  • Decomposition Sensitivity: The framework's success is heavily dependent on the quality of the initial decomposition. In the case of the Beijing Air Quality data, the highly irregular nature of the signal made it difficult for STL to extract a clean seasonal pattern, leading to more modest performance gains.

Discussion and Future Work

The research demonstrates that viewing exchangeability as a spectrum—rather than a binary property—allows for more nuanced uncertainty quantification. By isolating the "Remainder" component, practitioners can apply powerful standard CP tools to the noise while using specialized adaptive tools for the trend.

However, several areas for improvement remain. The current approach uses a simple additive sum for recomposition; future iterations could explore multiplicative strategies for series where variance scales with the signal. Additionally, the framework currently focuses on a single seasonal period. Extending this to handle nested cycles (e.g., simultaneous daily and weekly patterns) and tuning the "conservatism" of the aggregated intervals are promising avenues for future research.

Conclusion

Conformal Time Series Decomposition provides a modular, model-agnostic framework for providing uncertainty guarantees in forecasting. By acknowledging the different statistical regimes within a single time series, the method offers a path toward more reliable and efficient prediction intervals in complex, real-world applications.

Summary updated

Abstract

Conformal prediction offers a practical framework for distribution-free uncertainty quantification, providing finite-sample coverage guarantees under relatively mild assumptions on data exchangeability. However, these assumptions cease to hold for time series due to their temporally correlated nature. In this work, we present a novel use of conformal prediction for time series forecasting that incorporates time series decomposition. This approach allows us to model different temporal components individually. By applying specific conformal algorithms to each component and then merging the obtained prediction intervals, we customize our methods to account for the different exchangeability regimes underlying each component. Our decomposition-based approach is thoroughly discussed and empirically evaluated on synthetic and real-world data. We find that the method provides promising results on well-structured time series, but can be limited by factors such as the decomposition step for more complex data.
BibTeX
@inproceedings{prinzhorn2024conformal,
  author = {Derck W. E. Prinzhorn and Thijmen Nijdam and Putri A. van der Linden and Alexander Timans},
  title = {{Conformal time series decomposition with component-wise exchangeability}},
  booktitle = {COPA},
  year = {2024},
  url = {https://openreview.net/forum?id=yJfM2k7hQp},
}
Figure