Conformal prediction bands for two-dimensional functional time series

Niccolò Ajroldi, Jacopo Diquigiovanni, Matteo Fontana, Simone VantiniComput. Stat. Data Anal 2023 · DOI

TL;DR

This paper introduces a novel method for forecasting time-evolving surfaces (two-dimensional functional time series) with reliable uncertainty quantification. By combining conformal prediction with functional autoregressive models, it provides prediction bands that offer guaranteed coverage, a crucial feature for applications where understanding prediction uncertainty is paramount.

What Problem Does This Paper Address?

Forecasting complex, time-varying data like surfaces (e.g., weather patterns, sea level) presents a significant challenge. While predicting the future value is important, accurately quantifying the uncertainty around that prediction is equally, if not more, critical for decision-making. Existing methods for uncertainty quantification often struggle with the high dimensionality and complex dependencies inherent in two-dimensional functional time series. This paper addresses this gap by extending conformal prediction, a robust framework for uncertainty quantification, to this challenging data setting, aiming to provide statistically sound prediction bands.

Methods

The core of this work lies in the fusion of two key methodologies:

  • Conformal Prediction (CP): This is a non-parametric framework that provides prediction sets with guaranteed coverage probabilities, regardless of the underlying data distribution or the predictive model used. The authors employ an inductive or split conformal approach, which is computationally efficient for high-dimensional data. They adapt existing CP techniques for functional time series to the two-dimensional case, defining a specific nonconformity measure tailored for surfaces. This measure quantifies how "unusual" a new observation is compared to the training data.
  • Functional Autoregressive (FAR) Models: To provide point predictions for the time-evolving surfaces, the authors extend the Functional Autoregressive (FAR) model of order one (FAR(1)) to the two-dimensional setting. This model captures temporal dependencies within the functional data. They propose and compare different estimation techniques for this 2D FAR(1) model, including methods based on functional principal components (FPCs) and a simpler concurrent model.

The prediction bands are constructed by combining the point predictions from the FAR(1) model with the uncertainty estimates derived from the conformal prediction framework.

Datasets

The paper uses two types of data:

  • Simulated Data: A simulated FAR(1) process is used to evaluate the performance of the proposed methods under controlled conditions. This allows for a direct comparison of different forecasting algorithms and the assessment of coverage and band size.
  • Black Sea Level Anomalies Dataset: This real-world dataset, collected daily by Copernicus Climate Change Service (C3S), consists of sea level anomalies over the Black Sea. This dataset is used to demonstrate the practical applicability of the proposed methodology.

Key Insights and Ideas

  • Conformal Prediction for Surfaces: The paper successfully extends conformal prediction, a powerful tool for uncertainty quantification, to the complex domain of two-dimensional functional time series. This opens up new possibilities for reliable forecasting of time-evolving surfaces.
  • Guaranteed Coverage: A key advantage of conformal prediction is its ability to provide prediction bands with guaranteed coverage, meaning that the true future surface will fall within the predicted band with a specified probability, regardless of the underlying data distribution. This is a significant improvement over methods that rely on strong distributional assumptions.
  • FAR(1) Extension: The adaptation of the FAR(1) model to two-dimensional functional data provides a suitable framework for capturing the temporal dynamics of surfaces, which is crucial for accurate point predictions.
  • Forecasting Algorithm Impact: The simulations highlight that while conformal prediction provides valid coverage even with suboptimal forecasting algorithms, the efficiency (i.e., the size of the prediction bands) is significantly improved by using accurate forecasting methods.
  • Data-Driven Modulation: The use of the functional standard deviation as a modulation function for the prediction bands allows for adaptive uncertainty quantification, resulting in narrower bands in regions of low variability and wider bands in regions of high variability.

Key Results

  • Empirical Coverage: The proposed conformal prediction bands consistently achieved empirical coverage close to the nominal confidence level across various simulation scenarios and for the real-world Black Sea dataset, even when using less-than-perfect forecasting algorithms.
  • Prediction Band Size: The size of the prediction bands generally decreased with increasing sample size and decreasing confidence level. Among the forecasting algorithms, those based on estimating the autoregressive operator (FAR(1)-EK and FAR(1)-EK+) produced the tightest prediction bands, comparable to an "oracle" predictor with perfect knowledge.
  • Real-World Application: The application to Black Sea level anomalies demonstrated the practical utility of the method, producing visually interpretable prediction bands that captured the observed surfaces.

Limitations and Future Work

  • Stationarity Assumption: The FAR(1) model, like many time series models, assumes stationarity. The paper acknowledges the lack of established stationarity tests for two-dimensional functional time series and resorts to differencing the data to achieve stationarity. Developing such tests would be a valuable future direction.
  • Model Misspecification: While conformal prediction offers robustness to model misspecification for coverage, the efficiency of the prediction bands is still influenced by the accuracy of the point predictor. The paper explores various FAR(1) estimation techniques, but further research into optimal forecasting for 2D functional time series could lead to even more efficient bands.
  • Generalization to FAR(p): The current work focuses on FAR(1). Extending the methodology to FAR(p) models with p > 1 for two-dimensional functional time series is a natural next step, although it presents estimation challenges.
  • Computational Cost: While inductive conformal prediction is more efficient than transductive methods, dealing with high-dimensional functional data can still be computationally intensive. Further optimization of the algorithms could be beneficial.

Summary updated

BibTeX
@article{ajroldi2023conformal,
  author = {Niccolò Ajroldi and Jacopo Diquigiovanni and Matteo Fontana and Simone Vantini},
  title = {{Conformal prediction bands for two-dimensional functional time series}},
  journal = {Comput. Stat. Data Anal},
  year = {2023},
  doi = {10.1016/j.csda.2023.107821},
}
Figure