Abstract:Conformal prediction (CP) has been a popular method for uncertainty quantification because it is distribution-free, model-agnostic, and theoretically sound. For forecasting problems in supervised learning, most CP methods focus on building prediction intervals for univariate responses. In this work, we develop a sequential CP method called $\texttt{MultiDimSPCI}$ that builds prediction $\textit{regions}$ for a multivariate response, especially in the context of multivariate time series, which are not exchangeable. Theoretically, we estimate $\textit{finite-sample}$ high-probability bounds on the conditional coverage gap. Empirically, we demonstrate that $\texttt{MultiDimSPCI}$ maintains valid coverage on a wide range of multivariate time series while producing smaller prediction regions than CP and non-CP baselines.

What problem does this paper attempt to address?

### Main Problem Addressed by the Paper This paper primarily addresses the issue of uncertainty quantification in multidimensional time series forecasting, particularly in the context of non-exchangeable data (such as multivariate time series) using the **Conformal Prediction (CP)** method to construct prediction regions. ### Core Contributions 1. **Proposed a new sequential conformal prediction method** (MultiDimSPCI), which can establish ellipsoidal prediction regions for multivariate time series. This method dynamically re-estimates the size of the ellipsoid during the testing phase to ensure that the prediction regions are both adaptive and accurate. 2. **Provided finite-sample high-probability bounds** to assess the coverage gap of the constructed prediction regions, which do not rely on the exchangeability of observations. 3. **Empirical validation** shows that on multivariate time series data (up to 20 dimensions), MultiDimSPCI can produce smaller prediction regions than CP and other non-CP baseline methods without sacrificing coverage. ### Method Overview - **Ellipsoidal Uncertainty Sets**: The paper first defines uncertainty sets based on ellipsoidal shapes and calibrates the radius of the ellipsoid using the conformal prediction method for univariate time series. This method considers the covariance matrix of the residuals and adaptively estimates the quantiles of the non-conformity scores through the Sequential Conformal Inference (SPCI) method. - **Comparison with Copula Methods**: The paper compares the proposed method with Copula methods, noting that Copula methods require searching for multidimensional vectors and return hyper-rectangular prediction sets, whereas the proposed MultiDimSPCI method only needs to estimate the covariance matrix of the residuals and produces smaller prediction regions. - **Theoretical Analysis**: The paper also provides theoretical guarantees on conditional coverage, demonstrating that when using the empirical quantile function as the quantile regression predictor, finite-sample high-probability bounds can be obtained. ### Experimental Results - Experiments demonstrate that MultiDimSPCI can maintain effective coverage on multivariate time series data (up to 20 dimensions) while producing smaller prediction regions compared to other CP and non-CP baseline methods. In summary, this paper proposes a novel and effective method for uncertainty quantification in multivariate time series forecasting and provides substantial theoretical and empirical support for its approach.

Conformal prediction for multi-dimensional time series by ellipsoidal sets

Sequential Predictive Conformal Inference for Time Series

Conformal Prediction Regions for Time Series using Linear Complementarity Programming

Conformal Prediction: A Data Perspective

Copula Conformal Prediction for Multi-step Time Series Forecasting

Conformal Prediction Intervals with Temporal Dependence

Multi-model Ensemble Conformal Prediction in Dynamic Environments

An Information Theoretic Perspective on Conformal Prediction

Spatial Conformal Inference through Localized Quantile Regression

Robust Conformal Prediction under Distribution Shift via Physics-Informed Structural Causal Model

Efficient Conformal Prediction under Data Heterogeneity

Split Conformal Prediction and Non-Exchangeable Data

Conformal pid control for time series prediction

Predictive Inference with Feature Conformal Prediction

Efficient Conformal Prediction via Cascaded Inference with Expanded Admission

Conformalized Deep Splines for Optimal and Efficient Prediction Sets

Posterior Conformal Prediction

Conformal Predictions for Longitudinal Data

JANET: Joint Adaptive predictioN-region Estimation for Time-series