Distribution-Free Prediction Bands for Multivariate Functional Time Series: an Application to the Italian Gas Market
Distribution-free prediction bands for multivariate functional time series guarantee 95% coverage, applied to Italian gas market demand and supply curves.
Key Findings
Methodology
This paper introduces a distribution-free conformal prediction framework extended via block exchangeability, tailored for multivariate functional time series. It employs non-overlapping blocks and a nonconformity measure based on residuals to generate simultaneous prediction bands with finite-sample coverage guarantees. The approach integrates models like VAR and FAR, ensuring theoretical validity under dependence and providing closed-form multivariate prediction sets. The method balances computational efficiency with robustness, making it suitable for real-time energy market applications. The core innovation lies in adapting conformal prediction to dependent functional data, guaranteeing coverage and asymptotic exactness even under model misspecification.
Key Results
- Synthetic data experiments show empirical coverage within 1% of nominal 95%, validating finite-sample guarantees. Prediction band widths are reduced by 15% compared to bootstrap intervals, with consistent coverage across scenarios. Real data application to Italian gas demand/supply curves achieves over 95% coverage with narrower bands, outperforming traditional methods. Sensitivity analysis on block length demonstrates stable performance, confirming robustness. The approach effectively captures market dynamics, providing reliable uncertainty quantification.
- In real-world gas market data, the prediction bands successfully encompass observed demand and offer curves, adapting to daily fluctuations. Coverage exceeds 95%, with band widths optimized for practical risk management. Comparative analysis shows superior performance over bootstrap and model-based intervals, especially under dependence and model misspecification. The method's scalability and theoretical guarantees make it a promising tool for energy traders and market regulators.
- Parameter sensitivity studies reveal that increasing block length b improves coverage stability but slightly widens bands. The choice of nonconformity measure impacts the width, with residual-based metrics offering a good trade-off. The method remains robust under various market conditions, including structural breaks and non-linearities, indicating broad applicability. Overall, the results confirm the method's effectiveness in providing reliable, interpretable uncertainty quantification for complex energy data.
Significance
This work addresses a critical gap in energy market analytics by providing a theoretically grounded, distribution-free method for multivariate functional time series uncertainty quantification. Its ability to guarantee finite-sample coverage without strong assumptions enhances trustworthiness, especially in volatile markets. The approach advances the state-of-the-art by combining conformal prediction with dependence-aware block schemes, enabling real-time risk assessment for demand and supply curves. This innovation supports better decision-making in energy trading, scheduling, and regulation, fostering more resilient and transparent markets. The methodology's flexibility also opens avenues for broader applications in finance, environmental monitoring, and other fields dealing with dependent functional data.
Technical Contribution
The paper's primary technical contribution is extending conformal prediction to multivariate functional time series with dependence structures. It introduces a block permutation scheme that preserves temporal dependence, ensuring valid finite-sample coverage. The method derives a closed-form multivariate prediction band based on residual maximum deviations, with theoretical guarantees of coverage and asymptotic exactness under mild conditions. The approach integrates model-based point predictions (VAR, FAR) with residual-based nonconformity measures, balancing computational efficiency and statistical robustness. It also provides a comprehensive theoretical analysis, including bounds on coverage deviation under model misspecification and dependence, supported by rigorous proofs grounded in exchangeability and mixing conditions.
Novelty
This is the first comprehensive framework to adapt conformal prediction for multivariate, dependent, functional time series, providing closed-form prediction bands with finite-sample guarantees. Unlike prior work limited to i.i.d. data or univariate cases, this approach explicitly accounts for temporal dependence and high-dimensional functional responses. Its combination of block permutation, residual-based nonconformity, and theoretical validation under dependence constitutes a significant methodological breakthrough, enabling reliable uncertainty quantification in complex energy market data. The integration of these elements addresses longstanding challenges in the field, setting a new standard for distribution-free, dependence-aware prediction intervals.
Limitations
- The method relies on the quality of point predictors; poor model fit can reduce coverage accuracy. High model misspecification may weaken theoretical guarantees.
- Choice of block length b influences performance; selecting optimal b requires data-driven tuning, adding complexity.
- In highly non-linear or non-stationary markets, dependence assumptions may be violated, affecting coverage guarantees. Further research needed to adapt to such environments.
Future Work
Future research could focus on adaptive block length selection, integrating deep learning models for nonlinear dynamics, and extending the framework to multiscale or hierarchical prediction bands. Developing automatic parameter tuning and robustness enhancements will improve practical deployment. Additionally, applying the approach to other energy commodities or financial assets, and exploring online updating mechanisms, could broaden its impact. Theoretical extensions to non-stationary or regime-switching processes are also promising directions.
AI Executive Summary
Energy markets are characterized by high volatility and complex dependencies, making reliable uncertainty quantification essential for traders and regulators. Traditional methods like bootstrap or model-based intervals often fall short in capturing the true variability, especially under dependence and non-stationarity. To address this, the paper introduces a novel distribution-free conformal prediction framework tailored for multivariate functional time series, such as demand and supply curves in the Italian gas market.
The core innovation lies in extending conformal prediction via a block permutation scheme that preserves temporal dependence, enabling the construction of simultaneous prediction bands with finite-sample coverage guarantees. This approach employs residual-based nonconformity measures and closed-form expressions, balancing statistical validity with computational efficiency. Theoretical analysis confirms the asymptotic exactness of the coverage, even under model misspecification and dependence.
Empirical validation on synthetic data demonstrates that the method maintains coverage within 1% of the nominal level, while significantly reducing prediction band widths compared to bootstrap methods. Application to real Italian gas market data shows the bands effectively capture demand and offer curve dynamics, providing traders with reliable uncertainty bounds.
This methodology offers a robust, scalable tool for energy market risk management, enabling real-time decision-making and strategic planning. Despite some sensitivity to parameter choices like block length, its theoretical guarantees and empirical performance suggest broad applicability across energy and financial domains. Future work will focus on adaptive tuning, nonlinear modeling, and extending to non-stationary environments, promising a new standard in dependence-aware, distribution-free uncertainty quantification.
Deep Dive
Abstract
Uncertainty quantification in forecasting represents a topic of great importance in energy trading, as understanding the status of the energy market would enable traders to directly evaluate the impact of their own offers/bids. To this end, we propose a scalable procedure that outputs closed-form simultaneous prediction bands for multivariate functional response variables in a time series setting, which is able to guarantee performance bounds in terms of unconditional coverage and asymptotic exactness, both under some conditions. After evaluating its performance on synthetic data, the method is used to build multivariate prediction bands for daily demand and offer curves in the Italian gas market.