Distribution-Free Prediction Bands for Multivariate Functional Time Series: an Application to the Italian Gas Market
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.
Tom: We're looking at the paper: "Distribution-Free Prediction Bands for Multivariate Functional Time Series".
Jane: Here's the gist of the paper. Distribution-Free Prediction Bands for Multivariate Functional Time Series: an Application to the Italian Gas Market proposes a scalable procedure that outputs closedform 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 <ref:2107.00527#pg2> The aim of the present work is explictly tackle the need of uncertainty quantification in the prediction of gas supply and demand curves, by reaching a synthesis between two rich and ebullient fields of methodological literature (namely, set forecasting for (multivariate) FTS and finite-sample valid distribution-free prediction) by proposing distribution-free prediction bands for multivariate functional time series guaranteeing finite-sample performance bounds in terms of coverage and asymptotic exactness, i.e.
Tom: coverage asymptotically equal to the nominal confidence level <ref:2107.00527#pg4> Methodological Framework The framework considered is very general since it includes the case of univariate functional response variable (when p = 1) as a special case, but it also allows the p > 1 domains Qj and images of Yt j to be greatly different when j changes <ref:2107.00527#pg3> The authors aim to introduce a procedure able to create simultaneous prediction sets for YT+1 (i.e.
Jane: prediction sets holding for the multivariate random function YT+1 globally, and not only for its jth component YT+1, j) based on the information provided by Z1,..., ZT and by XT+1 and ensuring performance bounds in terms of unconditional coverage <ref:2107.00527#pg5> A specific aim is to find a prediction set having a particular shape, known as multivariate functional prediction band
19: , which is defined as y = (y1,..., yp) ∈ Y p j=1 L∞(Qj): yj(q) ∈ Bj(q), ∀ j ∈ 1,..., p, ∀q ∈ Qj <ref:2107.00527#pg6> Time Series Dependence Handling Moving away from the requirements of data exchangeability, a relatively new contribution by
20: shows an extension of Conformal methods to time dependent data <ref:2107.00527#pg4> The authors present a modification of the Conformal inference able to account for time series dependence by generalizing the Conformal approach traditionally used when the regression pairs Z1,..., ZT+1 are i.i.d.
Tom: by randomizing blocks of observations <ref:2107.00527#pg5> Specifically, they extend the non-overlapping blocking scheme proposed by Chernozhukov et al.
Paper summary: Tom: So to wrap up this paper, the authors of "Distribution-Free Prediction Bands for Multivariate Functional Time Series: an Application to the Italian Gas Market" have basically put together a scalable procedure for generating closed form simultaneous prediction bands.
Jane: They’re tackling uncertainty quantification in forecasting demand and supply curves in a multivariate functional time series framework, which is a new and relevant tool for energy trading.
Lu: What they've done is extend the non-overlapping blocking scheme from Chernozhukov et al to the split context to create simultaneous prediction bands for future multivariate random functions <ref:2107.00527#pg5>.
Meng: The core idea is that Theorem two point one provides a theoretically sound prediction framework based on assumptions similar to those introduced by Chernozhukov et al <ref:2107.00527#pg5> <ref:2107.00527#pg1>.
Tom: They found finite-sample performance bounds in terms of unconditional coverage and asymptotic exactness under some specific conditions involving the oracle nonconformity measure A star and the nonconformity measure A <ref:2107.00527#pg5>.
Jane: And they showed that this approach can also be applied to the multivariate functional context because of the split process and the specific nonconformity measure they used <ref:2107.00527#pg5>.
Lu: The authors conclude that Theorem two point one provides a prediction framework based on assumptions similar to those introduced by Chernozhukov et al <ref:2107.00527#pg5> <ref:2107.00527#pg1>.
Meng: They suggest that the prediction region naturally induced by their method is a promising tool, both for theory and for application in trading <ref:2107.00527#pg5>.
Tom: It's essentially saying this method can be profitably included in traders’ tool kits, which is the big implication here <ref:2107.00527#pg5>.
Conclusion: Tom: So, what is the big picture here? What are these authors actually trying to achieve with this work?
Jane: They're basically showing a way to put uncertainty quantification onto forecasting demand and supply curves in a complex, multivariate time series setting. It’s about making predictions reliable even when you don't know the exact statistical distribution of your data.
Lu: The main contribution is proposing these prediction bands that offer performance guarantees—specifically, finite-sample bounds for coverage and asymptotic exactness—in this functional context.
Meng: But what does that actually mean for us at scale? If we're deploying this in a real energy market, how robust are these guarantees when the underlying data isn't perfectly behaved?
Tom: Well, they show it works under certain conditions. The application to the Italian gas market used a rolling window of ninety days of data to create these simultaneous prediction bands for daily offer and demand curves.
Jane: And the empirical results were pretty close to what they promised; only about two out of sixty three confidence intervals missed the nominal level, which is actually quite good for a complex system like that.
Lu: It extends existing ideas from other time-dependent data methods by modifying blocking schemes to handle time series dependence in a split framework.
Meng: So, if you had to deploy this right now, what’s the practical trade-off? Is it more complicated than just using standard VAR models?
Tom: It’s about getting those simultaneous bands for multiple variables at once. The implication is that this could be a useful tool for traders who need to quantify risk in a multivariate environment.
Jane: It really suggests that you don't always need perfect distributional assumptions to get meaningful bounds on your forecasts, which is a big deal in real-world data science.
Meng: It’s promising, but I still have my reservations about how much of the data generation process has to be perfectly modeled for these guarantees to hold tight.
Lu: That's where the theoretical work continues—finding ways to make those assumptions even more flexible while keeping the performance bounds intact.
Tom: Right, so it’s a solid theoretical foundation being built, and now we look at how that translates into actual decision-making tools.
Jacopo Diquigiovannia, Matteo Fontanac, Simone Vantinic
Department of Statistical Sciences, University of Padova · European Commission, Joint Research Centre (JRC), Ispra · MOX - Department of Mathematics, Politecnico di Milano
stat.ME, stat.AP
Submitted: 2021-07-01
Updated: 2026-11-15
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 81/100
The gist: Distribution-Free Prediction Bands for Multivariate Functional Time Series: an Application to the Italian Gas Market proposes a scalable procedure that outputs closedform simultaneous prediction…
Key concepts
- Multivariate Functional Time Series (FTS)
- This refers to a time series where each observation is a function of multiple variables over time. The paper deals with predicting these complex, multi-dimensional functions simultaneously, which is crucial for understanding supply and demand curves in the gas market.
- Distribution-Free Prediction Bands
- These are prediction intervals that do not require assuming a specific underlying probability distribution for the data. They provide robust bounds on how accurate a forecast is, ensuring reliable performance regardless of the exact statistical nature of the time series.
- Conformal Methods Extension (Split Framework)
- The authors modify conformal inference to handle time-dependent data by randomizing blocks of observations. This allows them to apply distribution-free prediction techniques even when the data points are not independent and identically distributed, making it suitable for real-world time series.
- Asymptotic Exactness
- This means that as the sample size grows very large, the coverage rate of the prediction bands approaches the nominal confidence level. It ensures that in theory, a prediction band set at 95% confidence will cover the true value approximately 95% of the time.
Terminology
Summary
Distribution-Free Prediction Bands for Multivariate Functional Time Series: an Application to the Italian Gas Market proposes a scalable procedure that outputs closedform 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 <ref:2107. p. 0 Distribution-Free Prediction Bands for Multivariate Functional
Time Series: an Application to the Italian Gas Market
Jacopo Diquigiovannia,∗
, Matteo Fontanac,b, Simone Vantinic
aDepartment of Statistical Sciences, University of Padova, Italy
bnow at European Commission, Joint Research Centre (JRC), Ispra, (VA), Italy
cMOX - Department of Mathematics, Politecnico di Milano, Italy
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 closedform 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. ebullient fields of methodological literature (namely, set forecasting for (multivariate) FTS and
finite-sample valid distribution-free prediction) by proposing distribution-free prediction bands
for multivariate functional time series guaranteeing finite-sample performance bounds in terms
of coverage and asymptotic exactness, i. coverage asymptotically equal to the nominal confidence level. ebullient fields of methodological literature (namely, set forecasting for (multivariate) FTS and
finite-sample valid distribution-free prediction) by proposing distribution-free prediction bands
for multivariate functional time series guaranteeing finite-sample performance bounds in terms
of coverage and asymptotic exactness, i. coverage asymptotically equal to the nominal confidence level.Where the paper says it
Where the paper says it
Where the paper says it
Methodological Framework
The framework considered is very general since it includes the case of univariate functional response variable (when p = 1) as a special case, but it also allows the p > 1 domains Qj and images of Yt j to be greatly different when j changes <ref:2107.00527#pg3> The authors aim to introduce a procedure able to create simultaneous prediction sets for YT+1 (i. p. 2 Let µ
j
(xt j) = E(Yt j|Xt j = xt j) be the
regression map for the j-th component evaluated at xt j, and similarly let us define the scalar
quantity [µ
j
(xt j)](q) = E(Yt j(q)|Xt j = xt j). Our aim is to introduce a procedure able to create simultaneous prediction sets for YT+1 (i. Our aim is to introduce a procedure able to create simultaneous prediction sets for YT+1 (i. prediction sets holding for the multivariate random function YT+1 globally, and not only for its jth component YT+1, j) based on the information provided by Z1, . prediction sets holding for the multivariate random function YT+1 globally, and not only for its jth component YT+1, j) based on the information provided by Z1, . , ZT and by XT+1 and ensuring
performance bounds in terms of unconditional coverage. , ZT and XT+1 as CT,1−α (XT+1) for any significance level α ∈ (0, 1), the
purpose is to obtain prediction sets whose unconditional coverage P
YT+1 ∈ CT,1−α (XT+1)
is
close to the nominal confidence level 1−α under mild conditions on the data generating process. Within the possible kinds of prediction set, the specific aim, moreover, is to find a prediction
set having a particular shape, known as multivariate functional prediction band [19], which is
defined as
y = (y1, . vantini@polimi. it (Simone Vantini)
Preprint submitted to Elsevier January 17, 2024
arXiv:2107.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Time Series Dependence Handling
Moving away from the requirements of data exchangeability, a relatively new contribution by [20] shows an extension of Conformal methods to time dependent data <ref:2107.00527#pg4> The authors present a modification of the Conformal inference able to account for time series dependence by generalizing the Conformal approach traditionally used when the regression pairs Z1,. p. 3 , 15, 25, 17], we present a modification of the Conformal inference
able to account for time series dependence. Intuitively, the idea is to generalize the Conformal
approach traditionally used when the regression pairs Z1, . , ZT+1 are i. by randomizing
blocks of observations. by randomizing
blocks of observations. Specifically, we extend the non-overlapping blocking scheme proposed
by Chernozhukov et al. Specifically, we extend the non-overlapping blocking scheme proposed
by Chernozhukov et al. [20] to the Semi-Off-Line Inductive (simply known as Split) framework. Intuitively, the idea is the one introduced by Split Conformal Prediction: after randomly
dividing the observed data into the training and calibration sets, the prediction set CT,1−α (xT+1) is
defined as the set of all y ∈
Qp
j=1
L
∞(Qj) such that (xT+1, y) is similar - in terms of nonconformity
measure A - to the training set {zh : h ∈ I1} compared to the conformity of the elements of
the calibration set to the same training set. Differently from the Split Conformal Prediction
framework, the permutation scheme here proposed randomizes the elements of the calibration
set by considering blocks of observations of length b, and it computes the nonconformity scores
for (l + 1)/b − 1 elements of the calibration set (one for each block) and for zT+1.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Performance Guarantees
The permutation scheme Π guarantees that, if the regression pairs Z1,..., ZT+1 are i.i.d. (or even exchangeable), the prediction sets obtained are finite-sample valid, i. p. 4 , ZT+1 are i. (or even exchangeable), the prediction sets obtained are finite-sample
valid, i. vantini@polimi. it (Simone Vantini)
Preprint submitted to Elsevier January 17, 2024
arXiv:2107. The proof can be trivially obtained by generalizing the well-established
result holding in the Split Conformal framework [15]. As in the Conformal setting, the result
concerning the validity of the prediction sets induced by the permutation scheme Π can be enriched by proving that, if the nonconformity scores have a continuous joint distribution, then
P
YT+1 ∈ CT,1−α (XT+1)
= 1 −
⌊α(l+1)/b⌋
(l+1)/b
, i. If the regression pairs are not exchangeable, as in our case, the aforementioned results do
not hold. Nevertheless, two desirable properties can still be obtained under some conditions:
finite-sample performance bounds in terms of unconditional coverage and asymptotic exactness
(i. Nevertheless, two desirable properties can still be obtained under some conditions:
finite-sample performance bounds in terms of unconditional coverage and asymptotic exactness
(i. unconditional coverage asymptotically equal to the nominal confidence level 1 − α). the unconditional coverage is equal to an easyto-compute value and it is not only greater than or equal to 1 − α. Also in this case, the proof
can be trivially obtained by generalizing an existing result concerning Conformal Prediction [see
Theorem 1 of 18]. the unconditional coverage is equal to an easyto-compute value and it is not only greater than or equal to 1 − α. Also in this case, the proof
can be trivially obtained by generalizing an existing result concerning Conformal Prediction [see
Theorem 1 of 18]. In so doing, {δ1¯l
, γ1¯l
} should depend on ¯l. 1 provides, under some conditions, finite-sample performance bounds in terms
of unconditional coverage regardless the value of b and it guarantees that the prediction sets are
asymptotically exact since the right side of Inequality (1) converges to 0 when m,
¯l → 0.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Application Results
The application is presented in Section 4 while Section 5 presents conclusions and draws further developments <ref:2107.00527#pg4> The authors apply the procedure to a specific Italian gas market, namely the MGS (Mercato Gas in Stoccaggio), in order to create simultaneous prediction bands for the daily offer and demand curves <ref:2107. p. 0 To this end, we propose a scalable procedure that outputs closedform 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. Our aim is to introduce a procedure able to create simultaneous prediction sets for YT+1 (i. prediction sets holding for the multivariate random function YT+1 globally, and not only for its jth component YT+1, j) based on the information provided by Z1, . vantini@polimi. it (Simone Vantini)
Preprint submitted to Elsevier January 17, 2024
arXiv:2107. Specifically,
the empirical coverage is simply computed as the fraction of the N = 5000 replications in which
yT+1 belongs to CT,1−α (xT+1), and the confidence interval is reported in order to provide an
idea of the variability of the phenomenon, rather than to make inferential conclusion on the
unconditional coverage in the various settings. The evidence is quite satisfactory as for any value
of b, sample size T and model the empirical coverage is close to 1 − α = 0. The evidence is quite satisfactory as for any value
of b, sample size T and model the empirical coverage is close to 1 − α = 0. 75, as suggested
by the fact that only 2 out of the 63 confidence intervals (the cells in bold in the tables) do not
include the nominal confidence level.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Conclusion and Utility
The present work deals with the issue of forecasting demand and supply curves in on the Gas Market, proposing methods able to quantify uncertainty in the multivariate functional time series prediction framework, a new and particularly relevant workhorse in the energy trading field <ref:2107.00527#pg5> The approach developed extends the non-overlapping blocking scheme proposed by Chernozhukov et al. p. 3 by randomizing
blocks of observations. Specifically, we extend the non-overlapping blocking scheme proposed
by Chernozhukov et al. Our aim is to introduce a procedure able to create simultaneous prediction sets for YT+1 (i. prediction sets holding for the multivariate random function YT+1 globally, and not only for its jth component YT+1, j) based on the information provided by Z1, . Secondly, the regression estimators have
a fundamental impact on the computational cost: indeed, the procedure here developed is highly
scalable since, conditional on the computational cost required to obtain the regression estimates
and sj,I1
, the time needed to compute the prediction set increases linearly with T by assuming
the ratio T/l and b fixed, and consequently the computational effort is mainly determined by the
regression estimators used. The strategy proposed in this Section represents a theoretically sound framework to obtain
prediction bands when dealing with multivariate functional time series. vantini@polimi. it (Simone Vantini)
Preprint submitted to Elsevier January 17, 2024
arXiv:2107.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Pruning redundant reasoning tokens
The paper is structured as follows: Section 2 presents the proposed method and its theoretical underpinnings in detail, while in Section 3 and 3.1 we present a simulation study to assess the properties of the method and its results <ref:2107. p. 2 The paper is structured as follows: Section 2 presents the proposed method and
its theoretical underpinnings in detail, while in Section 3 and 3. 1 we present a simulation study
to assess the properties of the method and its results. 1 we present a simulation study
to assess the properties of the method and its results. The application is presented in Section 4
while Section 5 presents conclusions and draws further developments.Where the paper says it
Where the paper says it
How the training signal is built
The authors consider a data generating process defined as Yt1(q) = g′(q) · Y¯ t = Y¯ t1 + Y¯ t2 sin(2πq) √ 1/2 + Y¯ t3 cos(2πq) √ 1/2 with Y¯′t = [Y¯t1, Y¯t2, Y¯t3] a VAR(2) process <ref:2107.00527#pg5> The point predictions are obtained by considering both g(q) and Ψ¯1, Ψ¯2 as known <ref:2107. p. 7 A graphical
representation of a replication with T = 25 is provided in Figure 1. Seven models are taken into account:
• Oracle Model The point predictions are obtained by considering both g(q) and Ψ¯
1, Ψ¯
2 as
known. • VAR Models The point predictions are obtained by considering g(q) as known and Ψ¯
1, Ψ¯
2
as unknown. Specifically, the point prediction for Yt1(q) is obtained by estimating r
matrices Ψ¯
1, . • FAR Models The point predictions are obtained by considering both g(q) and Ψ¯
1, Ψ¯
2 as
unknown. The point predictions are obtained by fitting (on the training set, as usual) a
concurrent function-on-function autoregressive model of order r ∈ {1, 2, 3}, i. vantini@polimi. it (Simone Vantini)
Preprint submitted to Elsevier January 17, 2024
arXiv:2107.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the gains come from
The authors evaluate the procedure by taking into account scenarios characterized by a decreasing knowledge of the data generating process, including three widespread kinds of model misspecification: the misspecification due to omitted relevant variable [VAR Model with r = 1, see 27], misspecification due to inclusion of irrelevant variable [VAR Model with r = 3, see 27] and the functional form misspecification [FAR Models, see 28] <ref:2107.00527#pg5> The simulation study shows that the Oracle model outperforms the VAR models and the FAR models for every value of T, b, as expected <ref:2107. p. 9 Also the training set size has a relevant
impact on the phenomenon since one is justified in expecting the band size to decrease when
m grows because more accurate regression estimates provide smaller nonconformity scores, as
suggested by analyzing the three couples (T, b) = {(25, 1), (50, 3), (100, 6)} in which the number
of nonconformity scores computed is constant. The Oracle model outperforms the VAR models and the FAR models for every value of T, b,
as expected. Specifically,
the empirical coverage is simply computed as the fraction of the N = 5000 replications in which
yT+1 belongs to CT,1−α (xT+1), and the confidence interval is reported in order to provide an
idea of the variability of the phenomenon, rather than to make inferential conclusion on the
unconditional coverage in the various settings. The evidence is quite satisfactory as for any value
of b, sample size T and model the empirical coverage is close to 1 − α = 0. The result provided by the simulation study is particularly
appealing since it suggests that an appropriate coverage is reached also when the sample size is
very small, a fact that allows the procedure to be applied in many practical frameworks. Although comparing the size of prediction bands obtained in scenarios characterized by (potentially) different unconditional coverages may lead to misleading conclusions, in light of the
evidence provided so far we evaluate this aspect when the model, the value of T and the value of
b vary.Where the paper says it
Where the paper says it
Where the paper says it
The authors conclude that the prediction region naturally induced by the method described in Section 2 represents a promising tool - both from a theoretical and an application point of view - that can be profitably included in traders’ tool kit <ref:2107.00527#pg5> The procedure inherits the guarantees for the unconditional coverage in terms of finite-sample performance bounds and of asymptotic exactness under some conditions concerning the oracle nonconformity measure A∗ and the nonconformity measure A, but can be also satisfactorily applied to the multivariate functional context due to the Split process and to the specific nonconformity measure used [firstly introduced by 19] <ref:2107. p. 5 1 provides, under some conditions, finite-sample performance bounds in terms
of unconditional coverage regardless the value of b and it guarantees that the prediction sets are
asymptotically exact since the right side of Inequality (1) converges to 0 when m,
¯l → 0. In view
of this, the last, fundamental step in describing a strategy which can be used in practical applications is the definition of the nonconformity measure, a nontrivial issue in the functional setting.Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Improved AI system can generate closedform simultaneous prediction bands for multivariate functional response variables in a time series setting, guaranteeing performance bounds in terms of unconditional coverage and asymptotic exactness under some conditions.
-
The improved system can construct multivariate functional prediction bands for daily demand and offer curves in the Italian gas market using distribution-free prediction bands derived from a modification of Conformal inference that accounts for time series dependence via a Split framework extension.
-
The system can obtain finite-sample performance bounds in terms of unconditional coverage and asymptotic exactness even when the regression pairs are not exchangeable, as detailed in Theorem 2.1, under specific conditions related to the nonconformity measures and model misspecification tolerances.
-
The improved system can automatically obtain a prediction region for (Qt, Pt) by considering the region where the prediction band for the offer curve and that for the demand curve overlap, leading to empirical coverage values close to 1 minus alpha even when sample sizes are small or models are misspecified.
-
The system can directly evaluate the impact of any extra offer/bid on tomorrow's offer and demand curves by adding it to the predicted demand curve, inducing a change in the multivariate prediction band and consequently in the prediction region for the intersection point, allowing for speculative analysis.
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.
Sources
- Bootstrap Prediction Bands for Functional Time Series
- Conformal Prediction: a Unified Review of Theory and New Challenges
- The Importance of Being a Band: Finite-Sample Exact Distribution-Free Prediction Sets for Functional Data
- Conformal prediction for time series
Related papers
- Neural Estimation of Summary Statistics for Simulation-Based Inference with Non-Asymptotic Guarantees
- Spectral divide-and-conquer MCMC for long stationary time series
- Improving the Efficiency of Subgroup Analysis in Randomized Controlled Trials with TMLE
- Personalized Federated Vector Autoregression with Personalization Diversity
- Doubly Robust and Efficient Calibration of Prediction Sets for Right-Censored Time-to-Event Outcomes
- Covariance-Calibrated Prediction-Powered Inference