On the modelling and prediction of high-dimensional functional time series

arXiv:2406.00700 · stat.ME, econ.EM, math.ST, stat.TH · Submitted 2024-06-02 · Read on arXiv

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: I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: We're looking at the paper: "On the modelling and prediction of high-dimensional functional time series".

Jane: Here's the gist of the paper. The paper proposes a two-step procedure for modeling and predicting high-dimensional functional time series where p, the number of function-valued time series, is large relative to n, the length of the time series.

Tom: The first step involves a segmentation transformation, seeking a linear transformation Y tp = A Z tp, where A is an invertible constant matrix.

Jane: This transformation aims to segment the original high-dimensional functional time series into q groups such that any two subseries from different groups are uncorrelated both contemporaneously and serially.

Tom: The estimation of the transformation matrix A is achieved through eigenanalysis of a positive definite matrix derived from the double integral of quadratic forms in the autocovariance functions of Y tp.

Paper summary: Tom: So, this paper is called "On the modelling and prediction of high-dimensional functional time series," and it tackles a problem where you have a huge number of different time series, p, compared to how long your actual data record is, n.

Jane: Right, so when we talk about functional time series—those are records that aren't just single numbers but whole curves over time, like temperature readings from twenty different weather stations simultaneously—this paper is looking at what happens when there are too many of those curves.

Lu: The main idea they propose is a two-step procedure to handle this high-dimensional setting where p is larger than n. They start with a segmentation transformation, which turns the original messy functional data into simpler groups that don't interact much with each other.

Meng: Segmentation sounds like it could be useful for managing complexity, but what exactly are they trying to achieve with this transformation? Is it just sorting them by how they look?

Tom: No, not just sorting. They want to find a linear transformation matrix A that takes the original high-dimensional functional time series and segments it into several groups where any two subseries from different groups are uncorrelated, both at the same time and over time too.

Jane: That sounds like they're trying to separate the noise or the underlying patterns so each group can be modeled independently instead of everything being tangled up together.

Lu: Exactly. They estimate this transformation matrix A by doing eigenanalysis on a positive definite matrix that comes from looking at the double integral of quadratic forms in the autocovariance functions of those original time series.

Meng: Eigenanalysis sounds computationally heavy when you're dealing with potentially infinite dimensions, so how do they handle that part? Is it feasible to run that analysis when p is really big?

Tom: That’s where their second step comes in. After segmentation, they take each of those subseries and try to find a finite-dimensional dynamic structure for them on its own.

Jane: So, once they have these separated groups, what's the next level of modeling they use for each individual group?

Paper summary: Lu: They assume that this dynamic structure can be represented in a specific way, called a latent decomposition where you break down the subseries into some dynamics and some white noise.

Meng: White noise is standard, but they also make a specific assumption about how the dynamics relate to the noise over time lags. What does that mean for prediction?

Tom: It means that the part representing the past dynamics of that group is uncorrelated with its future white noise component, which simplifies forecasting immensely because you can model them separately.

Jane: But how do they actually use this structure to make a prediction for the original data? Is it just plugging in numbers from these separate models?

Tom: They do apply the learned dynamics for each group individually. For every subseries, they pick the best variable autoregressive model based on something called the AIC criterion to get an h-step ahead prediction.

Meng: So, you run multiple different VAR models on each group and pick the one with the lowest AIC score for a specific forecast horizon? That sounds like a lot of computational work at scale.

Jane: It is intensive, but it’s structured because they are treating each group as its own small problem instead of one giant impossible problem.

Lu: And then they take those predicted vector time series from all the groups and transform them back using that original matrix A to get the final prediction for the entire original high-dimensional functional time series.

Tom: The paper claims this approach significantly outperforms existing methods like SegV, UniV, and GSY across different data types, including things like age-specific mortality and energy consumption data.

Jane: That’s a big claim because those datasets are often messy and complex to model correctly at this scale. What makes it stand out from what's already out there?

Lu: The authors argue that the segmentation transformation they introduced is a real step forward from earlier ideas about decorrelation for vector time series, adapting it specifically to functional data where the dimension is infinite.

Meng: So, you’re saying their specific way of splitting the data into uncorrelated groups is what makes a difference in the accuracy of your prediction compared to just running standard methods on one massive set?

Paper summary: Tom: It seems that incorporating that serial dependence information through this lagged autocovariance approach in the second step is what really boosts its predictive power over simpler dimension reduction techniques.

Jane: So, for someone who just wants to know what this means practically, it suggests a way to take massive, complicated functional data sets and break them down into manageable pieces so you can forecast them with much more accuracy than before.

Lu: Theoretically, they've established some consistency results for the segmentation step in theorem one and uniform convergence rates for their dimension reduction estimation in theorem two.

Tom: And they also identified the dynamic space spanned by those eigenfunctions in theorem three, along with uniform convergence rates for identifying that space itself in theorem four.

Meng: Those theoretical guarantees are important, but what about the practical reality? How much computational overhead does this two-step process introduce when you actually deploy it on a system like mine?

Tom: Well, the paper shows how it scales, suggesting that while the setup is complex, the individual modeling steps for each group are relatively standard and can be done independently.

Jane: So, to wrap up on "On the modelling and prediction of high-dimensional functional time series," they are essentially proposing a structured way to decompose high-dimensional functional data into uncorrelated parts, model those parts separately using dynamic structures, and then stitch the predictions back together.

Lu: And it suggests that this method has shown strong predictive performance on real datasets compared to other established techniques.

Meng: It’s good that they have concrete results, but I still worry about deploying something that relies on eigenanalysis of a potentially huge matrix before I even know if the underlying assumptions hold up perfectly.

Tom: The paper does mention the limitations, specifically that it relies on the assumption of stationarity for their dynamic structure estimation and that they are focusing on linear dynamics in their second step.

Jane: So, while they've made some strong claims about performance and theory, the method is limited by assuming stationarity and sticking to linear models for the dynamics.

Lu: That’s a fair limitation to point out—the work is focused on what it can do under those specific assumptions regarding the functional time series structure.

Conclusion: Tom: So, we’re wrapping up this look at "On the modelling and prediction of high-dimensional functional time series."

Jane: It boils down to taking those massive, complex records with tons of different time series and breaking them into smaller groups that don't mess with each other.

Tom: Right. The authors are showing you a two-step process, first segmenting the data using some matrix transformation, then modeling each segment separately using a dynamic structure.

Jane: So they’re essentially saying you don’t have to tackle the whole giant problem at once; you can handle pieces individually and stitch the results back together.

Tom: Exactly. The main implication here is that for functional data with way too many series compared to the time span, this structured approach beats existing methods in terms of prediction errors on real datasets like energy consumption or mortality.

Jane: It means that if you've got huge amounts of complex time series, you can use this technique to get much more accurate forecasts than just running standard, single-model approaches.

Tom: I’m looking at the authors’ theoretical work too—they proved some consistency results for the segmentation step and uniform convergence rates for how they estimate the dynamic structure.

Jane: That means it isn't just a clever trick; there are mathematical proofs showing that this decomposition method is actually reliable under certain conditions.

Tom: But there’s still that caveat, as they state, it relies on assuming the underlying dynamics are linear and stationary within those groups.

Jane: So, for folks who only listen to the show, this is about getting a robust way to make sense of really big data sets by imposing a mathematical structure onto the chaos.

Tom: If you want to see how they actually did this segmentation using eigenanalysis on the autocovariance functions, that’s what we're going to look at next.

Joint Laboratory of Data Science and Business Intelligence, Southwestern University of Finance and Economics · Academy of Mathematics and Systems Science, Chinese Academy of Sciences · Business School, University of Sydney · Faculty of Business and Economics, The University of Hong Kong · Department of Statistics, London School of Economics

stat.ME, econ.EM, math.ST, stat.TH

Submitted: 2024-06-02

Updated: 2026-11-12

Comments: 35 pages, 1 figure, 7 tables

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 91/100

The gist: The paper proposes a two-step procedure for modeling and predicting high-dimensional functional time series where p, the number of function-valued time series, is large relative to n…

Key concepts

Segmentation Transformation
This step applies a linear transformation ($Y_{tp} = A Z_{tp}$) to group the original functional time series into $q$ distinct subseries. The goal is to ensure that series from different groups are uncorrelated both at the same time point and across time lags, simplifying subsequent modeling.
Eigenanalysis of Quadratic Forms
The transformation matrix $A$ is estimated by analyzing a positive definite matrix derived from the double integral of quadratic forms in the autocovariance functions. This mathematical technique identifies patterns in how different functional series correlate over time, guiding the segmentation process.
Latent Decomposition ($Z_{plq} = X_{plq} oldsymbol{ ho}_{plq}$)
Each transformed subseries is modeled as a product of a dynamic structure ($X_{plq}$) and white noise ($oldsymbol{ ho}_{plq}$). This decomposition separates the underlying, finite-dimensional dynamics from random fluctuations, allowing for independent modeling of each group.
Variational Multivariate FPCA
This technique is used to find the dynamic structure ($X_{plq}$) for each subseries. It is a dimension reduction method that captures the most important latent components of the series' dynamics while ensuring the resulting structure has specific properties related to time lags.

Terminology

Summary

The paper proposes a two-step procedure for modeling and predicting high-dimensional functional time series where p, the number of function-valued time series, is large relative to n, the length of the time series. p. 0

Where the paper says it

On the Modelling and Prediction of High-Dimensional Functional Time Series Jinyuan Chang1,2, Qin Fang3 , Xinghao Qiao4 , and Qiwei Yao5 1Joint Laboratory of Data Science and Business Intelligence, Southwestern University of Finance and Economics, Chengdu, Sichuan, China 2Academy of Mathematics and Systems Science, Chinese Academy of Sciences, Beijing, China 3Business School, University of Sydney, Sydney, Australia 4Faculty of Business and Economics, The University of Hong Kong, Hong Kong 5Department of Statistics, London School of Economics, London, U. Abstract We propose a two-step procedure to model and predict high-dimensional functional time series, where the number of function-valued time series p is large in relation to the length of time series n.

The first step involves a segmentation transformation, seeking a linear transformation Y tp = A Z tp, where A is an invertible constant matrix. p. 1

Where the paper says it

To achieve more effective dimension reduction and better predictive performance, we propose in this paper a two-step approach. Our first step is a segmentation transformation step in which we seek for a linear transformation Ytp¨q “ AZtp¨q, where A is a p ˆ p invertible constant matrix, such that the transformed series Ztp¨q “ tZ p1q t p¨qJ , .

This transformation aims to segment the original high-dimensional functional time series into q groups such that any two subseries from different groups are uncorrelated both contemporaneously and serially. p. 0
Where the paper says it

Abstract We propose a two-step procedure to model and predict high-dimensional functional time series, where the number of function-valued time series p is large in relation to the length of time series n. Our first step performs an eigenanalysis of a positive definite matrix, which leads to a one-to-one linear transformation for the original highdimensional functional time series, and the transformed curve series can be segmented into several groups such that any two subseries from any two different groups are uncorrelated both contemporaneously and serially.

The estimation of the transformation matrix A is achieved through eigenanalysis of a positive definite matrix derived from the double integral of quadratic forms in the autocovariance functions of Y tp. p. 2
Where the paper says it

Hence each Z piq t can be modelled and forecasted separately as far as the linear dynamics is concerned. Under the stationarity assumption, the estimation of the transformation matrix A boils down to the eigenanalysis of a positive definite matrix defined by the double integral of quadratic forms in the autocovariance functions of Ytp¨q.

This segmentation effectively transforms serial correlations across different series into autocorrelations within each identified subseries, allowing each group to be modeled separately regarding linear dynamics. p. 3
Where the paper says it

Our paper makes useful contributions on multiple fronts. Firstly, the segmentation transformation in the first step transforms the serial correlations across different series into the autocorrelations within each of the identified q subseries.

The second step focuses on establishing a finite-dimensional dynamical structure for each transformed subseries Z plq independently. p. 0

Where the paper says it

Consequently in our second step those groups are handled separately without the information loss on the overall linear dynamic structure. The second step is devoted to establishing a finite-dimensional dynamical structure for all the transformed functional time series within each group.

This structure is characterized by a latent decomposition Z plq = X plq epsilon plq, where X plq represents the dynamics, epsilon plq is white noise with zero mean and zero covariance for any vectors and time lags, and tX plq is uncorrelated with t epsilon l for t at least 1. p. 2
Where the paper says it

Our second step is to identify a finite-dimensional dynamic structure for each transformed subseries Z plq t p¨q separately, which is based on a latent decomposition Z plq t puq “ X plq t puq ` ε plq t puq, u P U , (1) where X plq t p¨q represents the dynamics of Z plq t p¨q, ε plq t p¨q is white noise with Etε plq t puqu “ 0 and Etε plq t puqε plq s pvq J u “ 0 for any pu, vq P U 2 and t ‰ s, and tX plq t p¨qun t“1 are uncorrelated with tε plq t p¨qun t“1 . Furthermore we assume that the dynamic structure of X plq t p¨q admits a vector time series presentation via a variational multivariate FPCA.

The paper assumes that the dynamic structure of X plq can be represented as a vector time series via a variational multivariate FPCA. p. 2
Where the paper says it

Our second step is to identify a finite-dimensional dynamic structure for each transformed subseries Z plq t p¨q separately, which is based on a latent decomposition Z plq t puq “ X plq t puq ` ε plq t puq, u P U , (1) where X plq t p¨q represents the dynamics of Z plq t p¨q, ε plq t p¨q is white noise with Etε plq t puqu “ 0 and Etε plq t puqε plq s pvq J u “ 0 for any pu, vq P U 2 and t ‰ s, and tX plq t p¨qun t“1 are uncorrelated with tε plq t p¨qun t“1 . Furthermore we assume that the dynamic structure of X plq t p¨q admits a vector time series presentation via a variational multivariate FPCA.

The overall modeling and forecasting are realized by applying the dynamics learned for each group separately. p. 0

Where the paper says it

Consequently in our second step those groups are handled separately without the information loss on the overall linear dynamic structure. The second step is devoted to establishing a finite-dimensional dynamical structure for all the transformed functional time series within each group.

Specifically, for each subseries Z plq, the best VAR model is selected based on the AIC criterion to achieve an h-step ahead prediction, denoted as Z hat plq n. p. 2
Where the paper says it

An additional permutation on the components of Ztp¨q will be specified in order to identify the latent group structure. Our second step is to identify a finite-dimensional dynamic structure for each transformed subseries Z plq t p¨q separately, which is based on a latent decomposition Z plq t puq “ X plq t puq ` ε plq t puq, u P U , (1) where X plq t p¨q represents the dynamics of Z plq t p¨q, ε plq t p¨q is white noise with Etε plq t puqu “ 0 and Etε plq t puqε plq s pvq J u “ 0 for any pu, vq P U 2 and t ‰ s, and tX plq t p¨qun t“1 are uncorrelated with tε plq t p¨qun t“1 .

The final prediction for the original functional time series Y npq is then obtained by transforming this result back using the estimated transformation matrix A and the predicted vector time series Z hat npq. p. 0
Where the paper says it

Abstract We propose a two-step procedure to model and predict high-dimensional functional time series, where the number of function-valued time series p is large in relation to the length of time series n. Our first step performs an eigenanalysis of a positive definite matrix, which leads to a one-to-one linear transformation for the original highdimensional functional time series, and the transformed curve series can be segmented into several groups such that any two subseries from any two different groups are uncorrelated both contemporaneously and serially.

The paper makes several key contributions. p. 3

Where the paper says it

Our paper makes useful contributions on multiple fronts. Firstly, the segmentation transformation in the first step transforms the serial correlations across different series into the autocorrelations within each of the identified q subseries.

Firstly, the segmentation transformation addresses the decorrelation idea from Chang et al. p. 3
Where the paper says it

Therefore, the proposed transformation can always be used as an initial step in modelling high-dimensional functional time series. Secondly, though the segmentation transformation is motivated from the decorrelation idea of Chang et al.

(2018) for vector time series but adapts it to functional data, introducing complexities due to infinite dimensionality. p. 3
Where the paper says it

Secondly, though the segmentation transformation is motivated from the decorrelation idea of Chang et al. (2018) for vector time series, its adaption to the functional setting introduces additional methodological and theoretical complexities and requires innovative advancements in both methodology and theory due to the intrinsic infinite-dimensionality of functional data.

Secondly, the second step utilizes a nonzero lagged autocovariance-based dimension reduction approach to effectively incorporate serial dependence information relevant for prediction. p. 4
Where the paper says it

Its theoretical analysis involves establishing novel inequalities between functional versions of matrix norms. Thirdly, the nonzero lagged autocovariance-based dimension reduction approach in the second step makes the good use of the serial dependence information in our estimation, which is most relevant in time series prediction.

The method is demonstrated to significantly outperform existing methods (SegV significantly outperforms UniV and GSY) across various settings, including age-specific mortality and energy consumption data, achieving the lowest prediction errors in simulations and on real datasets. p. 1
Where the paper says it

1 Introduction Functional time series typically refers to continuous-time records that are naturally divided into consecutive time intervals, such as days, months or years. With recent advances in data collection technology, multivariate or even high-dimensional functional time series arise ubiquitously in many applications, including daily pollution concentration curves over different locations, annual temperature curves at different stations, annual age-specific mortality rates for different countries, and intraday energy consumption trajectories from different households.

Theoretically, the analysis establishes group recovery consistency for the segmentation step (Theorem 1), uniform convergence rates for dimension reduction estimation (Theorem 2), identification of the dynamic space spanned by eigenfunctions (Theorem 3), and uniform convergence rates for this identification (Theorem 4). p. 4
Where the paper says it

Importantly, when p is fixed or moderately large, such method can be directly applied to the observed curve series Ytp¨q for dimension reduction and forecasting purposes. On the theory side, we demonstrate that our proposal exhibits appealing convergence properties despite the additional transformation and estimation errors arisen from the first step, which are not involved in Bathia et al.

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

  1. Enhanced dimensionality reduction via two-step procedure: The system can perform an eigenanalysis for segmentation transformation to achieve a one-to-one linear transformation of high-dimensional functional time series, followed by a finite-dimensional dynamic structure identification using variational multivariate FPCA on each segmented group.

  2. Improved group structure recovery: The system can estimate the number of groups q through an iterative process based on cross-autocovariance maxima and then identify the true segmentation by splitting graphs into isolated subgraphs, ensuring group recovery consistency.

  3. Superior predictive performance: The system can achieve better forecasting accuracy by utilizing the transformed series Ztpq, as demonstrated in Example 1 where SegV and FTSegV outperform direct forecasting methods, even under high-dimensional settings.

  4. Robustness to high dimensionality: The system can maintain good performance when p is large by employing functional thresholding estimation of autocovariance functions, which is shown to be more effective than classical estimation methods like NonT in large p scenarios.

  5. Consistent subspace identification: The system can reliably identify the dynamic space for each transformed subseries Zplq by using variational multivariate FPCA based on a variational Karhunen-Loève decomposition, ensuring the identified basis functions capture the most significant serial dependence.

Abstract

We propose a two-step procedure to model and predict high-dimensional functional time series, where the number of function-valued time series p is large in relation to the length of time series n. Our first step performs an eigenanalysis of a positive definite matrix, which leads to a one-to-one linear transformation for the original high-dimensional functional time series, and the transformed curve series can be segmented into several groups such that any two subseries from any two different groups are uncorrelated both contemporaneously and serially. Consequently in our second step those groups are handled separately without the information loss on the overall linear dynamic structure. The second step is devoted to establishing a finite-dimensional dynamical structure for all the transformed functional time series within each group. Furthermore the finite-dimensional structure is represented by that of a vector time series. Modelling and forecasting for the original high-dimensional functional time series are realized via those for the vector time series in all the groups. We investigate the theoretical properties of our proposed methods, and illustrate the finite-sample performance through both extensive simulation and two real datasets.

Related papers