Probabilistic Functional Neural Networks

Haixu Wang, Jiguo Cao · 2025

Read Probabilistic Functional Neural Networks by Haixu Wang, Jiguo Cao (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-21585.txt.

Introduction

Haixu Wang and Jiguo Cao propose in 2025 a probabilistic neural network for forecasting functional time series with many dimensions. Such data have nonlinear trends and high spatial size, which makes modeling hard. Standard tools struggle with nonstationarity and scale in this setting.

Their model, called ProFnet, mixes feedforward and deep networks with probability modeling. Forecasts are generated by sampling, which also gives uncertainty estimates for each prediction. It captures links across time and across regions in one framework. Applied to mortality rates in Japan, it beats rival methods and gives interpretable uncertainty.

Catalog record

CREATORHaixu Wang, Jiguo Cao
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE39 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2503-21585

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Probabilistic Functional Neural Networks Haixu Wang, Jiguo Cao Abstract High-dimensional functional time series (HDFTS) are often characterized by nonlinear trends and high spatial dimensions. Such data poses unique challenges for modeling and forecasting due to the nonlinearity, nonstationarity, and high dimensionality. We propose a novel probabilistic functional neural network ( ProFnet ) to address these challenges. ProFnet integrates the strengths of feedforward and deep neural networks with probabilistic modeling. The model generates probabilistic forecasts using Monte Carlo sampling and also enables the quantification of uncertainty in predictions. While capturing both temporal and spatial dependencies across multiple regions, ProFnet offers a scalable and unified solution for large datasets. Applications to Japan’s mortality rates demonstrate superior performance. This approach enhances predictive accuracy and provides interpretable uncertainty estimates, making it a valuable tool for forecasting complex high-dimensional functional data and HDFTS. Keywords: High-dimensional functional time series, Probabilistic neural network, Functional data analysis, Gaussian Processes 1 Introduction High-dimensional functional time series (HDFTS) have emerged as a unique data structure in fields such as econometrics, public health, and environmental science. This data type involves observations of functional data, like curves or surfaces, across multiple regions over time. Analyzing and forecasting HDFTS presents unique challenges due to its inherent nonlinearity, nonstationarity, and high dimensionality, where the number of spatial units often exceeds temporal observations. For instance, consider monitoring mortality rate curves across Japan’s 47 prefectures from 1973 to 2022 as in Figure 1 . Each prefecture contributes a time series of functional data, and collectively, these form a high-dimensional dataset. This emerging type of functional data urges a unified and flexible method that should be robust and scalable. Figure 1: An example of high-dimensional functional time series. The data is a set of mortality rate curves for 47 prefectures in Japan from year 1973 to 2022. The color represents the year of the mortality rate curve, from blue (oldest) to red (most current). We would simultaneously observe a functional time series within each prefecture, and the collection of all prefectures forms a high-dimensional functional time series. The plot showcases the functional time series for 2 randomly chosen prefectures. Analyzing high-dimensional functional time series (HDFTS) often extends from established methods for univariate functional time series Aue et al., (2015) ; Hyndman and Ullah, (2007) . Traditional analyses typically fall into two categories: functional principal component analysis (FPCA) and functional regression models. The first approach focus on predicting future functional data by obtaining their projections on the estimated functional principal components, e.g., Jiao et al., (2023) ; Horváth and Kokoszka, (2012) ; Rubin and Panaretos, (2020) ; Horvath et al., (2013) . This approach assumes the stationarity and linearity of the data and heavily depends on lag-specific forecasting. Functional regression models, on the other hand, adopt a regression-based framework in which past functional observations serve as covariates to predict future data. This methodology, similar to a functional counterpart of autoregressive models, has been applied in studies like Hyndman and Shang, (2009) and Kargin and Onatski, (2008) . A nonparametric model has been discussed in Elias et al., (2022) . Although these models address some of the shortcomings of FPCA, they also rely on the same rigid assumptions of stationarity and linearity, limiting their applicability in real-world scenarios characterized by complex and non-stationary trends. Conventional FPCA focuses on univariate functional data and struggles with multivariate or high-dimensional settings. A common workaround is to concatenate multiple dimensions of functional data into a single functional object and apply univariate FPCA techniques Ramsay and Silverman, (2005) . While effective to some extent, this approach is computationally expensive and overlooks the discontinuous connecting points between functional variables. Some alternative FPCA methods are developed for multivariate functional data Chiou et al., (2014) ; Happ and Greven, (2018) ; Berrendero et al., (2011) . In recent years, attention has turned to explicitly tackling HDFTS. Some important work has been developed by Gao et al., (2019) ; Li et al., (2024) ; Jiménez-Varón et al., (2024) . In addition, Zhou and Dette, (2023) explored statistical inference, Tang et al., (2022) investigated clustering, and Hallin et al., (2023) employed factor models for forecasting. To tack non-stationarity, Hörmann et al., (2015) and Elayouty et al., (2022) employed a dynamic version of FPCA for forecasting. Despite these advancements, many remain direct extensions of univariate methods and fail to fully address the limitations of stationarity, linearity, and lag dependence. An alternative approach to modeling and forecasting high-dimensional functional time series (HDFTS) is the application of neural networks. Among these, recurrent neural networks (RNNs) have gained popularity for their ability to model temporal dependencies in time series data. RNN-based strategies have been extensively reviewed and discussed, highlighting their strengths and limitations in various forecasting contexts Hewamalage et al., (2021) . Recent developments, such as the robust forecasting model proposed in Zhang et al., (2023) and the sequence-to-sequence framework introduced in Sutskever et al., (2014) , have enhanced the utility of RNNs in handling complex temporal structures. For functional data, RNNs have been adapted to forecast functional time series, as demonstrated in Wang and Cao, (2023) . These models are highly effective at capturing nonlinearity and temporal patterns but face several significant challenges. First, RNN-based methods are typically lag-dependent, requiring a fixed network architecture specifically designed for certain time lags. Additionally, they often encounter difficulties during training and incur high computational costs, making them less suitable for HDFTS. Finally, a critical limitation of RNNs in the context of HDFTS is their inability to quantify forecast uncertainty. In many applications, particularly in areas such as public health and finance, uncertainty quantification is as important as the point forecast itself for enabling informed decision-making and robust statistical inference. In summary, HDFTS poses a complex data structure that demands flexible and scalable models for effective analysis and forecasting. Existing methodologies suffer from limitations of linearity and stationarity assumptions, and lack of scalability or uncertainty quantification. In this work, we propose a probabilistic functional neural network ( ProFnet ) for analyzing and forecasting HDFTS. To the best of our knowledge, this is the first attempt to build a probabilistic model and address the uncertainty of functional data with neural networks. We outline the key contributions to the forecasting problem and advantages over traditional functional data analysis (FDA) methods and RNN-based models: Probabilistic forecasting. The foremost contribution of this work is to build a probabilistic model for HDFTS. In fact, it can be seen as a generative model quantifying the forecast uncertainty through Monte Carlo samples and consequently generated functional data. Then, we can quantify the uncertainty through the distribution of generated functional data. This capability is essential for applications where reliable predictions and uncertainty estimates can support more informed decision-making. Compared to the existing approaches, ProFnet simultaneously addresses the complexity and nonlinearity of functional data while combining the strengths of neural networks and statistical inference. Joint modeling of temporal and spatial dependencies. In this work, we will focus on the application with dimensionality corresponding to the spatial dimension. The proposed method simultaneously captures temporal and spatial dependencies across regions, modeling how functional data for one region is influenced by its own history and that of other regions. This simultaneous modeling of all regions offers a significant

Read the complete text