Abstract:Modern time series forecasting methods, such as Transformer and its variants, have shown strong ability in sequential data modeling. To achieve high performance, they usually rely on redundant or unexplainable structures to model complex relations between variables and tune the parameters with large-scale data. Many real-world data mining tasks, however, lack sufficient variables for relation reasoning, and therefore these methods may not properly handle such forecasting problems. With insufficient data, time series appear to be affected by many exogenous variables, and thus, the modeling becomes unstable and unpredictable. To tackle this critical issue, in this paper, we develop a novel algorithmic framework for inferring the intrinsic latent factors implied by the observable time series. The inferred factors are used to form multiple independent and predictable signal components that enable not only sparse relation reasoning for long-term efficiency but also reconstructing the future temporal data for accurate prediction. To achieve this, we introduce three characteristics, i.e., predictability, sufficiency, and identifiability, and model these characteristics via the powerful deep latent dynamics models to infer the predictable signal components. Empirical results on multiple real datasets show the efficiency of our method for different kinds of time series forecasting. The statistical analysis validates the predictability of the learned latent factors.

What problem does this paper attempt to address?

The paper aims to address key challenges in time series forecasting, especially when dealing with real-world datasets that may contain limited observables and be influenced by many unknown factors. Existing modern time series modeling methods (such as Transformer and its variants) perform well when handling domain-specific sequence data (such as language and vision), but they may struggle to identify meaningful patterns and dependencies in data with complex spatiotemporal characteristics, leading to unstable and difficult-to-predict models. To address the above issues, the authors propose a novel algorithmic framework for inferring latent factors from observed time series data. These inferred factors are used to form multiple predictable and independent signal components, which not only accurately predict future time series but also support sparse relationship inference to improve long-term efficiency. Specifically, the authors introduce three key characteristics—predictability, sufficiency, and identifiability—and model these characteristics of the latent factors through a powerful deep latent dynamic model, thereby inferring predictable signal components. The main contributions of the paper are as follows: 1. Proposing a method to infer time-varying latent factors from observed data, forming multiple signal components with predictability, sufficiency, and identifiability, which can be used to construct future signals for time series forecasting. 2. Designing a novel approach to decouple latent factors and model the relationships between factors based on signal components, thereby mapping the joint distribution for more efficient approximation. 3. Achieving better performance compared to existing techniques in long-term series forecasting and stock trend prediction tasks. In short, the core of this paper lies in developing a method that can automatically discover easily predictable latent factors from complex time series data, to better understand and predict future data.

Discovering Predictable Latent Factors for Time Series Forecasting

Foreformer: an Enhanced Transformer-Based Framework for Multivariate Time Series Forecasting

Hidformer: Hierarchical Dual-Tower Transformer Using Multi-Scale Mergence for Long-Term Time Series Forecasting

Enhancing Time Series Forecasting: A Hierarchical Transformer with Probabilistic Decomposition Representation

Probabilistic Decomposition Transformer for Time Series Forecasting

Generating Time Series by Using Latent Space

Are Transformers Effective for Time Series Forecasting?

Itransformer: Inverted Transformers Are Effective for Time Series Forecasting

Two Steps Forward and One Behind: Rethinking Time Series Forecasting with Deep Learning

A hybrid framework for multivariate long-sequence time series forecasting

SageFormer: Series-Aware Framework for Long-term Multivariate Time Series Forecasting

Towards Long-Term Time-Series Forecasting: Feature, Pattern, and Distribution

Adapt to Small-Scale and Long-Term Time Series Forecasting with Enhanced Multidimensional Correlation

Sufficient Forecasting Using Factor Models

Latent Diffusion Transformer for Probabilistic Time Series Forecasting

Adaptive Multi-Scale Decomposition Framework for Time Series Forecasting

DRFormer: Multi-Scale Transformer Utilizing Diverse Receptive Fields for Long Time-Series Forecasting

Predicting Time Series from Short-Term High-Dimensional Data.

Leveraging 2D Information for Long-term Time Series Forecasting with Vanilla Transformers