Transformer-based models have gained increasing attention in time series research, driving interest in Large Language Models (LLMs) and foundation models for time series analysis. As the field moves toward multi-modality, Large Vision Models (LVMs) are emerging as a promising direction. In the past, the effectiveness of Transformer and LLMs in time series has been debated. When it comes to LVMs, a similar question arises: are LVMs truely useful for time series analysis? To address it, we design and conduct the first principled study involving 4 LVMs, 8 imaging methods, 18 datasets and 26 baselines across both high-level (classification) and low-level (forecasting) tasks, with extensive ablation analysis. Our findings indicate LVMs are indeed useful for time series classification but face challenges in forecasting. Although effective, the contemporary best LVM forecasters are limited to specific types of LVMs and imaging methods, exhibit a bias toward forecasting periods, and have limited ability to utilize long look-back windows. We hope our findings could serve as a cornerstone for future research on LVM- and multimodal-based solutions to different time series tasks.
The paper introduces a novel conditional independence (CI) based method for linear and nonlinear, lagged and contemporaneous causal discovery from observational time series in the causally sufficient case. Existing CI-based methods such as the PC algorithm and also common methods from other frameworks suffer from low recall and partially inflated false positives for strong autocorrelation which is an ubiquitous challenge in time series. The novel method, PCMCI$^+$, extends PCMCI [Runge et al., 2019b] to include discovery of contemporaneous links. PCMCI$^+$ improves the reliability of CI tests by optimizing the choice of conditioning sets and even benefits from autocorrelation. The method is order-independent and consistent in the oracle case. A broad range of numerical experiments demonstrates that PCMCI$^+$ has higher adjacency detection power and especially more contemporaneous orientation recall compared to other methods while better controlling false positives. Optimized conditioning sets also lead to much shorter runtimes than the PC algorithm. PCMCI$^+$ can be of considerable use in many real world application scenarios where often time resolutions are too coarse to resolve time delays and strong autocorrelation is present.
Recent years have witnessed a growing interest for time series foundation models, with a strong emphasis on the forecasting task. Yet, the crucial task of out-of-domain imputation of missing values remains largely underexplored. We propose a first step to fill this gap by leveraging implicit neural representations (INRs). INRs model time series as continuous functions and naturally handle various missing data scenarios and sampling rates. While they have shown strong performance within specific distributions, they struggle under distribution shifts. To address this, we introduce MoTM (Mixture of Timeflow Models), a step toward a foundation model for time series imputation. Building on the idea that a new time series is a mixture of previously seen patterns, MoTM combines a basis of INRs, each trained independently on a distinct family of time series, with a ridge regressor that adapts to the observed context at inference. We demonstrate robust in-domain and out-of-domain generalization across diverse imputation scenarios (e.g., block and pointwise missingness, variable sampling rates), paving the way for adaptable foundation imputation models.
Transformer-based models have recently become dominant in Long-term Time Series Forecasting (LTSF), yet the variations in their architecture, such as encoder-only, encoder-decoder, and decoder-only designs, raise a crucial question: What Transformer architecture works best for LTSF tasks? However, existing models are often tightly coupled with various time-series-specific designs, making it difficult to isolate the impact of the architecture itself. To address this, we propose a novel taxonomy that disentangles these designs, enabling clearer and more unified comparisons of Transformer architectures. Our taxonomy considers key aspects such as attention mechanisms, forecasting aggregations, forecasting paradigms, and normalization layers. Through extensive experiments, we uncover several key insights: bi-directional attention with joint-attention is most effective; more complete forecasting aggregation improves performance; and the direct-mapping paradigm outperforms autoregressive approaches. Furthermore, our combined model, utilizing optimal architectural choices, consistently outperforms several existing models, reinforcing the validity of our conclusions. We hope these findings offer valuable guidance for future research on Transformer architectural designs in LTSF. Our code is available at https://github.com/HALF111/TSF_architecture.
Diffusion models, a type of generative model, have shown promise in time series forecasting. But they face limitations like rigid source distributions and limited sampling paths, which hinder their performance. Flow matching offers faster generation, higher-quality outputs, and greater flexibility, while also possessing the ability to utilize valuable information from the prediction errors of prior models, which were previously inaccessible yet critically important. To address these challenges and fully unlock the untapped potential of flow matching, we propose Conditional Guided Flow Matching (CGFM). CGFM extends flow matching by incorporating the outputs of an auxiliary model, enabling a previously unattainable capability in the field: learning from the errors of the auxiliary model. For time series forecasting tasks, it integrates historical data as conditions and guidance, constructs two-sided conditional probability paths, and uses a general affine path to expand the space of probability paths, ultimately leading to improved predictions. Extensive experiments show that CGFM consistently enhances and outperforms state-of-the-art models, highlighting its effectiveness in advancing forecasting methods.
Real-world time series are often governed by complex nonlinear dynamics. Understanding these underlying dynamics is crucial for precise future prediction. While deep learning has achieved major success in time series forecasting, many existing approaches do not explicitly model the dynamics. To bridge this gap, we introduce DeepEDM, a framework that integrates nonlinear dynamical systems modeling with deep neural networks. Inspired by empirical dynamic modeling (EDM) and rooted in Takens' theorem, DeepEDM presents a novel deep model that learns a latent space from time-delayed embeddings, and employs kernel regression to approximate the underlying dynamics, while leveraging efficient implementation of softmax attention and allowing for accurate prediction of future time steps. To evaluate our method, we conduct comprehensive experiments on synthetic data of nonlinear dynamical systems as well as real-world time series across domains. Our results show that DeepEDM is robust to input noise, and outperforms state-of-the-art methods in forecasting accuracy. Our code is available at: https://abrarmajeedi.github.io/deep_edm.
With advancements in computing and communication technologies, the Internet of Things (IoT) has seen significant growth. IoT devices typically collect data from various sensors, such as temperature, humidity, and energy meters. Much of this data is temporal in nature. Traditionally, data from IoT devices is centralized for analysis, but this approach introduces delays and increased communication costs. Federated learning (FL) has emerged as an effective alternative, allowing for model training across distributed devices without the need to centralize data. In many applications, such as smart home energy and environmental monitoring, the data collected by IoT devices across different locations can exhibit significant variation in trends and seasonal patterns. Accurately forecasting such non-stationary, non-linear time-series data is crucial for applications like energy consumption estimation and weather forecasting. However, these data variations can severely impact prediction accuracy. The key contributions of this paper are: (1) Investigating how non-linear, non-stationary time-series data distributions, like generalized extreme value (gen-extreme) and log norm distributions, affect FL performance. (2) Analyzing how different detrending techniques for non-linear time-series data influence the forecasting model's performance in a FL setup. We generated several synthetic time-series datasets using non-linear data distributions and trained an LSTM-based forecasting model using both centralized and FL approaches. Additionally, we evaluated the impact of detrending on real-world datasets with non-linear time-series data distributions. Our experimental results show that: (1) FL performs worse than centralized approaches when dealing with non-linear data distributions. (2) The use of appropriate detrending techniques improves FL performance, reducing loss across different data distributions.
The zero-shot capabilities of foundation models (FMs) for time series forecasting offer promising potentials in conformal prediction, as most of the available data can be allocated to calibration. This study compares the performance of Time Series Foundation Models (TSFMs) with traditional methods, including statistical models and gradient boosting, within a conformal prediction setting. Our findings highlight two key advantages of TSFMs. First, when the volume of data is limited, TSFMs provide more reliable conformalized prediction intervals than classic models, thanks to their superior predictive accuracy. Second, the calibration process is more stable because more data are used for calibration. Morever, the fewer data available, the more pronounced these benefits become, as classic models require a substantial amount of data for effective training. These results underscore the potential of foundation models in improving conformal prediction reliability in time series applications, particularly in data-constrained cases. All the code to reproduce the experiments is available.
Currently, iTransformer is one of the most popular and effective models for multivariate time series (MTS) forecasting. Thanks to its inverted framework, iTransformer effectively captures multivariate correlation. However, the inverted framework still has some limitations. It diminishes temporal interdependency information, and introduces noise in cases of nonsignificant variable correlation. To address these limitations, we introduce a novel data augmentation method on inverted framework, called DAIF. Unlike previous data augmentation methods, DAIF stands out as the first real-time augmentation specifically designed for the inverted framework in MTS forecasting. We first define the structure of the inverted sequence-to-sequence framework, then propose two different DAIF strategies, Frequency Filtering and Cross-variation Patching to address the existing challenges of the inverted framework. Experiments across multiple datasets and inverted models have demonstrated the effectiveness of our DAIF.
In this paper, we investigate the distillation of time series reasoning capabilities into small, instruction-tuned language models as a step toward building interpretable time series foundation models. Leveraging a synthetic dataset of mean-reverting time series with systematically varied trends and noise levels, we generate natural language annotations using a large multimodal model and use these to supervise the fine-tuning of compact Qwen models. We introduce evaluation metrics that assess the quality of the distilled reasoning - focusing on trend direction, noise intensity, and extremum localization - and show that the post-trained models acquire meaningful interpretive capabilities. Our results highlight the feasibility of compressing time series understanding into lightweight, language-capable models suitable for on-device or privacy-sensitive deployment. This work contributes a concrete foundation toward developing small, interpretable models that explain temporal patterns in natural language.
With the advent of Transformers, time series forecasting has seen significant advances, yet it remains challenging due to the need for effective sequence representation, memory construction, and accurate target projection. Time series forecasting remains a challenging task, demanding effective sequence representation, meaningful information extraction, and precise future projection. Each dataset and forecasting configuration constitutes a distinct task, each posing unique challenges the model must overcome to produce accurate predictions. To systematically address these task-specific difficulties, this work decomposes the time series forecasting pipeline into three core stages: input sequence representation, information extraction and memory construction, and final target projection. Within each stage, we investigate a range of architectural configurations to assess the effectiveness of various modules, such as convolutional layers for feature extraction and self-attention mechanisms for information extraction, across diverse forecasting tasks, including evaluations on seven benchmark datasets. Our models achieve state-of-the-art forecasting accuracy while greatly enhancing computational efficiency, with reduced training and inference times and a lower parameter count. The source code is available at https://github.com/RobertLeppich/REP-Net.
Many dynamic processes such as telecommunication and transport networks can be described through discrete time series of graphs. Modelling the dynamics of such time series enables prediction of graph structure at future time steps, which can be used in applications such as detection of anomalies. Existing approaches for graph prediction have limitations such as assuming that the vertices do not to change between consecutive graphs. To address this, we propose to exploit time series prediction methods in combination with an adapted form of flux balance analysis (FBA), a linear programming method originating from biochemistry. FBA is adapted to incorporate various constraints applicable to the scenario of growing graphs. Empirical evaluations on synthetic datasets (constructed via Preferential Attachment model) and real datasets (UCI Message, HePH, Facebook, Bitcoin) demonstrate the efficacy of the proposed approach.
Time series with missing or irregularly sampled data are a persistent challenge in machine learning. Many methods operate on the frequency-domain, relying on the Fast Fourier Transform (FFT) which assumes uniform sampling, therefore requiring prior interpolation that can distort the spectra. To address this limitation, we introduce a differentiable Lomb--Scargle layer that enables a reliable computation of the power spectrum of irregularly sampled data. We integrate this layer into a novel score-based diffusion model (LSCD) for time series imputation conditioned on the entire signal spectrum. Experiments on synthetic and real-world benchmarks demonstrate that our method recovers missing data more accurately than purely time-domain baselines, while simultaneously producing consistent frequency estimates. Crucially, our method can be easily integrated into learning frameworks, enabling broader adoption of spectral guidance in machine learning approaches involving incomplete or irregular data.
Large Language Models (LLMs) have shown remarkable performance across diverse tasks without domain-specific training, fueling interest in their potential for time-series forecasting. While LLMs have shown potential in zero-shot forecasting through prompting alone, recent studies suggest that LLMs lack inherent effectiveness in forecasting. Given these conflicting findings, a rigorous validation is essential for drawing reliable conclusions. In this paper, we evaluate the effectiveness of LLMs as zero-shot forecasters compared to state-of-the-art domain-specific models. Our experiments show that LLM-based zero-shot forecasters often struggle to achieve high accuracy due to their sensitivity to noise, underperforming even simple domain-specific models. We have explored solutions to reduce LLMs' sensitivity to noise in the zero-shot setting, but improving their robustness remains a significant challenge. Our findings suggest that rather than emphasizing zero-shot forecasting, a more promising direction would be to focus on fine-tuning LLMs to better process numerical sequences. Our experimental code is available at https://github.com/junwoopark92/revisiting-LLMs-zeroshot-forecaster.
As a prominent data modality task, time series forecasting plays a pivotal role in diverse applications. With the remarkable advancements in Large Language Models (LLMs), the adoption of LLMs as the foundational architecture for time series modeling has gained significant attention. Although existing models achieve some success, they rarely both model time and frequency characteristics in a pretraining-finetuning paradigm leading to suboptimal performance in predictions of complex time series, which requires both modeling periodicity and prior pattern knowledge of signals. We propose MoFE-Time, an innovative time series forecasting model that integrates time and frequency domain features within a Mixture of Experts (MoE) network. Moreover, we use the pretraining-finetuning paradigm as our training framework to effectively transfer prior pattern knowledge across pretraining and finetuning datasets with different periodicity distributions. Our method introduces both frequency and time cells as experts after attention modules and leverages the MoE routing mechanism to construct multidimensional sparse representations of input signals. In experiments on six public benchmarks, MoFE-Time has achieved new state-of-the-art performance, reducing MSE and MAE by 6.95% and 6.02% compared to the representative methods Time-MoE. Beyond the existing evaluation benchmarks, we have developed a proprietary dataset, NEV-sales, derived from real-world business scenarios. Our method achieves outstanding results on this dataset, underscoring the effectiveness of the MoFE-Time model in practical commercial applications.
Portfolio optimization is a fundamental challenge in quantitative finance, requiring robust computational tools that integrate statistical rigor with practical implementation. We present skfolio, an open-source Python library for portfolio construction and risk management that seamlessly integrates with the scikit-learn ecosystem. skfolio provides a unified framework for diverse allocation strategies, from classical mean-variance optimization to modern clustering-based methods, state-of-the-art financial estimators with native interfaces, and advanced cross-validation techniques tailored for financial time series. By adhering to scikit-learn's fit-predict-transform paradigm, the library enables researchers and practitioners to leverage machine learning workflows for portfolio optimization, promoting reproducibility and transparency in quantitative finance.
Multi-task and few-shot time series forecasting tasks are commonly encountered in scenarios such as the launch of new products in different cities. However, traditional time series forecasting methods suffer from insufficient historical data, which stems from a disregard for the generalized and specific features among different tasks. For the aforementioned challenges, we propose the Feature-Adaptive Time Series Forecasting Framework (FAF), which consists of three key components: the Generalized Knowledge Module (GKM), the Task-Specific Module (TSM), and the Rank Module (RM). During training phase, the GKM is updated through a meta-learning mechanism that enables the model to extract generalized features across related tasks. Meanwhile, the TSM is trained to capture diverse local dynamics through multiple functional regions, each of which learns specific features from individual tasks. During testing phase, the RM dynamically selects the most relevant functional region from the TSM based on input sequence features, which is then combined with the generalized knowledge learned by the GKM to generate accurate forecasts. This design enables FAF to achieve robust and personalized forecasting even with sparse historical observations We evaluate FAF on five diverse real-world datasets under few-shot time series forecasting settings. Experimental results demonstrate that FAF consistently outperforms baselines that include three categories of time series forecasting methods. In particular, FAF achieves a 41.81\% improvement over the best baseline, iTransformer, on the CO$_2$ emissions dataset.
Multivariate time series anomaly detection (MTS-AD) is critical in domains like healthcare, cybersecurity, and industrial monitoring, yet remains challenging due to complex inter-variable dependencies, temporal dynamics, and sparse anomaly labels. We introduce mTSBench, the largest benchmark to date for MTS-AD and unsupervised model selection, spanning 344 labeled time series across 19 datasets and 12 diverse application domains. mTSBench evaluates 24 anomaly detection methods, including large language model (LLM)-based detectors for multivariate time series, and systematically benchmarks unsupervised model selection techniques under standardized conditions. Consistent with prior findings, our results confirm that no single detector excels across datasets, underscoring the importance of model selection. However, even state-of-the-art selection methods remain far from optimal, revealing critical gaps. mTSBench provides a unified evaluation suite to enable rigorous, reproducible comparisons and catalyze future advances in adaptive anomaly detection and robust model selection.
Fault diagnosis in Cyber-Physical Systems (CPSs) is essential for ensuring system dependability and operational efficiency by accurately detecting anomalies and identifying their root causes. However, the manual modeling of faulty behaviors often demands extensive domain expertise and produces models that are complex, error-prone, and difficult to interpret. To address this challenge, we present a novel unsupervised fault diagnosis methodology that integrates collective anomaly detection in multivariate time series, process mining, and stochastic simulation. Initially, collective anomalies are detected from low-level sensor data using multivariate time-series analysis. These anomalies are then transformed into structured event logs, enabling the discovery of interpretable process models through process mining. By incorporating timing distributions into the extracted Petri nets, the approach supports stochastic simulation of faulty behaviors, thereby enhancing root cause analysis and behavioral understanding. The methodology is validated using the Robotic Arm Dataset (RoAD), a widely recognized benchmark in smart manufacturing. Experimental results demonstrate its effectiveness in modeling, simulating, and classifying faulty behaviors in CPSs. This enables the creation of comprehensive fault dictionaries that support predictive maintenance and the development of digital twins for industrial environments.
Federated Learning (FL) has emerged as a transformative paradigm for distributed machine learning while preserving data privacy. However, existing approaches predominantly focus on model heterogeneity and aggregation techniques, largely overlooking the fundamental impact of dataset size characteristics on federated training dynamics. This paper introduces Size-Based Adaptive Federated Learning (SAFL), a novel progressive training framework that systematically organizes federated learning based on dataset size characteristics across heterogeneous multi-modal data. Our comprehensive experimental evaluation across 13 diverse datasets spanning 7 modalities (vision, text, time series, audio, sensor, medical vision, and multimodal) reveals critical insights: 1) an optimal dataset size range of 1000-1500 samples for federated learning effectiveness; 2) a clear modality performance hierarchy with structured data (time series, sensor) significantly outperforming unstructured data (text, multimodal); and 3) systematic performance degradation for large datasets exceeding 2000 samples. SAFL achieves an average accuracy of 87.68% across all datasets, with structured data modalities reaching 99%+ accuracy. The framework demonstrates superior communication efficiency, reducing total data transfer to 7.38 GB across 558 communications while maintaining high performance. Our real-time monitoring framework provides unprecedented insights into system resource utilization, network efficiency, and training dynamics. This work fills critical gaps in understanding how data characteristics should drive federated learning strategies, providing both theoretical insights and practical guidance for real-world FL deployments in neural network and learning systems.