The long horizon forecasting (LHF) problem has come up in the time series literature for over the last 35 years or so. This review covers aspects of LHF in this period and how deep learning has incorporated variants of trend, seasonality, fourier and wavelet transforms, misspecification bias reduction and bandpass filters while contributing using convolutions, residual connections, sparsity reduction, strided convolutions, attention masks, SSMs, normalization methods, low-rank approximations and gating mechanisms. We highlight time series decomposition techniques, input data preprocessing and dataset windowing schemes that improve performance. Multi-layer perceptron models, recurrent neural network hybrids, self-attention models that improve and/or address the performances of the LHF problem are described, with an emphasis on the feature space construction. Ablation studies are conducted over the ETTm2 dataset in the multivariate and univariate high useful load (HUFL) forecasting contexts, evaluated over the last 4 months of the dataset. The heatmaps of MSE averages per time step over test set series in the horizon show that there is a steady increase in the error proportionate to its length except with xLSTM and Triformer models and motivate LHF as an error propagation problem. The trained models are available here: https://bit.ly/LHFModelZoo
QCAAPatchTF is a quantum attention network integrated with an advanced patch-based transformer, designed for multivariate time series forecasting, classification, and anomaly detection. Leveraging quantum superpositions, entanglement, and variational quantum eigensolver principles, the model introduces a quantum-classical hybrid self-attention mechanism to capture multivariate correlations across time points. For multivariate long-term time series, the quantum self-attention mechanism can reduce computational complexity while maintaining temporal relationships. It then applies the quantum-classical hybrid self-attention mechanism alongside a feed-forward network in the encoder stage of the advanced patch-based transformer. While the feed-forward network learns nonlinear representations for each variable frame, the quantum self-attention mechanism processes individual series to enhance multivariate relationships. The advanced patch-based transformer computes the optimized patch length by dividing the sequence length into a fixed number of patches instead of using an arbitrary set of values. The stride is then set to half of the patch length to ensure efficient overlapping representations while maintaining temporal continuity. QCAAPatchTF achieves state-of-the-art performance in both long-term and short-term forecasting, classification, and anomaly detection tasks, demonstrating state-of-the-art accuracy and efficiency on complex real-world datasets.
We introduce a temporal feature encoding architecture called Time Series Representation Model (TSRM) for multivariate time series forecasting and imputation. The architecture is structured around CNN-based representation layers, each dedicated to an independent representation learning task and designed to capture diverse temporal patterns, followed by an attention-based feature extraction layer and a merge layer, designed to aggregate extracted features. The architecture is fundamentally based on a configuration that is inspired by a Transformer encoder, with self-attention mechanisms at its core. The TSRM architecture outperforms state-of-the-art approaches on most of the seven established benchmark datasets considered in our empirical evaluation for both forecasting and imputation tasks. At the same time, it significantly reduces complexity in the form of learnable parameters. The source code is available at https://github.com/RobertLeppich/TSRM.
We introduce Toto, a time series forecasting foundation model with 151 million parameters. Toto uses a modern decoder-only architecture coupled with architectural innovations designed to account for specific challenges found in multivariate observability time series data. Toto's pre-training corpus is a mixture of observability data, open datasets, and synthetic data, and is 4-10$\times$ larger than those of leading time series foundation models. Additionally, we introduce BOOM, a large-scale benchmark consisting of 350 million observations across 2,807 real-world time series. For both Toto and BOOM, we source observability data exclusively from Datadog's own telemetry and internal observability metrics. Extensive evaluations demonstrate that Toto achieves state-of-the-art performance on both BOOM and on established general purpose time series forecasting benchmarks. Toto's model weights, inference code, and evaluation scripts, as well as BOOM's data and evaluation code, are all available as open source under the Apache 2.0 License available at https://huggingface.co/Datadog/Toto-Open-Base-1.0 and https://github.com/DataDog/toto.
Sufficiently modeling the correlations among variables (aka channels) is crucial for achieving accurate multivariate time series forecasting (MTSF). In this paper, we propose a novel technique called Temporal Query (TQ) to more effectively capture multivariate correlations, thereby improving model performance in MTSF tasks. Technically, the TQ technique employs periodically shifted learnable vectors as queries in the attention mechanism to capture global inter-variable patterns, while the keys and values are derived from the raw input data to encode local, sample-level correlations. Building upon the TQ technique, we develop a simple yet efficient model named Temporal Query Network (TQNet), which employs only a single-layer attention mechanism and a lightweight multi-layer perceptron (MLP). Extensive experiments demonstrate that TQNet learns more robust multivariate correlations, achieving state-of-the-art forecasting accuracy across 12 challenging real-world datasets. Furthermore, TQNet achieves high efficiency comparable to linear-based methods even on high-dimensional datasets, balancing performance and computational cost. The code is available at: https://github.com/ACAT-SCUT/TQNet.
This paper presents $\mathbf{OLinear}$, a $\mathbf{linear}$-based multivariate time series forecasting model that operates in an $\mathbf{o}$rthogonally transformed domain. Recent forecasting models typically adopt the temporal forecast (TF) paradigm, which directly encode and decode time series in the time domain. However, the entangled step-wise dependencies in series data can hinder the performance of TF. To address this, some forecasters conduct encoding and decoding in the transformed domain using fixed, dataset-independent bases (e.g., sine and cosine signals in the Fourier transform). In contrast, we utilize $\mathbf{OrthoTrans}$, a data-adaptive transformation based on an orthogonal matrix that diagonalizes the series' temporal Pearson correlation matrix. This approach enables more effective encoding and decoding in the decorrelated feature domain and can serve as a plug-in module to enhance existing forecasters. To enhance the representation learning for multivariate time series, we introduce a customized linear layer, $\mathbf{NormLin}$, which employs a normalized weight matrix to capture multivariate dependencies. Empirically, the NormLin module shows a surprising performance advantage over multi-head self-attention, while requiring nearly half the FLOPs. Extensive experiments on 24 benchmarks and 140 forecasting tasks demonstrate that OLinear consistently achieves state-of-the-art performance with high efficiency. Notably, as a plug-in replacement for self-attention, the NormLin module consistently enhances Transformer-based forecasters. The code and datasets are available at https://anonymous.4open.science/r/OLinear
There has been a recent surge of interest in time series modeling using the Transformer architecture. However, forecasting multivariate time series with Transformer presents a unique challenge as it requires modeling both temporal (cross-time) and variate (cross-variate) dependencies. While Transformer-based models have gained popularity for their flexibility in capturing both sequential and cross-variate relationships, it is unclear how to best integrate these two sources of information in the context of the Transformer architecture while optimizing for both performance and efficiency. We re-purpose the Transformer architecture to effectively model both cross-time and cross-variate dependencies. Our approach begins by embedding each variate independently into a variate-wise representation that captures its cross-time dynamics, and then models cross-variate dependencies through attention mechanisms on these learned embeddings. Gating operations in both cross-time and cross-variate modeling phases regulate information flow, allowing the model to focus on the most relevant features for accurate predictions. Our method achieves state-of-the-art performance across 13 real-world datasets and can be seamlessly integrated into other Transformer-based and LLM-based forecasters, delivering performance improvements up to 20.7\% over original models. Code is available at this repository: https://github.com/nyuolab/Gateformer.
Multivariate time series forecasting is crucial across various industries, where accurate extraction of complex periodic and trend components can significantly enhance prediction performance. However, existing models often struggle to capture these intricate patterns. To address these challenges, we propose FilterTS, a novel forecasting model that utilizes specialized filtering techniques based on the frequency domain. FilterTS introduces a Dynamic Cross-Variable Filtering Module, a key innovation that dynamically leverages other variables as filters to extract and reinforce shared variable frequency components across variables in multivariate time series. Additionally, a Static Global Filtering Module captures stable frequency components, identified throughout the entire training set. Moreover, the model is built in the frequency domain, converting time-domain convolutions into frequency-domain multiplicative operations to enhance computational efficiency. Extensive experimental results on eight real-world datasets have demonstrated that FilterTS significantly outperforms existing methods in terms of prediction accuracy and computational efficiency.
Multivariate time series forecasting (MTSF) endeavors to predict future observations given historical data, playing a crucial role in time series data management systems. With advancements in large language models (LLMs), recent studies employ textual prompt tuning to infuse the knowledge of LLMs into MTSF. However, the deployment of LLMs often suffers from low efficiency during the inference phase. To address this problem, we introduce TimeKD, an efficient MTSF framework that leverages the calibrated language models and privileged knowledge distillation. TimeKD aims to generate high-quality future representations from the proposed cross-modality teacher model and cultivate an effective student model. The cross-modality teacher model adopts calibrated language models (CLMs) with ground truth prompts, motivated by the paradigm of Learning Under Privileged Information (LUPI). In addition, we design a subtractive cross attention (SCA) mechanism to refine these representations. To cultivate an effective student model, we propose an innovative privileged knowledge distillation (PKD) mechanism including correlation and feature distillation. PKD enables the student to replicate the teacher's behavior while minimizing their output discrepancy. Extensive experiments on real data offer insight into the effectiveness, efficiency, and scalability of the proposed TimeKD.
The Transformer model has shown strong performance in multivariate time series forecasting by leveraging channel-wise self-attention. However, this approach lacks temporal constraints when computing temporal features and does not utilize cumulative historical series effectively.To address these limitations, we propose the Structured Channel-wise Transformer with Cumulative Historical state (SCFormer). SCFormer introduces temporal constraints to all linear transformations, including the query, key, and value matrices, as well as the fully connected layers within the Transformer. Additionally, SCFormer employs High-order Polynomial Projection Operators (HiPPO) to deal with cumulative historical time series, allowing the model to incorporate information beyond the look-back window during prediction. Extensive experiments on multiple real-world datasets demonstrate that SCFormer significantly outperforms mainstream baselines, highlighting its effectiveness in enhancing time series forecasting. The code is publicly available at https://github.com/ShiweiGuo1995/SCFormer
Data-driven benchmarks have led to significant progress in key scientific modeling domains including weather and structural biology. Here, we introduce the Zebrafish Activity Prediction Benchmark (ZAPBench) to measure progress on the problem of predicting cellular-resolution neural activity throughout an entire vertebrate brain. The benchmark is based on a novel dataset containing 4d light-sheet microscopy recordings of over 70,000 neurons in a larval zebrafish brain, along with motion stabilized and voxel-level cell segmentations of these data that facilitate development of a variety of forecasting methods. Initial results from a selection of time series and volumetric video modeling approaches achieve better performance than naive baseline methods, but also show room for further improvement. The specific brain used in the activity recording is also undergoing synaptic-level anatomical mapping, which will enable future integration of detailed structural information into forecasting methods.
Recent advancements have progressively incorporated frequency-based techniques into deep learning models, leading to notable improvements in accuracy and efficiency for time series analysis tasks. However, the Mid-Frequency Spectrum Gap in the real-world time series, where the energy is concentrated at the low-frequency region while the middle-frequency band is negligible, hinders the ability of existing deep learning models to extract the crucial frequency information. Additionally, the shared Key-Frequency in multivariate time series, where different time series share indistinguishable frequency patterns, is rarely exploited by existing literature. This work introduces a novel module, Adaptive Mid-Frequency Energy Optimizer, based on convolution and residual learning, to emphasize the significance of mid-frequency bands. We also propose an Energy-based Key-Frequency Picking Block to capture shared Key-Frequency, which achieves superior inter-series modeling performance with fewer parameters. A novel Key-Frequency Enhanced Training strategy is employed to further enhance Key-Frequency modeling, where spectral information from other channels is randomly introduced into each channel. Our approach advanced multivariate time series forecasting on the challenging Traffic, ECL, and Solar benchmarks, reducing MSE by 4%, 6%, and 5% compared to the previous SOTA iTransformer. Code is available at this GitHub Repository: https://github.com/Levi-Ackman/ReFocus.
The diversity of time series applications and scarcity of domain-specific data highlight the need for time-series models with strong few-shot learning capabilities. In this work, we propose a novel training scheme and a transformer-based architecture, collectively referred to as TimePFN, for multivariate time-series (MTS) forecasting. TimePFN is based on the concept of Prior-data Fitted Networks (PFN), which aims to approximate Bayesian inference. Our approach consists of (1) generating synthetic MTS data through diverse Gaussian process kernels and the linear coregionalization method, and (2) a novel MTS architecture capable of utilizing both temporal and cross-channel dependencies across all input patches. We evaluate TimePFN on several benchmark datasets and demonstrate that it outperforms the existing state-of-the-art models for MTS forecasting in both zero-shot and few-shot settings. Notably, fine-tuning TimePFN with as few as 500 data points nearly matches full dataset training error, and even 50 data points yield competitive results. We also find that TimePFN exhibits strong univariate forecasting performance, attesting to its generalization ability. Overall, this work unlocks the power of synthetic data priors for MTS forecasting and facilitates strong zero- and few-shot forecasting performance.
Pre-trained foundation models (FMs) have shown exceptional performance in univariate time series forecasting tasks. However, several practical challenges persist, including managing intricate dependencies among features and quantifying uncertainty in predictions. This study aims to tackle these critical limitations by introducing adapters; feature-space transformations that facilitate the effective use of pre-trained univariate time series FMs for multivariate tasks. Adapters operate by projecting multivariate inputs into a suitable latent space and applying the FM independently to each dimension. Inspired by the literature on representation learning and partially stochastic Bayesian neural networks, we present a range of adapters and optimization/inference strategies. Experiments conducted on both synthetic and real-world datasets confirm the efficacy of adapters, demonstrating substantial enhancements in forecasting accuracy and uncertainty quantification compared to baseline methods. Our framework, AdaPTS, positions adapters as a modular, scalable, and effective solution for leveraging time series FMs in multivariate contexts, thereby promoting their wider adoption in real-world applications. We release the code at https://github.com/abenechehab/AdaPTS.
Multivariate Time Series Forecasting (MTSF) plays a crucial role across diverse fields, ranging from economic, energy, to traffic. In recent years, deep learning has demonstrated outstanding performance in MTSF tasks. In MTSF, modeling the correlations among different channels is critical, as leveraging information from other related channels can significantly improve the prediction accuracy of a specific channel. This study systematically reviews the channel modeling strategies for time series and proposes a taxonomy organized into three hierarchical levels: the strategy perspective, the mechanism perspective, and the characteristic perspective. On this basis, we provide a structured analysis of these methods and conduct an in-depth examination of the advantages and limitations of different channel strategies. Finally, we summarize and discuss some future research directions to provide useful research guidance. Moreover, we maintain an up-to-date Github repository (https://github.com/decisionintelligence/CS4TS) which includes all the papers discussed in the survey.
Generative models have gained significant attention in multivariate time series forecasting (MTS), particularly due to their ability to generate high-fidelity samples. Forecasting the probability distribution of multivariate time series is a challenging yet practical task. Although some recent attempts have been made to handle this task, two major challenges persist: 1) some existing generative methods underperform in high-dimensional multivariate time series forecasting, which is hard to scale to higher dimensions; 2) the inherent high-dimensional multivariate attributes constrain the forecasting lengths of existing generative models. In this paper, we point out that discrete token representations can model high-dimensional MTS with faster inference time, and forecasting the target with long-term trends of itself can extend the forecasting length with high accuracy. Motivated by this, we propose a vector quantized framework called Hierarchical Discrete Transformer (HDT) that models time series into discrete token representations with l2 normalization enhanced vector quantized strategy, in which we transform the MTS forecasting into discrete tokens generation. To address the limitations of generative models in long-term forecasting, we propose a hierarchical discrete Transformer. This model captures the discrete long-term trend of the target at the low level and leverages this trend as a condition to generate the discrete representation of the target at the high level that introduces the features of the target itself to extend the forecasting length in high-dimensional MTS. Extensive experiments on five popular MTS datasets verify the effectiveness of our proposed method.
Variate tokenization, which independently embeds each variate as separate tokens, has achieved remarkable improvements in multivariate time series forecasting. However, employing self-attention with variate tokens incurs a quadratic computational cost with respect to the number of variates, thus limiting its training efficiency for large-scale applications. To address this issue, we propose VarDrop, a simple yet efficient strategy that reduces the token usage by omitting redundant variate tokens during training. VarDrop adaptively excludes redundant tokens within a given batch, thereby reducing the number of tokens used for dot-product attention while preserving essential information. Specifically, we introduce k-dominant frequency hashing (k-DFH), which utilizes the ranked dominant frequencies in the frequency domain as a hash value to efficiently group variate tokens exhibiting similar periodic behaviors. Then, only representative tokens in each group are sampled through stratified sampling. By performing sparse attention with these selected tokens, the computational cost of scaled dot-product attention is significantly alleviated. Experiments conducted on public benchmark datasets demonstrate that VarDrop outperforms existing efficient baselines.
This paper presents \textbf{FreEformer}, a simple yet effective model that leverages a \textbf{Fre}quency \textbf{E}nhanced Trans\textbf{former} for multivariate time series forecasting. Our work is based on the assumption that the frequency spectrum provides a global perspective on the composition of series across various frequencies and is highly suitable for robust representation learning. Specifically, we first convert time series into the complex frequency domain using the Discrete Fourier Transform (DFT). The Transformer architecture is then applied to the frequency spectra to capture cross-variate dependencies, with the real and imaginary parts processed independently. However, we observe that the vanilla attention matrix exhibits a low-rank characteristic, thus limiting representation diversity. This could be attributed to the inherent sparsity of the frequency domain and the strong-value-focused nature of Softmax in vanilla attention. To address this, we enhance the vanilla attention mechanism by introducing an additional learnable matrix to the original attention matrix, followed by row-wise L1 normalization. Theoretical analysis~demonstrates that this enhanced attention mechanism improves both feature diversity and gradient flow. Extensive experiments demonstrate that FreEformer consistently outperforms state-of-the-art models on eighteen real-world benchmarks covering electricity, traffic, weather, healthcare and finance. Notably, the enhanced attention mechanism also consistently improves the performance of state-of-the-art Transformer-based forecasters.
We introduce an interpretable deep learning model for multivariate time series forecasting that prioritizes both predictive performance and interpretability - key requirements for understanding complex physical phenomena. Our model not only matches but often surpasses existing interpretability methods, achieving this without compromising accuracy. Through extensive experiments, we demonstrate its ability to identify the most relevant time series and lags that contribute to forecasting future values, providing intuitive and transparent explanations for its predictions. To minimize the need for manual supervision, the model is designed so one can robustly determine the optimal window size that captures all necessary interactions within the smallest possible time frame. Additionally, it effectively identifies the optimal model order, balancing complexity when incorporating higher-order terms. These advancements hold significant implications for modeling and understanding dynamic systems, making the model a valuable tool for applied and computational physicists.
In recent years, the application of transformer-based models in time-series forecasting has received significant attention. While often demonstrating promising results, the transformer architecture encounters challenges in fully exploiting the temporal relations within time series data due to its attention mechanism. In this work, we design eXponential Patch (xPatch for short), a novel dual-stream architecture that utilizes exponential decomposition. Inspired by the classical exponential smoothing approaches, xPatch introduces the innovative seasonal-trend exponential decomposition module. Additionally, we propose a dual-flow architecture that consists of an MLP-based linear stream and a CNN-based non-linear stream. This model investigates the benefits of employing patching and channel-independence techniques within a non-transformer model. Finally, we develop a robust arctangent loss function and a sigmoid learning rate adjustment scheme, which prevent overfitting and boost forecasting performance. The code is available at the following repository: https://github.com/stitsyuk/xPatch.