InstructTime++ · Preprint
Extends InstructTime with implicit feature enhancement via statistical feature extraction and vision-language-based image captioning for superior classification performance.
Paper →Research Projects
Time Series Classification & Forecasting
Our research spans multiple directions: multimodal language modeling for classification, reasoning-based forecasting, self-supervised representation learning, and comprehensive surveys.
Time Series Classification
InstructTime · WSDM 2025 (Best of WSDM)
Reformulates time series classification as a multimodal generative task, converting continuous sequences into discrete tokens and leveraging pre-trained language models.
Paper →
HiTime · ACM TIST
Hierarchical multimodal LLMs with semantic space alignment for time series classification: hierarchical sequence encoding, coarse-to-fine cross-modal alignment, and generative classification via parameter-efficient fine-tuning.
Paper →
TableTime · CIKM 2025
Reformulates multivariate time series classification as zero-shot table understanding via large language models, minimizing information loss through tabular representation.
Paper →
TimeMAE · WSDM 2026
Self-supervised representation learning with decoupled masked autoencoders, learning enriched contextual representations through bidirectional encoding over sub-series.
Paper →
CrossTimeNet · WSDM 2025
Cross-domain self-supervised pre-training with language models for transferable time series representations; vector quantization tokenization and corrupted-region recovery for classification and forecasting downstream tasks.
Code →
ConvTimeNet · WWW 2025
Hierarchical fully convolutional model with deformable patch layers for adaptive perception of local patterns and multi-scale dependency modeling.
Paper →
FormerTime · WWW 2023
Hierarchical multi-scale representation for multivariate time series classification, capturing temporal dependencies at multiple granularities.
Paper →Time Series Forecasting
CastMind
Interaction-driven agentic reasoning framework for cognition-inspired forecasting, organizing prediction into multi-stage workflow with context preparation and reflective evaluation.
Paper →
OneCast
Structured decomposition and modular generation for cross-domain time series forecasting: decomposes into seasonal (lightweight projection with basis functions) and trend (semantic-aware tokenizer + masked discrete diffusion) components for generalization across domains.
Paper →
MemCast
Memory-driven forecasting with experience-conditioned reasoning, learning historical patterns, reasoning wisdom, and general laws from training data.
Code →
Time-R1
Slow-thinking approach with reinforced LLMs for time series forecasting, employing two-stage reinforcement fine-tuning to enhance multi-step reasoning ability.
Paper →
TokenCast
LLM-driven framework for context-aware forecasting via symbolic discretization, transforming continuous sequences into temporal tokens for unified representation.
Paper →
TimeReasoner · WSDM 2026
Empirical study on slow-thinking LLMs reasoning over temporal patterns, demonstrating non-trivial zero-shot forecasting capabilities in capturing trends and contextual shifts.
Paper →
TimeDART · ICML 2025
Diffusion autoregressive transformer for self-supervised time series representation, unifying causal transformer encoding with denoising diffusion for global and local patterns.
Paper →
FDF · IJCAI 2025
Flexible decoupled framework that decomposes time series into trend and seasonal components, employing conditional denoising (diffusion) for the fluctuating seasonal part and enhanced linear models for the smooth trend, trained end-to-end.
Paper →
HMF · IJCAI 2025
Hybrid multi-factor network with dynamic sequence modeling for early warning of intraoperative hypotension; decomposes physiological signals into trend and seasonal components with patch-based Transformers and symmetric normalization.
Paper →
PIR · NeurIPS 2025
Instance-aware post-hoc revision framework for improving forecasting via identification and revision of biased instances using contextual information.
Paper →
GPHT · SIGKDD 2024
Generative pretrained hierarchical transformer for forecasting, employing mixed dataset pretraining and autoregressive generation for arbitrary horizon settings.
Paper →
SAN · NeurIPS 2023
Adaptive normalization for non-stationary time series forecasting from a temporal slice perspective, eliminating non-stationarity at local slice level.
Paper →Survey
A Comprehensive Survey of Time Series Forecasting · Preprint
Examines key entities in TSF and their characteristics, introduces a general problem formulation and challenge analysis, proposes a taxonomy classifying methodologies from preprocessing and forecasting perspectives, and highlights emerging topics like transfer learning and trustworthy forecasting.
Paper →Timeline
Milestones in Time Series Research
Key publications from 2023 to 2026 across classification, forecasting, and representation learning.
2023
SAN · NeurIPS 2023
Introduced adaptive normalization for non-stationary time series forecasting, addressing distribution shifts through temporal slice-level normalization.
2023
FormerTime · WWW 2023
Proposed hierarchical multi-scale representation for multivariate time series classification, establishing foundation for multi-granularity temporal modeling.
2024
GPHT · SIGKDD 2024
Developed generative pretrained hierarchical transformer enabling single model for arbitrary horizon forecasting through mixed dataset pretraining.
2025
InstructTime · WSDM 2025 (Best of WSDM)
Awarded Best of WSDM for reformulating classification as multimodal generative task, demonstrating potential for universal foundation models in time series.
2025
InstructTime++ · Preprint
Extends InstructTime with implicit feature enhancement for superior time series classification.
2025
CrossTimeNet · WSDM 2025
Cross-domain pre-training with LMs for transferable time series representations via vector quantization tokenization and self-supervised recovery objective.
2025
TimeDART · ICML 2025
Unified diffusion and autoregressive paradigms for self-supervised representation learning, capturing both global trends and local patterns.
2025
TableTime · CIKM 2025
Enabled zero-shot classification through table understanding reformulation, achieving natural alignment with LLM semantic space.
2025
CastMind
Interaction-driven agentic reasoning framework for cognition-inspired time series forecasting.
2025
MemCast
Memory-driven forecasting with experience-conditioned reasoning from historical patterns and reasoning wisdom.
2025
Time-R1
Slow-thinking approach with reinforced LLMs for multi-step reasoning in time series forecasting.
2025
TokenCast
LLM-driven context-aware forecasting via symbolic discretization and unified semantic space.
2025
PIR · NeurIPS 2025
Model-agnostic framework for instance-aware post-hoc revision, effectively mitigating instance-level errors in forecasting.
2025
ConvTimeNet · WWW 2025
Demonstrated viability of pure convolutional models with hierarchical architecture and deformable patch layers for time series analysis.
2025
HiTime · ACM TIST (Accepted Nov 2025)
Hierarchical multimodal LLMs with semantic space alignment for time series classification, bridging temporal representations with generative reasoning.
2025
FDF · IJCAI 2025
Flexible decoupled framework combining diffusion for seasonal component and enhanced linear models for trend, trained end-to-end.
2025
HMF · IJCAI 2025
Hybrid multi-factor network for early warning of intraoperative hypotension via dynamic sequence modeling and trend-seasonal decomposition.
2025
OneCast
Structured decomposition and modular generation for cross-domain time series forecasting: seasonal projection with basis functions and trend modeling via semantic-aware tokenizer and masked discrete diffusion.
2026
TimeMAE · WSDM 2026
Advanced self-supervised learning with decoupled masked autoencoders, learning transferable representations through bidirectional encoding.
2026
TimeReasoner · WSDM 2026
Empirical study revealing slow-thinking LLMs' capabilities in temporal reasoning, paving way for reasoning-based forecasting paradigms.
—
Comprehensive Survey of Time Series Forecasting
Concepts, challenges, taxonomy, and future directions for time series forecasting research.
Resources
Research Directions & Contributions
Multimodal Language Modeling
Converting time series to discrete tokens and leveraging pre-trained language models for classification and forecasting tasks.
Reasoning-Based Forecasting
Slow-thinking approaches and agentic reasoning frameworks that enable multi-step reasoning over temporal patterns.
Self-Supervised Learning
Masked autoencoders, diffusion models, and hierarchical transformers for learning transferable time series representations.
Comprehensive Surveys
In-depth analysis of time series forecasting concepts, challenges, architectural trends, and future directions.
Advancing Time Series Understanding
We explore novel paradigms for time series analysis, from multimodal language modeling to reasoning-based forecasting. Our research addresses fundamental challenges in temporal data understanding, including non-stationarity, long-term dependencies, and contextual integration.
Multimodal Integration
Bridging numerical sequences with contextual features through symbolic discretization and alignment strategies for enhanced understanding.
Reasoning Capabilities
Developing slow-thinking approaches and memory-driven frameworks that enable explicit reasoning over temporal dynamics and dependencies.
Transferable Representations
Self-supervised learning methods that learn generalizable patterns across diverse time series domains and applications.