Time series analysis has historically been a fragmented landscape. If you want to predict the next 24 hours of electricity usage, you build a forecasting model. If you want to detect a faulty heartbeat in an ECG stream, you build an anomaly detection model. If you want to identify a physical activity from accelerometer data, you build a classification model.
This task-specific siloing stands in stark contrast to natural language processing or computer vision, where a single foundation model can seamlessly transition between generation, classification, and structural analysis. The emerging frontier in time series research is cross-task generalization: engineering representations and architectures that can natively transcend the boundary between predictive (generative) and discriminative tasks.
Why the Divide Exists: The Geometry of Representations
To understand why cross-task generalization is difficult, we must look at what these tasks demand from a model's latent space.
Forecasting models focus heavily on local temporal continuity and periodic trends. For example, Gerald Woo et al. developed CoST (CoST: Contrastive Learning of Disentangled Seasonal-Trend Representations for Time Series Forecasting), which explicitly disentangles seasonal and trend components to optimize multi-step lookahead windows. However, a representation perfectly tuned to isolate smooth trends and periodic waves may discard the exact phase shifts or high-frequency anomalies needed for classification or outlier detection.
Conversely, classification models often compress the entire temporal dimension into a single global embedding vector. As shown by Xiang Zhang et al. in their graph-guided network RAINDROP (Graph-Guided Network for Irregularly Sampled Multivariate Time Series), classification frameworks excel at mapping irregular sensor relationships to discrete labels, but they do not preserve the sequential autoregressive structure required to project continuous values into the future.
Unified Self-Supervised Learning as a Bridge
How do we build a representation that satisfies both worlds? The answer lies in multi-objective self-supervised learning (SSL). Instead of training a model on a single proxy task (like masked autoencoding), researchers are designing compound objectives that force the latent space to encode multiple facets of the data simultaneously.
Heejeong Choi et al. demonstrated this intuition in Multi-task self-supervised time-series representation learning by combining three distinct consistency objectives: contextual, temporal, and transformation-based. By forcing the network to remain invariant to artificial transformations while simultaneously tracking temporal progression, the learned embeddings become robust enough to serve downstream classification, forecasting, and anomaly detection alike.
Without this multi-faceted pressure, single-objective SSL frameworks often remain trapped in their original domain. For instance, the Transformer framework proposed by George Zerveas et al. (A Transformer-based Framework for Multivariate Time Series Representation Learning) proved highly effective for classification and regression but left generative tasks like forecasting and imputation largely unaddressed due to its representation structure. Similarly, the universal feature learning framework by Huiyuan Chen et al. (Toward a Foundation Model for Time Series Data) demonstrated massive success in downstream classification by pre-training on diverse datasets, yet extending such representations to long-term prediction remains an open challenge.
The Channel Interaction Dilemma
A major roadblock to cross-task generalization is how multivariate time series handle relationships between different variables (channels).
In forecasting, treating channels independently often yields surprisingly resilient results because it prevents cross-channel noise from corrupting temporal trends. Kibok Lee (Channel Normalization for Time Series Channel Identification) demonstrates that channel-specific normalization significantly boosts forecasting accuracy by preserving channel identifiability. To re-introduce necessary interactions, Hanyin Cheng et al. introduced CoRA (CoRA: Boosting Time Series Foundation Models for Multivariate Forecasting through Correlation-aware Adapter) to adaptively capture inter-channel dependencies specifically for forecasting.
However, for classification or anomaly detection, the interaction between channels is often the primary signal (e.g., a sudden misalignment between blood pressure and heart rate indicates an anomaly). A model optimized strictly for channel-independent forecasting fails when cross-task transfer requires deep, discriminative cross-channel reasoning.
The Generative Foundation Model Paradigm
The ultimate manifestation of cross-task generalization is the Time Series Foundation Model (TSFM). Rather than adapting a small model to multiple tasks, the goal is to pre-train a massive generative model that treats all time series problems as a unified sequence-to-sequence problem.
In Timer (Timer: Generative Pre-trained Transformers Are Large Time Series Models), Yong Liu demonstrates that a large GPT-style decoder pre-trained on billions of time points can natively handle forecasting, imputation, and anomaly detection within a single generative framework. By framing anomaly detection as a high reconstruction error or low next-token probability, generative models can repurpose their predictive capabilities for diagnostic tasks.
Yet, even these large-scale models face a structural bottleneck: incorporating discrete classification tasks into a purely next-token autoregressive prediction framework remains an open challenge. To bridge this, multimodal frameworks like ChatTime, developed by Chengsen Wang et al. (ChatTime: A Unified Multimodal Time Series Foundation Model Bridging Numerical and Textual Data), map numerical time series into a textual token space. This allows the model to leverage the broad reasoning capabilities of Large Language Models to perform both generative forecasting and discriminative reasoning (such as trend and outlier classification) under a single universal paradigm.