TimePerceiver: An Encoder-Decoder Framework for Generalized Time-Series Forecasting

TL;DR

TimePerceiver introduces a unified encoder–decoder framework for generalized time-series forecasting that extends beyond predicting only the future. By formulating a generalized objective over arbitrary input and target temporal segments and coupling it with a patch-based embedding strategy, a latent bottleneck encoder, and a query-based decoder, the approach flexibly handles extrapolation, interpolation, and imputation. Empirical results across eight real-world datasets show consistent, state-of-the-art performance and strong efficiency due to the latent bottleneck reducing attention complexity. The work emphasizes aligning architectural design with the generalized forecasting task, offering a pathway toward more holistic, end-to-end time-series models with practical impact in diverse domains.

Abstract

In machine learning, effective modeling requires a holistic consideration of how to encode inputs, make predictions (i.e., decoding), and train the model. However, in time-series forecasting, prior work has predominantly focused on encoder design, often treating prediction and training as separate or secondary concerns. In this paper, we propose TimePerceiver, a unified encoder-decoder forecasting framework that is tightly aligned with an effective training strategy. To be specific, we first generalize the forecasting task to include diverse temporal prediction objectives such as extrapolation, interpolation, and imputation. Since this generalization requires handling input and target segments that are arbitrarily positioned along the temporal axis, we design a novel encoder-decoder architecture that can flexibly perceive and adapt to these varying positions. For encoding, we introduce a set of latent bottleneck representations that can interact with all input segments to jointly capture temporal and cross-channel dependencies. For decoding, we leverage learnable queries corresponding to target timestamps to effectively retrieve relevant information. Extensive experiments demonstrate that our framework consistently and significantly outperforms prior state-of-the-art baselines across a wide range of benchmark datasets. The code is available at https://github.com/efficient-learning-lab/TimePerceiver.

Figures (5)

• Figure 1: (a) The standard time-series forecasting task aims to predict only the future values from past observations. In contrast, (b) our generalized task formulation aims to predict not only the future, but also the past and missing values based on arbitrary contextual information.
• Figure 2: An overview of our TimePerceiver framework.
• Figure 3: Relative MSE reduction (%) compared to the standard formulation with varying sampling strategies for target patches under input length $L = 384$ and two prediction lengths $H \in \{96, 336\}$. Sampling strategies are (a) fully contiguous: only one contiguous segment in $\mathcal{J}_\text{patch}$, (b) fully disjoint: selected target indices are fully arbitrary, and (c) mixed: two equal-sized segments in $\mathcal{J}_\text{patch}$. Higher values indicate greater improvements over the standard formulation.
• Figure 4: Cross-attention maps from the encoder and decoder on ETTh1 and ETTm1. For each dataset, the left image shows attention from latent tokens to input patches (encoder), and the right shows attention from query tokens to input patches (decoder). In encoder maps, the x-axis shows flattened input patch indices ($C \times |\mathcal{I}_\text{patch}|$), and the y-axis represents latent tokens ($M$). In decoder maps, the x-axis is temporal input positions ($|\mathcal{I}_\text{patch}|$, $L{=}384$), and the y-axis is query tokens aligned with prediction patches ($|\mathcal{J}_\text{patch}|$, $H{=}96$). Each token spans 12 hours (ETTh1) or 6 hours (ETTm1).
• Figure 5: (a) and (b) compare computational costs resulting from different encoder structures, including GPU memory usage and epoch time per input length. (c) shows the difference in parameter count between our model and other commonly used models with decoder structures based on linear projection, across various forecasting horizons.