Research map / Time series / Long-term forecasting

CATS: Enhancing Multivariate Time Series Forecasting by Constructing Auxiliary Time Series as Exogenous Variables

ICML 2024 · Time series › Long-term forecasting · approach: Graph-based multivariate forecasting · 6 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes CATS, which builds auxiliary time series (ATS) from multivariate inputs with continuity, sparsity and variability principles, letting a simple 2-layer MLP predictor match or beat state-of-the-art multivariate forecasters at lower cost.

Problem

Multivariate time series forecasting (MTSF) where univariate models often beat multivariate ones; existing multivariate models fail to capture inter-series relationships effectively and overfit when such relationships are weak.

Approach

Generate auxiliary time series from the original series via several constructors (conv, non-overlapping conv, independent conv, linear, identity, embedding) that act as exogenous variables, regularized by a continuity loss, channel and temporal sparsity gates, and an OTS shortcut, then project them back to the OTS output; works as a 2D temporal-contextual attention-like mechanism around a simple p

Key results

• CATS (2L) achieves the lowest average MSE percentage (43.9% vs 44.9% for ARM, the second best) and best average rank 1.25 across nine datasets, winning 28 of 36 sub-experiments.
• On the MultiX extended benchmarks, CATS reaches 59.6% average MSE percentage vs 70.5% for ARM and wins 12 of 16 sub-experiments.
• CATS (2L) has 2.47M params and 382M FLOPs at L_P=96 versus 7.89M params and 426M FLOPs for ARM, and 15.5M params for Autoformer.

Measured effects reported in the paper

Surprising findings

• A basic identity-mapping ATS constructor with an independent linear predictor, with no trainable ATS parameters, already turns a univariate model into one that solves a shifting multivariate problem.
• Removing the OTS shortcut or continuity loss degrades performance notably on Multi and ETTh2, showing ATS-OTS division of labor matters.

Limitations

• Number of ATS channels is fixed in advance and cannot grow automatically during training if weak-but-complex relationships need more ATS.
• Gains are concentrated on datasets with strong inter-series dependencies (Multi/MultiX); improvements on ETT and Weather are marginal.
• Evaluated only on forecasting; transfer to NLP or audio is proposed but untested.

Experimental setup

Datasets: ETTm1 ETTm2 ETTh1 ETTh2 Electricity Traffic Weather Exchange Multi MultiX (Multi20, Multi50, Multi100, Multi200)

Benchmarks: ETTm1 ETTm2 ETTh1 ETTh2 Electricity Traffic Weather Exchange Multi Multi20 Multi50 Multi100

Metrics: MSE MAE average MSE percentage relative to Repeat average rank win count

Baselines: ARM PatchTST DLinear FEDformer Autoformer Informer Repeat

Builds on: ARM DLinear PatchTST

Closest papers by meaning

• ARM: Refining Multivariate Forecasting with Adaptive Temporal-Contextual Learning (ICLR 2024 · similarity 0.90)
• TimePFN: Effective Multivariate Time Series Forecasting with Synthetic Data (AAAI 2025 · similarity 0.88)
• Revisiting Attention for Multivariate Time Series Forecasting (AAAI 2025 · similarity 0.88)
• TimeXer: Empowering Transformers for Time Series Forecasting with Exogenous Variables (NeurIPS 2024 · similarity 0.87)
• TimeCMA: Towards LLM-Empowered Multivariate Time Series Forecasting via Cross-Modality Alignment (AAAI 2025 · similarity 0.87)
• Revitalizing Multivariate Time Series Forecasting: Learnable Decomposition with Inter-Series Dependencies and Intra-Series Variations Modeling (ICML 2024 · similarity 0.87)