TL;DR
TimeBridge resolves a long-standing dilemma in multivariate time series forecasting: whether to eliminate non-stationarity to avoid short-term spurious regressions or preserve it to capture long-term cointegration. By introducing a dual-attention mechanism, it mitigates volatile fluctuations within individual variables while preserving non-stationary trends across variables to model long-term relationships. This hybrid approach yields state-of-the-art performance across diverse benchmarks, including highly volatile traffic and financial datasets.
What Problem Does This Paper Address?
Real-world multivariate time series—such as stock prices, weather metrics, or traffic flows—are notoriously non-stationary, meaning their underlying statistical properties change over time due to shifting trends and sudden fluctuations. This creates a classic catch-22 for forecasting models:
- If you remove non-stationarity (e.g., via normalization), you prevent the model from hallucinating false correlations (spurious regressions) caused by short-term noise. However, you also destroy the long-term shared trends (cointegration) that naturally exist between different variables.
- If you retain non-stationarity, the model can track long-term equilibrium trends, but it easily gets tripped up by local, highly random short-term fluctuations.
Existing forecasting methods typically choose one side of the coin, either over-stabilizing the data or failing to protect against spurious correlations. TimeBridge addresses this gap by recognizing that non-stationarity must be handled differently depending on the temporal scale.
Methods
The authors propose TimeBridge, a framework that decouples short-term and long-term dependency modeling to handle non-stationarity adaptively.
The architecture consists of four main components:
- Patch Embedding: The input sequence for each variable is segmented into smaller, non-overlapping patches. This localizes short-term fluctuations within discrete tokens.
- Integrated Attention: Operating under a Channel Independent paradigm, this module applies patch-wise normalization to strip away short-term non-stationarity. It safely models temporal dependencies within each individual variable without the risk of spurious regressions.
- Patch Downsampling: An MLP and attention mechanism aggregate the patches. This reduces the total patch count while enriching each remaining patch with broader, global context.
- Cointegrated Attention: Operating under a Channel Dependent paradigm, this module intentionally retains non-stationarity. It analyzes the downsampled patches across different variables to capture long-term cointegration and tracking trends.
Datasets
The framework was validated across a comprehensive set of real-world benchmarks:
- Long-term Forecasting: The Electricity Transformer Temperature datasets (ETTh1, ETTh2, ETTm1, ETTm2), Weather, Electricity, Traffic, and Solar-Energy.
- Short-term Forecasting: The PeMS traffic network datasets (PeMS03, PeMS04, PeMS07, and PeMS08), which capture complex spatiotemporal traffic network data in California.
- Financial Forecasting: The CSI 500 (Chinese stock market) and S&P 500 (U.S. stock market) indices, spanning from 2018 to 2023.
Key Insights and Ideas
- Scale Separation of Non-Stationarity: The core "aha moment" of the paper is that non-stationarity isn't a monolith. Short-term non-stationarity is mostly noise that disrupts local causality, whereas long-term non-stationarity contains the vital structural signal that binds different variables together over time.
- The CI/CD Alignment: Short-term causal relationships primarily occur within a single variable over time, making Channel Independent (CI) modeling ideal. Conversely, long-term equilibrium dynamics occur across different variables, requiring Channel Dependent (CD) modeling. TimeBridge elegantly maps its dual attention modules to these distinct properties.
- Stationary Attention Maps: By using normalized patches as the Query and Key but the original unnormalized patches as the Value in the Integrated Attention block, the model generates a stable attention map without needing complex, error-prone denormalization steps later.
Key Results
TimeBridge consistently achieved state-of-the-art results across all evaluated tasks:
- Long-term Forecasting: On the massive Traffic dataset (862 channels), TimeBridge handled high dimensionality exceptionally well. Compared to top baselines DeformableTST, ModernTCN, and TimeMixer, TimeBridge reduced Mean Squared Error (MSE) by 1.85%, 5.56%, and 13.66%, respectively.
- Short-term Forecasting: On the PeMS datasets, where purely channel-independent models typically suffer due to strong inter-variable traffic dependencies, TimeBridge outperformed all baselines, including the recent state-of-the-art TimeMixer.
- Financial Portfolio Backtesting: When deployed in a buy-hold-sell trading strategy on the CSI 500 and S&P 500, TimeBridge achieved significantly higher Annual Return Ratios (ARR) and Sharpe Ratios (ASR) than classic econometrics and deep learning baselines, demonstrating robust market efficiency.
Limitations and Future Work
- Downsampling Sensitivity: The model's performance is sensitive to the patch downsampling rate ($M$). While datasets with hundreds of channels and strong cointegration (like Solar and Traffic) benefit from more downsampled patches, datasets with fewer channels (like Weather) suffer if downsampled too aggressively.
- Hyperparameter Tuning: Currently, balancing the downsampling rate and patch sizes requires manual tuning based on the specific dataset's characteristics, rather than utilizing a fully automated or dynamic allocation strategy.
Summary updated