Time Series Forecasting: Unpacking the Latest Breakthroughs – From MLPs to LLMs, and Beyond!
Latest 9 papers on time series forecasting: May. 2, 2026
Time series forecasting is the heartbeat of countless industries, from energy grids to financial markets, and fluid dynamics simulations. Accurately predicting future values from historical data is a challenge constantly pushing the boundaries of AI/ML. Recent research reveals exciting advancements, tackling issues from model complexity and interpretability to real-world operational challenges and high-performance computing needs. Let’s dive into some of the latest breakthroughs that are reshaping the landscape of time series prediction.
The Big Idea(s) & Core Innovations
One striking theme emerging from recent work is the push for more efficient, robust, and interpretable models, sometimes even challenging the dominance of complex architectures. For instance, the paper “ITS-Mina: A Harris Hawks Optimization-Based All-MLP Framework with Iterative Refinement and External Attention for Multivariate Time Series Forecasting” by Pourya Zamanvaziri et al. from Shahid Beheshti University, Iran, introduces ITS-Mina, an all-MLP framework that surprisingly outperforms Transformer-based models on several multivariate datasets. Their key insight? Simple MLP architectures, combined with iterative refinement via shared-parameter residual mixer loops and a linear-complexity external attention mechanism, can achieve state-of-the-art performance with less computational overhead. They even use Harris Hawks Optimization for automatic dropout tuning, showing that nature-inspired metaheuristics can optimize continuous hyperparameters effectively.
Challenging the conventional wisdom in a different way, “DecompKAN: Decomposed Patch-KAN for Long-Term Time Series Forecasting” by Naveen Mysore introduces DecompKAN, an attention-free architecture leveraging Kolmogorov-Arnold Networks (KANs) for long-term forecasting. This model prioritizes interpretability, allowing direct visualization of learned 1D nonlinear transformations via B-spline KAN edge functions. A core takeaway is that the overall pipeline design (decomposition, patching, normalization) often matters more than the specific nonlinear layer (KAN vs. attention vs. linear), especially for physics-structured data.
However, the application of KANs to time series isn’t without its caveats. “Autocorrelation Reintroduces Spectral Bias in KANs for Time Series Forecasting” by Chen Zeng et al. from Southeast University, China, reveals a critical issue: KANs, despite theoretical claims of overcoming spectral bias under independent inputs, suffer from it in time series due to temporal autocorrelation. Their solution, DCT-KAN, uses Discrete Cosine Transform (DCT) preprocessing to decorrelate inputs, significantly reducing low-frequency preference and improving high-frequency component recovery.
Meanwhile, Large Language Models (LLMs) are also stepping into the forecasting arena. “CastFlow: Learning Role-Specialized Agentic Workflows for Time Series Forecasting” by Bokai Pan et al. from the University of Science and Technology of China, proposes CastFlow, an agentic framework that transforms forecasting from a static one-shot generation to a dynamic, multi-step workflow. It uses role-specialized LLMs (a frozen LLM for reasoning, a fine-tuned one for numerical forecasting) with an evidence-guided iterative refinement process, demonstrating how LLMs can bring advanced reasoning to numerical tasks. The framework is trained using a two-stage workflow combining supervised fine-tuning (SFT) and reinforcement learning with verifiable rewards (RLVR).
Extending on the success of Mamba models, “AdaMamba: Adaptive Frequency-Gated Mamba for Long-Term Time Series Forecasting” by Xudong Jiang et al. from Tongji University, addresses cross-domain heterogeneity. AdaMamba endogenizes adaptive frequency-domain analysis directly into Mamba’s state-space update. By extending the temporal forgetting gate to a unified time-frequency forgetting gate, it dynamically calibrates state transitions based on learned frequency importance, while retaining Mamba’s linear complexity.
Under the Hood: Models, Datasets, & Benchmarks
This collection of papers highlights several crucial models, datasets, and benchmarking initiatives:
- ITS-Mina: An all-MLP architecture with shared-parameter residual mixer loops and an external attention module. Evaluated on Traffic, Electricity, ETTh1/2, ETTm1/2 datasets.
- CastFlow: A dynamic agentic framework using a combination of frozen and fine-tuned LLMs, supported by a memory module and a multi-view toolkit for temporal pattern extraction. Code is available at https://github.com/Forever-Pan/CastFlow.
- DecompKAN: A B-spline KAN-based architecture for interpretability, utilizing trend-residual decomposition and learned instance normalization. Tested on PPG-DaLiA, Weather, Solar, ECL, Traffic, ETTh1/2, ETTm1/2 datasets.
- DCT-KAN: A modification of KANs, employing Discrete Cosine Transform (DCT) preprocessing to mitigate spectral bias in time series. References FastKAN implementation.
- AdaMamba: An extension of Mamba state-space models, integrating adaptive frequency gating directly into the state update mechanism. Tested on ETT (h1/2, m1/2), Weather, ILI, Exchange Rate, KnowAir-V2, NOAA datasets. Code at https://github.com/XDjiang25/AdaMamba.
- FETS Benchmark: This crucial work by Marco Obermeier et al. from Julius-Maximilians-Universität Würzburg, Germany, introduces a new benchmark for Foundation Models in Energy Time Series (FETS). It systematically evaluates 54 energy datasets across 9 categories, featuring Chronos-2, TimesFM, TiRex, FlowState, TabPFN-TS, and comparing them against XGBoost and Random Forest. The dataset collection relies on sources like ENTSO-E, Open-Meteo, CAISO, and more. Found at https://arxiv.org/pdf/2604.22328.
- Energy-Arena: A dynamic benchmarking platform for operational energy forecasting by Max Kleinebrahm et al. from Karlsruhe Institute of Technology, Germany. Available at https://Energy-Arena.org, it offers forward-looking evaluation, standardized challenges, and persistent leaderboards for both point and probabilistic forecasts, preventing information leakage inherent in static benchmarks. It integrates with external data providers like ENTSO-E.
- Distributed CFD Simulations: “A Study on the Performance of Distributed Training of Data-driven CFD Simulations” by Sergio Iserte et al. from Universitat Jaume I, Spain, focuses on distributed training for RNN-based time series forecasting in Computational Fluid Dynamics (CFD). They use RNNs with LSTM layers and evaluate Horovod and TensorFlow MultiworkerStrategy on HPC clusters, showing significant GPU-enabled speedups. Code is at https://github.com/AlejandroGB13/CFD_AI.
- Power System Forecasting Benchmark: Muhy Eddin Za’tera and Bri-Mathias Hodge from University of Colorado Boulder and NREL, in “Empirical Assessment of Time-Series Foundation Models For Power System Forecasting Applications”, rigorously benchmark TimesFM, Chronos-Bolt, Moirai-L, MOMENT, TTM, TFT, TimeXer, LSTM, and CNN for solar, wind, and load forecasting on the high-resolution ARPA-E PERFORM dataset for ERCOT.
Impact & The Road Ahead
These advancements have profound implications. The success of all-MLP models like ITS-Mina suggests that architectural simplicity, combined with clever design patterns, can yield high performance at reduced computational costs. This is crucial for deploying AI/ML models in resource-constrained environments. The drive for interpretability, as seen with DecompKAN, makes black-box models more trustworthy and allows domain experts to gain insights into their predictions, fostering greater adoption in critical sectors.
The integration of LLMs into forecasting workflows, as demonstrated by CastFlow, opens new avenues for leveraging general-purpose reasoning with specialized numerical precision, shifting towards more adaptive and intelligent forecasting agents. AdaMamba’s adaptive frequency gating in Mamba models offers a powerful way to handle the inherent heterogeneity in time series data, ensuring robust performance across diverse domains.
The new benchmarks, FETS and Energy-Arena, are particularly impactful. They address the critical need for standardized, dynamic, and forward-looking evaluation, moving research closer to real-world operational constraints. The FETS benchmark specifically highlights the superior performance of covariate-informed foundation models over traditional ML, even in zero-shot settings, indicating a powerful shift towards more generalizable models for energy applications, especially where data is scarce. However, the ERCOT power system benchmark by Za’tera and Hodge cautions that while foundation models are data-efficient, they still require fine-tuning for operational readiness in complex power systems, and transformers currently excel at leveraging multivariate weather inputs. This suggests a future for hybrid approaches.
Collectively, this research points towards a future where time series forecasting models are not only more accurate and efficient but also more interpretable, adaptable, and robust to diverse, real-world conditions. The journey continues, with a clear focus on making these powerful tools more accessible and effective for all.
Share this content:
Post Comment