Time Series Forecasting: Unlocking New Potential with Adaptive LLMs and Physics-Guided AI
Latest 5 papers on time series forecasting: Sep. 13, 2026
Time series forecasting is the bedrock of decision-making in everything from finance to climate science, yet it presents enduring challenges: dynamic shifts, sparse data, and the sheer complexity of underlying processes. The past year has seen remarkable strides, particularly in harnessing the power of Large Language Models (LLMs) and integrating domain-specific knowledge to create more robust, accurate, and adaptable forecasting systems. Let’s dive into some groundbreaking research that’s reshaping this exciting field.
The Big Idea(s) & Core Innovations
The overarching theme in recent research is adaptation and integration: adapting powerful models like LLMs to the unique demands of numerical sequences, and integrating deep learning with scientific principles. A key innovation comes from KAIST, Daejeon, Republic of Korea, with their paper, “Compositional Spectral Prompts for LLM-based Online Time Series Forecasting” which introduces CoSPOT. This framework tackles online time series forecasting in non-stationary environments by guiding frozen LLMs with learnable compositional spectral prompts. Instead of continuous fine-tuning or memory buffers, CoSPOT decomposes time series into frequency bases, allowing LLMs to generalize to unseen patterns by combining a finite set of learned frequency basis prompts. This is a game-changer for handling distribution shifts efficiently without retraining massive models.
Building on the LLM adaptation theme, researchers from the Department of Electrical and Computer Engineering, Isfahan University of Technology, Iran present PRICE in their paper, “PRICE: A Systematic Study of LLM Adaptation Choices for Bitcoin Price Forecasting”. They demonstrate that specific adaptation choices, often considered secondary, are critical for LLM performance in numerical forecasting. Their systematic study reveals that integer-rounded numerical representation dramatically reduces token fragmentation while preserving price scale, and recursive multi-step inference provides better accuracy for short horizons. This work highlights that general-purpose LLMs can even outperform specialized time-series foundation models with the right adaptation strategies.
Meanwhile, Sun Yat-sen University, Guangzhou, China, and others address the challenges of multi-agent news-driven forecasting with “CompEvo: Competition-Induced Evolution for Multi-Agent in News-Driven Time Series Forecasting”. CompEvo leverages evolutionary game theory to prevent “thought degeneration” where agents converge to similar behaviors. By formulating news-driven time series forecasting as an incomplete-information evolutionary game, agents are driven by competitive pressure to explore diverse evidence-seeking strategies, maintaining specialization and improving accuracy by a significant margin. This brings theoretical guarantees to multi-agent systems, ensuring robust and diverse forecasting behaviors.
Shifting to a more physics-guided approach, researchers from Wuhan University of Technology and Wuhan University introduce MZ-Rain in their paper, “MZ-Rain: Moisture-Budget-Guided Zero-Inflated Model for Station-Level Precipitation Nowcasting”. This framework integrates the moisture budget equation as a structural prior to decompose precipitation formation into specific pathways (moisture storage, transport, evaporation, persistence), each modeled by dedicated sLSTM branches. MZ-Rain also employs adaptive Tweedie modeling with auxiliary occurrence supervision to handle the severe zero-inflation and mixed discrete-continuous nature of precipitation data, achieving substantial improvements in precipitation nowcasting.
Finally, moving to the foundational architecture, Amirhosein Azarpour from the Department of Computer Science, Shahid Beheshti University, Tehran, Iran proposes RecKAN in “RecKAN: Kolmogorov-Arnold Networks with a Learnable Recursive Polynomial Basis”. This novel variant of Kolmogorov–Arnold Networks introduces a learnable second-order polynomial recurrence for generating basis functions. Unlike fixed bases, RecKAN dynamically adapts its functional capacity, even recovering classical polynomial families or evolving into new bases better suited for complex local structures, demonstrating superior performance across various benchmarks, including time series.
Under the Hood: Models, Datasets, & Benchmarks
These innovations are supported by a combination of new techniques and established resources:
- CoSPOT utilizes frozen Large Language Models (LLMs), specifically adapted with compositional spectral prompts, and demonstrates superior performance on real-world datasets in extended online phases and cross-dataset scenarios. The code is publicly available at https://github.com/seungyoon-Choi/CoSPOT.
- PRICE adapts a quantized LLaMA-3 model with techniques like LoRA for parameter-efficient fine-tuning, recursive multi-step inference, integer-rounded numerical representation, Context-Task-Format (CTF) prompting, and zero-temperature decoding. The work leverages the
unsloth/llama-3-8b-Instruct-bnb-4bitmodel and encourages exploration via https://huggingface.co/unsloth/llama-3-8b-Instruct-bnb-4bit and the https://github.com/thuml/Time-Series-Library. - CompEvo formulates multi-agent news-driven forecasting using evolutionary game theory principles with a three-stage gated generator for opponent-aware logic evolution. Their framework provides significant RMSE and MAPE reductions. Code can be found at https://github.com/Aliiina-z/CompEvo.
- MZ-Rain employs a physics-structured temporal framework with sLSTM branches and adaptive Tweedie modeling. It utilizes diverse meteorological data including the RainfallBench dataset, GNSS PWV data, GPM IMERG Final Precipitation Product, and ERA5 reanalysis products. A public code repository is available at https://anonymous.4open.science/r/MZ-Rain-7888.
- RecKAN introduces a learnable recursive polynomial basis for Kolmogorov-Arnold Networks, enhancing their functional capacity. The theoretical insights and experimental validations show superior performance over fixed-basis KANs and standard MLPs on image, text, and time-series benchmarks.
Impact & The Road Ahead
These advancements herald a new era for time series forecasting. The ability to effectively adapt LLMs for numerical prediction, as demonstrated by CoSPOT and PRICE, opens doors to leveraging vast pre-trained knowledge for diverse forecasting tasks without domain-specific pre-training. This could democratize high-performance forecasting. CompEvo’s game-theoretic approach provides a robust framework for multi-agent systems, which are increasingly vital in complex, dynamic environments like financial markets.
Furthermore, MZ-Rain showcases the critical importance of integrating scientific knowledge—like the moisture budget equation—into deep learning models for domains where physical processes are paramount. This physics-guided AI offers greater interpretability and robustness, especially in critical applications like climate modeling. Lastly, RecKAN’s innovation in learnable basis functions suggests a path toward more fundamentally adaptive and interpretable neural network architectures, promising better function approximation across all data types, including time series.
The road ahead involves further exploring the synergy between large foundation models and specialized, physics-informed architectures. We can anticipate more hybrid models that combine the generality of LLMs with the precision of domain-specific insights. The focus will be on even greater adaptability, robustness to unforeseen events, and the continued pursuit of interpretable, theoretically grounded solutions. The future of time series forecasting is not just about prediction; it’s about understanding and adapting to an ever-changing world with unprecedented intelligence.
Share this content:
Discover more from SciPapermill
Subscribe to get the latest posts sent to your email.
Post Comment