arXiv Science⌕ Search

arXiv subjects

Bokai Pan

Publications and source records attributed to Bokai Pan.

6 recordsLinked to original sources

CastWorld: Benchmarking Context-Aware Agentic Time Series Forecasting Beyond Predictive Accuracy

Agentic time series forecasting extends conventional approaches beyond one-shot numerical prediction toward autonomous, context-aware interaction. However, existing benchmarks offer limited integrated support for diverse contextual information, forecasting-specific adaptive interactions, and systematic evaluation of forecasting agents. In this paper, we introduce CastWorld, an agentic benchmark for context-aware time series forecasting. First, its context-rich dataset suite spans 8 domains and 25 real-world datasets, covering external variables, events, and historical cases under a unified context specification. Second, a lightweight, modular forecasting-native harness provides 49 core tools and a state-adaptive loop that tracks forecast versions and enables iterative evidence gathering and forecast revision. Intermediate forecasts can expose information gaps, while newly acquired evidence guides subsequent revisions and final version selection. Third, model-level evaluation assesses forecast quality, process execution, behavioral trustworthiness, and resource efficiency through decision trajectories and forecast-version histories. Also, systems-level evaluation compares predictive accuracy across forecasting systems. Experiments reveal distinct decision behaviors among agents with comparable accuracy. Additional interactions do not necessarily improve forecasts, and agents may fail to select their best previously generated predictions. Allowing evidence acquisition during forecast refinement improves task-aware forecast scores across all three evaluated models. CastWorld provides a reusable foundation for evaluating agentic time series forecasting under controlled conditions.

cs.LG↗

CastFSR: Fast--Slow--Reflect Agentic Reasoning for Context-Aware Time Series Forecasting

Time series forecasting is fundamental to temporal data mining, yet numerical models often struggle to account for changing contextual factors. A key challenge is determining when contextual evidence should guide forecast revisions while maintaining consistency with temporal patterns and domain knowledge. To address this, we propose CastFSR, a fast--slow--reflect agentic framework for context-aware time series forecasting. Fast-thinking forecasting analyzes numerical patterns, selects a suitable forecaster, and constructs a data-driven forecast prior. Slow deliberative reasoning incorporates relevant contextual evidence to selectively refine this prior, while reflective evaluation checks the revised forecast against temporal patterns, contextual evidence, and domain constraints. A shared state maintains forecasting priors, contextual evidence, intermediate forecasts, and feedback across stages. CastFSR supports both training-free LLM coordination and two-stage post-training for compact LLM coordinators. Workflow-aware trajectory distillation (WATD) learns state-conditioned tool use and forecasting decisions from teacher trajectories, while trajectory-level policy optimization (TLPO) further improves the coordinator through reinforcement learning guided by forecasting accuracy and workflow feedback. Experiments on multiple real-world datasets under an offline evaluation protocol using historical context demonstrate competitive performance against representative baselines. Our code is available at https://github.com/Xiaoyu-Tao/CastFSR.

cs.AI↗

CastFlow: Bridging Evidence Acquisition and Accurate Prediction in Agentic Time Series Forecasting

Time series forecasting is a fundamental task in temporal data analytics, and large language model (LLM)-based agents offer a promising paradigm through autonomous tool interaction and evidence-grounded reasoning. However, two fundamental challenges remain: how to acquire forecast-relevant evidence through adaptive tool interactions, and how to translate such evidence into accurate numerical predictions. To address these challenges, we propose CastFlow, a role-specialized agentic forecasting framework that bridges general-purpose tool reasoning and specialized temporal forecasting. Specifically, CastFlow introduces an Experience Memory that distills successful tool-use trajectories from training data, guiding a frozen LLM to orchestrate a multi-view toolkit for evidence acquisition. A lightweight specialized LLM then performs evidence-grounded temporal reasoning to refine an ensemble forecasting prior. To enhance this capability, we develop a two-stage training strategy combining supervised fine-tuning and reinforcement learning with verifiable rewards, where final forecasting outcomes of the overall workflow guide the optimization of the specialized model. Extensive experiments on ten real-world benchmarks demonstrate that CastFlow outperforms competitive baselines on nine datasets. Our findings highlight the importance of bridging autonomous evidence acquisition and specialized temporal reasoning for accurate and adaptive agentic forecasting. The source code is available at https://github.com/Forever-Pan/CastFlow.

cs.LG↗

CastClaw: A Human-in-the-Loop Autonomous Agent for Interactive Time Series Forecasting

Practical time-series forecasting requires domain knowledge, constraints, plausibility checks, and human intervention. We present CastClaw, an interactive forecasting system centered on an autonomous forecasting loop and supported by a forecasting-oriented harness. Users create and refine tasks in natural language through command-line and Web interfaces. An internal large language model (LLM) agent reads the task and evidence, selects a permitted tool action or request for human feedback, and uses checked results to decide whether to continue, revisit an analysis, switch a candidate, or stop. The harness supplies tool integration, controlled execution, state management, historical validation, and execution limits. npm distribution connects installation and configuration to interactive execution and forecast and report export. Across five public electricity-price datasets, CastClaw achieves the lowest point-estimate MSE and MAE among 16 baselines. A Nord Pool case study illustrates its inspectable decision process, while provincial electricity-load data from North China (January--June 2026) demonstrate offline task execution.

cs.LG↗

AlphaCast: A Human Wisdom-LLM Intelligence Co-Reasoning Framework for Time Series Forecasting

Time series forecasting is typically treated as a single-pass numerical prediction task based on historical observations. However, real-world forecasting often requires expert-like reasoning, where temporal patterns, contextual cues, case-based references, and domain knowledge are jointly considered and iteratively revised. Existing statistical, deep learning, foundation, and LLM-based forecasting models mainly rely on pattern fitting, direct prompting, or static context augmentation, leaving interactive and reflective forecasting insufficiently explored. In this work, we propose AlphaCast, a human wisdom-LLM intelligence co-reasoning framework for interactive time series forecasting. AlphaCast does not require real-time human intervention; instead, it operationalizes human wisdom as structured forecasting support, including time-series features, contextual cues, case-based references, and domain knowledge. Built upon an expert-like think-forecast-reflect workflow, AlphaCast comprises three agents: an Investigator that constructs forecasting context from available evidence, a Generator that produces context-grounded forecasts, and a Reflector that evaluates plausibility and provides feedback for refinement. Extensive experiments demonstrate that AlphaCast improves forecasting accuracy and produces interpretable forecast explanations, offering a promising step toward interactive and context-aware time series forecasting.

cs.AI↗

From Values to Tokens: An LLM-Driven Framework for Context-aware Time Series Forecasting via Symbolic Discretization

Time series forecasting plays a vital role in supporting decision-making across a wide range of critical applications, including energy, healthcare, and finance. Despite recent advances, forecasting accuracy remains limited due to the challenge of integrating historical numerical sequences with contextual features, which often comprise unstructured textual data. To address this challenge, we propose TokenCast, a large language model (LLM) driven framework that leverages language-based symbolic representations as a unified intermediary for context-aware time series forecasting. Specifically, TokenCast employs a discrete tokenizer to transform continuous numerical sequences into temporal tokens, enabling structural alignment with language-based inputs. To effectively bridge the semantic gap between modalities, both temporal and contextual tokens are embedded into a shared representation space via a pre-trained LLM, further optimized with generative objectives. Building upon this unified semantic space, the aligned LLM is subsequently fine-tuned in a supervised manner to predict future temporal tokens, which are then decoded back into the original numerical space. Extensive experiments on real-world datasets demonstrate the effectiveness of our framework and highlight its potential as a generative framework for context-aware time series forecasting. The code is available at https://github.com/Xiaoyu-Tao/TokenCast.

cs.LG↗