CastFlow: A Role-Specialized Agent Workflow for Time Series Forecasting

CastFlow is a dynamic agent forecasting framework that resolves the static generation limitations of traditional LLM forecasting methods through role-specialized design and a four-stage agent workflow. Its core advantages include multi-perspective time-series pattern extraction, multi-round context acquisition, iterative prediction optimization, and integrated prediction support. It aims to transform large language models from simple generation tools into agents capable of planning, acting, and reflecting, thereby enhancing the accuracy and reliability of time series forecasting.

Large language models have great potential in the field of time series forecasting, but traditional methods have four major limitations:

1. Limited time-series pattern extraction: Single-generation is difficult to capture complex periodic, trend, and seasonal patterns, and easily misses deep dynamic dependencies;
2. Single context feature acquisition: Only single-round acquisition of external information (such as holidays, weather) without dynamic adjustment of needs;
3. Lack of iterative optimization: No analysis-prediction-reflection-correction process like human experts;
4. No integrated prediction support: Failure to leverage the more robust advantages of multi-model integration.

CastFlow's core architecture is a four-stage agent workflow:

• Planning: Analyze task characteristics and formulate forecasting strategies (identify sequence types, determine context requirements, plan time span);
• Action: Execute strategies, call tools to obtain information, query similar cases, or retrieve past experiences;
• Prediction: Generate numerical predictions based on the integrated forecasting baseline;
• Reflection: Evaluate the rationality of results, and loop back for correction if problems are found. Support modules include:
• Memory Module: Store/retrieve past forecasting experiences to enable experience reuse;
• Multi-Perspective Toolkit: Provide statistical models (e.g., ARIMA), machine learning models (e.g., XGBoost), and combination strategies to build diagnostic evidence and integrated baselines.

CastFlow innovatively adopts a role-specialized design:

• Frozen General LLM: Responsible for general reasoning (understanding requirements, formulating strategies, explaining results, generating reports), retaining general knowledge and reasoning capabilities;
• Fine-tuned Domain-Specific LLM: Responsible for numerical prediction, performing evidence-guided generation based on the integrated baseline to ensure professional accuracy. The division of labor is clear, balancing general intelligence and forecasting professionalism.

To optimize the domain-specific LLM, a two-stage training approach is adopted:

1. Supervised Fine-Tuning (SFT): Train with labeled data to master basic forecasting skills and workflow norms;
2. Verifiable Reward Reinforcement Learning (RLVR): Further optimize model accuracy based on reward signals from comparing prediction results with real values.

Experiments validate the generalization ability on multi-domain datasets (energy consumption, traffic flow, financial markets, meteorological data, etc.), and the results show:

• Outperforms traditional statistical methods and pure neural network methods;
• Superior to other LLM-based forecasting schemes, proving that the paradigm upgrade of the agent workflow brings actual performance improvements.

Implications of CastFlow:

1. From single-generation to iterative optimization: Forecasting becomes a dynamic improvement process, suitable for high-reliability scenarios (e.g., energy scheduling, supply chain planning);
2. From single model to integrated intelligence: Mix the advantages of traditional statistics, machine learning, and LLMs to build a collaborative ecosystem;
3. From general to specialized: Combining general reasoning with specialized skills is a feasible path for complex task automation. In the future, advances in agent technology will drive time series forecasting to play a value in more key areas.