Technical Summary: MemCast – Memory-Driven Time Series Forecasting with Experience-Conditioned Reasoning

1. Problem Definition

Time series forecasting (TSF) is critical for decision-making in domains such as energy scheduling, financial trading, and healthcare. While recent Large Language Model (LLM)-based approaches have shown promise, existing methods suffer from two primary limitations:

1. Lack of Explicit Experience Accumulation: Most methods treat each forecasting instance as an isolated reasoning process. They fail to explicitly convert historical prediction outcomes, reasoning trajectories, and temporal regularities into reusable forecasting experience, hindering knowledge transfer across instances.
2. Absence of Continual Evolution: Current models lack mechanisms for lifelong improvement. Stored experience is not adaptively refined as new forecasting instances are observed, limiting the model's ability to evolve in non-stationary environments.

Traditional training-based methods (fine-tuning) often rely on small-scale LLMs due to computational constraints, while training-free methods (prompting) rely heavily on static prompts without accumulating dynamic experience.

2. Methodology: MemCast Framework

MemCast proposes a learning-to-memory framework that reformulates TSF as an experience-conditioned reasoning task. Instead of updating LLM parameters via backpropagation, MemCast accumulates structured forecasting experience in an external hierarchical memory during the training phase and utilizes this memory to guide reasoning during inference.

2.1. Hierarchical Memory Construction (Training Phase)

The framework extracts and organizes experience from the training set into three distinct layers:

• Historical Pattern Abstraction: Raw numerical time series are converted into compact semantic summaries using a summarization LLM. These summaries capture trend evolution, volatility, and peak values, creating a textual memory ($M_{his}$) that serves as concrete analogical references for rare scenarios or distribution shifts.
• Reasoning Wisdom Distillation: Inference trajectories (Chain-of-Thought) are analyzed and partitioned into successful and failed cases based on prediction errors. A filtering mechanism (combining semantic similarity and Dynamic Time Warping) distills these into reusable "wisdom" ($M_{rea}$), separating successful strategies from failure modes to guide future trajectory selection.
• General Law Induction: To prevent deviations from physical commonsense, the framework leverages LLMs to analyze extracted temporal features and induce general laws ($M_{gen}$). These laws act as domain-specific constraints (e.g., non-negativity, range bounds) for reflective iteration.

2.2. Experience-Conditioned Reasoning (Inference Phase)

During inference on the testing set, the frozen LLM utilizes the constructed memory through three mechanisms:

• History-Enhanced Reasoning: The system retrieves top-$k$ relevant historical patterns to augment the context, transforming the inference probability from $P(Y|X)$ to $P(Y|X, C_{sim})$. This grounds reasoning in similar past instances.
• Wisdom-Driven Trajectory Exploration: To address uncertainty, the model generates multiple candidate reasoning trajectories. A scoring function evaluates the semantic consistency of each trajectory against the retrieved "reasoning wisdom" to select the most reliable path, filtering out stochastic noise.
• Law-Based Reflection: Predicted sequences are evaluated against the induced general laws. If a prediction violates a constraint (e.g., physical limits), a corrective loop is triggered, feeding the violation details back to the LLM for self-correction until constraints are met.

2.3. Dynamic Confidence Adaptation

To enable continual evolution without test set distribution leakage, MemCast employs a dynamic confidence adaptation strategy. Instead of modifying memory content (which could bias the model), the system updates the confidence scores of individual memory entries.

• Upon observing ground truth, the system compares the forecast against a moving average baseline.
• If the forecast is successful ($L_{LLM} < L_{MA}$), the confidence scores of the contributing experiences are increased.
• This allows the system to adaptively refine the reliability of stored experience over time while maintaining a strict separation between training and test distributions.

3. Key Contributions

• Reformulation of TSF: The paper proposes shifting the paradigm from isolated instance prediction to experience-conditioned reasoning, explicitly accumulating and reusing forecasting experience.
• Hierarchical Memory Design: The introduction of a three-tier memory structure (Historical Patterns, Reasoning Wisdom, General Laws) that addresses different aspects of reasoning: instance-level analogies, trajectory-level logic, and global physical constraints.
• Continual Evolution without Retraining: The dynamic confidence adaptation strategy allows the model to evolve and improve its reliability on the fly without retraining parameters or leaking test data.
• Robustness to Distribution Shifts: By leveraging historical patterns and physical laws, the framework is designed to handle out-of-distribution (OOD) scenarios and non-stationary environments more effectively than standard attention-based or fine-tuned models.

4. Experimental Results

Extensive experiments were conducted on 10 diverse real-world datasets (including electricity prices, transformer temperatures, wind/solar power, and hydrological data) covering both short-term and long-term forecasting.

• Performance: MemCast consistently outperformed state-of-the-art baselines, including statistical methods (ARIMA, Prophet), deep learning models (PatchTST, iTransformer, DLinear), and other LLM-based approaches (Time-LLM, TimeReasoner, LSTPrompt). It achieved the lowest Mean Squared Error (MSE) and Mean Absolute Error (MAE) across most benchmarks.
• Ablation Studies:
  • Removing any of the three memory components (Patterns, Wisdom, Laws) resulted in performance degradation, confirming the necessity of the hierarchical structure.
  • Dynamic Confidence Adaptation: Variants without this strategy showed higher errors, particularly on volatile datasets, proving that updating entry reliability stabilizes reasoning.
  • Trajectory Exploration: Active selection of trajectories based on semantic consistency outperformed passive aggregation (mean ensemble) and single-pass generation.
• Scalability: The framework demonstrated backbone-agnostic capabilities, performing well across different LLMs (GPT-4o-mini, Gemini, DeepSeek), indicating that the memory mechanism effectively leverages the reasoning capacity of various models.

5. Significance and Claims

The paper claims that MemCast offers a promising step toward experience-driven learning for time series forecasting. Its significance lies in:

• Interpretability and Reliability: By externalizing reasoning traces and enforcing physical laws, the model mitigates generative hallucinations and provides controllable, interpretable predictions.
• Adaptability: The ability to evolve confidence weights allows the system to adapt to changing temporal patterns without the computational cost of retraining or the risk of data leakage.
• Paradigm Shift: It validates the hypothesis that treating TSF as a reasoning task conditioned on accumulated experience yields superior results compared to treating it as a pure pattern-matching or parameter-fitting problem.

The authors conclude that explicit experience abstraction and controlled memory utilization are critical for accurate forecasting in complex, real-world scenarios.