Research map / Time series / Foundation models for time series

TimeClaw: A Time-Series AI Agent with Exploratory Execution Learning

NeurIPS 2026 · Time series › Foundation models for time series · approach: LLM-based time series forecasting · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes TimeClaw, a frozen-LLM time-series agent that learns from metric-compared exploratory executions, uses task-aware tool dropout against tool-prior collapse, and reinjects hierarchical distilled experience at inference on MTBench.

Problem

Time-series LLM agents optimize execution-time capability and learn little from exploratory execution; completion-only supervision ignores quantitative differences among valid executions, and exploration suffers from premature tool-usage concentration.

Approach

An Explore-Compare-Distill-Reinject loop where valid candidate executions are ranked by task metrics (MAE/MSE), tool usage is controlled by frequency-aware dropout during exploration, and distilled memory/tool/skill artifacts are injected into prompts at inference while the base model stays frozen.

Key results

• TimeClaw achieves best results on 27 of 30 reported metrics across 17 MTBench tasks, per the paper's claim.
• On weather forecasting 7D MSE, TimeClaw reaches 10.719 vs 14.197 for GPT-5.4 (lower is better).
• On finance trend 7D 5-way accuracy, TimeClaw reaches 0.696 vs 0.569 for GPT-5.4.

Surprising findings

• The paper reports TimeClaw is worse than GPT-5.4 on 30-day finance MACD MSE and on long-horizon finance correlation, where Llama-3 and GPT-5.4 lead respectively.

Limitations

• Evaluated only on a fixed 17-task MTBench suite covering finance and weather.
• Exploration is token-expensive, generating multiple candidate paths per sample.
• Weaker on some long-horizon finance tasks (MACD 30D MSE, correlation long-horizon).

Experimental setup

Datasets: MTBench separately generated MTBench-style exploration corpus

Benchmarks: MTBench

Metrics: MAE MAPE MSE RMSE accuracy

Baselines: GPT-5.4 GPT-4o Gemini Claude DeepSeek Qwen-3 Llama-3 w/o Exp. ablation

Builds on: Toolformer ToolLLM MemSkill ASDA Skill-SD

Closest papers by meaning

• ForecastCompass: Guiding Agentic Forecasting with Adaptive Factor Memory (NeurIPS 2026 · similarity 0.87)
• TimeXL: Explainable Multi-modal Time Series Prediction with LLM-in-the-Loop (NeurIPS 2025 · similarity 0.86)
• TimeCAP: Learning to Contextualize, Augment, and Predict Time Series Events with Large Language Model Agents (AAAI 2025 · similarity 0.86)
• TimePerceiver: An Encoder-Decoder Framework for Generalized Time-Series Forecasting (NeurIPS 2025 · similarity 0.86)
• TIMING: Temporality-Aware Integrated Gradients for Time Series Explanation (ICML 2025 · similarity 0.85)
• Rethinking Post-Training Recipes for Multimodal Time-Series Forecasting (NeurIPS 2026 · similarity 0.85)