TS-Reasoner: Structured Reasoning for Complex Time Series Tasks

While most time series models aim for better accuracy in prediction or anomaly detection, TS-Reasoner redefines what it means to “understand” time series. TS-Reasoner is a compositional reasoning framework that goes far beyond traditional forecasting — enabling multi-step analysis, decision-making, and causal reasoning through structured LLM-guided programs.

This is not just another time series model. It’s a general-purpose reasoning engine for temporal data.

🔍 Motivation

Traditional models are excellent at single-shot predictions—but fall short when tasks become multi-faceted:

• “What caused the sales drop after a campaign?”
• “Will an intervention work based on past conditions?”
• “How do multiple signals jointly impact a future event?”

These types of compositional, constrained questions require breaking down the problem and combining multiple reasoning steps — precisely what TS-Reasoner is built for.

🧱 Framework Overview

TS-Reasoner is built on three key pillars:

1. Programmatic Task Decomposition

   TS-Reasoner uses LLMs to convert natural language queries into structured reasoning programs — sequences of modular operations that can include forecasting, comparison, filtering, and more.

2. Custom Reasoning Modules

   The framework includes plug-and-play modules for tasks such as:

   • Forecasting via time series models (e.g., PatchTST)

   • Constraint checking (e.g., “satisfy threshold X”)

   • Domain-based filtering or ranking

     Each module operates on temporal inputs and outputs, enabling intermediate supervision and traceable logic.

3. LLM-Guided Execution with Traceability

   A controller model interprets and executes the reasoning program step by step, optionally returning execution traces and natural language rationales, which makes the entire reasoning process interpretable and verifiable.

🧪 Real-World Use Cases

TS-Reasoner unlocks capabilities that are hard to achieve with conventional models: