🤖 AI

Using LLMs for Explainable, Data-Driven Insight Generation from Time Series

This paper proposes a domain-agnostic framework that generates grounded, natural language explanations for time series forecasts by extracting structured factors from historical data and conditioning generation on verifiable evidence, thereby achieving analyst-level quality in readability, consistency, and persuasiveness without requiring domain-specific fine-tuning.

Original authors: Ria Mundhra, Gustavo Sato dos Santos, Michael Benedikt

Published 2026-07-22

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Ria Mundhra, Gustavo Sato dos Santos, Michael Benedikt

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are looking at a weather forecast that predicts a storm. A simple forecast might just say, "Rain tomorrow." But if you're planning a wedding or a picnic, you need more than just the prediction; you need the why. You need to know if the storm is coming from a cold front, if the humidity is unusually high, or if a specific mountain range is trapping the clouds. In the world of data science, this "why" is called an explanation. For years, computers have been getting really good at predicting numbers—like stock prices or shipping costs—but they have been terrible at explaining why those numbers are moving. They act like a magic 8-ball: they give an answer, but they can't tell you the story behind it.

This is where Large Language Models (LLMs) come in. Think of these as super-smart robots that have read almost everything ever written. They are great at telling stories and writing sentences that sound human. Scientists have been trying to teach these robots to look at a pile of numbers (like a time series) and write a story about what's happening. But there's a big catch: these robots are notorious for "hallucinating." If you ask them to explain a stock market drop, they might invent a fake news story about a CEO quitting, even if that CEO never left their job. They make things up because they are trying to be creative, not because they are looking at the actual evidence. The big question for scientists has been: Can we build a robot that tells a story about numbers without making things up?

The Story of the Truthful Time-Traveler

This paper introduces a clever new framework designed to teach AI how to write grounded, honest stories about time-based data (like stock prices or shipping rates). The authors, a team from the University of Oxford and Vortexa, realized that the problem isn't that the AI can't write; it's that the AI is trying to guess the story instead of reading the evidence. Their solution is like giving the AI a strict set of rules and a "cheat sheet" of real facts before it starts writing.

The Three-Step Recipe for Truth

The authors built a modular framework (a set of building blocks) that works in three main stages. You can think of it as a factory line for creating trustworthy explanations.

Step 1: The Detective's Notebook (Historical Factor Extraction)
First, the team looked at thousands of real explanations written by human financial experts. These experts don't just say "the market went up"; they say "the market went up because unemployment dropped and a new tech law passed." The researchers used an AI to read these human stories and pull out the "factors"—the specific reasons behind the trends. They turned these messy human sentences into neat, structured lists.

The Analogy: Imagine a detective who reads old police reports. Instead of just reading the story, the detective pulls out every specific clue (like "rainy night," "broken window," "muddy footprint") and writes them down on a standardized card. Now, instead of a long story, they have a box of verified clues.

Step 2: The Evidence Locker (Grounded Generation)
Next, when the AI needs to write a new explanation for a forecast, it doesn't just guess. It goes to its "Evidence Locker." It looks at the specific time period it's talking about and pulls out the relevant clues from Step 1. It also grabs real-time data, like unemployment rates or news headlines, and translates them into simple text summaries.

The Analogy: This is like a student taking a test. Instead of trying to memorize the whole textbook and hoping they remember the right fact, the teacher gives them a specific list of notes and a calculator. The student must use only those notes to write their answer. If the notes say "it rained," the student can't write "it was sunny." This stops the AI from making up fake reasons.

Step 3: The Strict Editor (Scalable Evaluation)
Finally, the team needed a way to check if the AI's stories were good without hiring a hundred human experts to read every single one. They built a system that acts like a strict editor. This system checks three things:

Readability: Is the story easy to read?
Consistency: Does the story contradict itself? (e.g., saying "interest rates went up" and then "that caused prices to go up" when usually high rates make prices go down).
Persuasiveness: Does the story sound convincing?

The Analogy: Imagine a robot editor that checks your essay. If you say "I ate a pizza" and then "I didn't eat lunch," the editor flags it as a contradiction. If you use fancy words correctly, it gives you a high score. If you make up facts, it gives you a failing grade.

What They Found

The researchers tested this system on two very different worlds: the NASDAQ-100 (a list of 100 big tech companies' stock prices) and freight pricing (the cost of shipping goods by sea).

The Results: They found that when they forced the AI to use the "Evidence Locker" and the "Strict Editor," the stories it wrote were almost as good as the ones written by human experts.
- Readability: The AI's stories were just as clear and professional as the human ones.
- Logic: The AI made fewer logical mistakes when it was forced to stick to the evidence.
- Persuasiveness: In head-to-head comparisons, the AI's explanations were often judged to be just as convincing as the human experts' explanations.
The "Magic" Ingredient: The study showed that the most important part wasn't just having a smart AI, but having the right process. When they removed the "Evidence Locker" (the part that translates raw numbers into simple text), the AI started making more mistakes and contradicting itself. When they added real news articles to the mix, the stories became more convincing.

What This Means (and What It Doesn't)

The paper suggests that we don't need to train a brand-new, super-expensive AI from scratch to explain time series data. Instead, we can take existing smart AI models and give them a structured way to look at evidence, much like giving a writer a set of facts before they start drafting a novel.

However, the authors are careful to note that this isn't a perfect, solved problem yet. Their system works well for the two specific areas they tested (stocks and shipping), but they aren't sure if it will work exactly the same way for every type of data in the world. Also, the "Strict Editor" they used is another AI, which is good for speed but might not be as sharp as a real human expert.

In short, this paper shows that if you give a language model a map of the evidence and tell it to stick to the path, it can tell a story that is both fluent and honest. It's a step toward making AI a reliable partner in decision-making, rather than just a creative storyteller that might lie to you.

Technical Summary: Using LLMs for Explainable, Data-Driven Insight Generation from Time Series

Problem Statement

Time series forecasts are critical in high-stakes domains such as finance and logistics, yet they are rarely consumed without accompanying explanations. Traditionally, generating these explanations is a manual, expensive, and non-scalable process performed by domain experts who synthesize numerical patterns with external context. While Large Language Models (LLMs) offer a path to automation, their direct application to time series data is fraught with challenges. Time series data is numerical, temporally structured, and exhibits complex behaviors (autocorrelation, seasonality, regime shifts) that LLMs are not inherently trained to reason over. Naive application of LLMs often results in "hallucinations," where models generate factually incorrect, temporally inconsistent, or data-unsubstantiated narratives. Furthermore, the field lacks a scalable, rigorous framework for evaluating the quality of such explanations, as generic text metrics (e.g., BLEU, ROUGE) correlate poorly with human judgment regarding logical consistency and persuasiveness.

Methodology

The authors propose a modular, domain-agnostic framework (illustrated in Figure 1) designed to generate grounded natural language explanations. The framework operates in three distinct stages:

1. Historical Factor Extraction

The first stage aims to learn the structure of expert reasoning without requiring labeled data.

Input: A corpus of historical analyst-written explanations (AnX).
Process: The system segments AnX into chunks and uses an LLM to extract structured "factors." A factor is defined as a tuple $(n, e, A)$ , where $n$ is a descriptive name, $e$ is the explicit evidence span from the text, and $A$ is a set of attributes (e.g., time horizon, sector affected, price impact).
Clustering: Extracted factors are embedded and clustered (using UMAP and HDBSCAN) to identify reusable explanatory themes (e.g., "Unemployment Rate and Jobless Claims").
Quality Validation: To ensure the extracted attributes are meaningful, the authors employ a retrieval-based diagnostic task. A reasoning model attempts to identify the source AnX given only a subset of factor attributes. This serves as a proxy to select the most "high-signal" attributes for downstream use, filtering out noise.

2. Grounded Explanation Generation

The second stage generates the explanation (GenX) by conditioning the LLM on verifiable evidence rather than latent knowledge.

Evidence Retrieval: Identified factor clusters are mapped to concrete evidence. This includes quantitative signals (e.g., aligned auxiliary time series like unemployment rates) and qualitative signals (e.g., news articles regarding geopolitical events).
Time Series Encoding: Raw numerical data is not fed directly to the LLM. Instead, it undergoes a multi-step encoding process:
1. Multi-scale Representation: Data is represented at daily, weekly, and monthly resolutions.
2. STL Decomposition: Seasonal-Trend decomposition separates trend, seasonality, and residuals.
3. Natural Language Summarization: Statistics are converted into concise text dictionaries describing direction, magnitude, and cycles.
Structured Prompting: The encoded evidence and external context are interleaved into a structured prompt. Few-shot examples are used to anchor tone and structure, but the model acts primarily as a linguistic realizer for the supplied evidence, minimizing hallucination.

3. Scalable Evaluation and Refinement

The third stage evaluates the generated explanations along three axes without relying solely on expensive human expert review:

Readability: Measured via the Gunning-Fog Index (general) and a "Smart Cloze" test (domain-specific), where an LLM fills in blanks to assess the predictability and appropriateness of specialized terminology.
Logical Consistency: A contradiction-detection procedure converts explanations into assertions and checks them against a set of domain axioms (e.g., "rising interest rates cause downward pressure on indices"). High-confidence contradictions are flagged.
Persuasiveness: An LLM-as-a-judge compares generated explanations against analyst-written references to determine which is more convincing, with controls for positional bias and superficial cues.
Refinement Loop: Contradictions detected during evaluation are fed back as constraints to the generator to iteratively improve the output.

Key Results

The framework was evaluated on two case studies: the NASDAQ-100 index (primary) and freight pricing (secondary, using Vortexa data).

Readability: Generated explanations (GenX) matched analyst-written explanations (AnX) in both general and domain-specific readability. The "Smart Cloze" evaluation showed narrow performance gaps, indicating that the framework preserves appropriate industry terminology.
Consistency: There was a significant difference based on the underlying LLM's reasoning capacity. Explanations generated by a lightweight model (gpt-4o-mini) exhibited higher rates of internal contradictions compared to AnX. However, a stronger reasoning model (o3) substantially narrowed this gap, approaching the consistency levels of human experts.
Persuasiveness: In pairwise comparisons, analyst-written explanations were generally preferred over gpt-4o-mini outputs. However, o3-generated explanations were judged as equally persuasive or preferred in the majority of cases against the human references.
Ablation Study: The study revealed that different components of the pipeline serve distinct functions:
- Time Series Encoding was the most critical factor for reducing internal contradictions and ensuring temporal consistency.
- External News Summaries primarily improved persuasiveness by grounding claims in real-world events.
- Few-Shot Examples had the strongest impact on stylistic alignment and domain-specific readability.
- The full configuration (all components enabled) performed strongly across all metrics, though trade-offs existed (e.g., context saturation slightly affecting individual metric peaks).

Significance and Claims

The paper claims that grounded explanation generation for time series forecasting can be achieved at scale without domain-specific fine-tuning.

By explicitly constraining LLM generation to verifiable evidence derived from structured factor extraction and compact time series encoding, the framework successfully mitigates the hallucination problems typically associated with applying LLMs to numerical data. The authors demonstrate that their approach produces explanations that are comparable to expert-written narratives in terms of readability, logical consistency, and persuasiveness.

The framework is presented as a general blueprint for transparent, evidence-backed decision support in diverse fields beyond finance, including logistics, supply-chain forecasting, epidemiological surveillance, and infrastructure monitoring. The authors remain modest, noting that the evaluation relies on the availability of high-quality historical signals and that LLM-based judges, while scalable, are imperfect proxies for human expertise. Future work is suggested to include reinforcement-based fine-tuning and adaptive attribute selection.