Story Is All You Need | Lin Liu, Wealthsimple

Story Is All You Need | Lin Liu, Wealthsimple

by Lin Liu

Applied Sciences & Engineering Literature & Writing DNBiography and non-fiction proseDNPReportage, journalism or collected columns
🎙 Lin Liu 👥 5K 📅 October 20, 2025 ⏱ 23 min 👁 67 📄 expert opinion 🧭 2026-08-15
Available in: English (current) Français

Keywords

data-as-storyLLMpredictive modelingfeature engineeringMLOps

Summary

Lin Liu, Director of Data Science at Wealthsimple, presents a novel approach to predictive analytics called ‘data-as-story’. Instead of traditional feature engineering, raw structured data (e.g., transaction logs) is transformed into coherent text narratives, which are then fed directly to LLMs to generate predictions. The talk illustrates this with examples like identifying parents from credit card transactions and predicting credit default. Liu claims this method can outperform rule-based systems and traditional ML models, citing a 40% lift in conversion rates for marketing audiences. The approach leverages LLMs’ common sense and external knowledge, reducing the need for labeled data and extensive feature engineering. However, Liu acknowledges limitations: data privacy concerns, hallucination risks, and the need for careful prompt engineering. He recommends using this method for low-risk decisions like marketing, and suggests combining it with traditional ML for better results. The talk includes practical implementation tips, such as batching stories to avoid LLM counting errors and fine-tuning models for specific domains. Overall, the presentation offers a creative and promising direction for applying LLMs in predictive tasks, but it is based on anecdotal evidence and lacks rigorous validation.

185 words

Critical Evaluation

Value of the Information & Strength of the Argument

The talk introduces a genuinely novel and creative approach to predictive modeling, which is valuable for practitioners. The argumentation is coherent: Liu identifies a common problem (lack of labels, cumbersome feature engineering) and proposes a solution that leverages LLMs’ strengths. He supports his claims with a concrete example (parent detection) and a case study (marketing conversion lift), though the latter is only briefly mentioned without detailed methodology. The reasoning is logical, but the evidence is largely anecdotal and lacks rigorous statistical validation. The speaker also honestly discusses limitations and risks, which strengthens the credibility of the presentation.

Scientific Rigor, Source Quality, Title Accuracy

The talk is based on the speaker’s professional experience and internal projects at Wealthsimple, but no external sources or references are cited. The description provides a link to the MLOps World conference, but no specific papers or studies are mentioned. The title ‘Story Is All You Need’ is catchy and accurately reflects the core concept, though it may overstate the universality of the approach. The content is presented as an expert opinion rather than a peer-reviewed study, so the scientific rigor is moderate. The speaker does not provide detailed metrics or comparisons, which limits the ability to assess the effectiveness of the method.

215 words

Title / Content Match

The title 'Story Is All You Need' is catchy and accurately reflects the core concept of using narrative data with LLMs for prediction.

Quality & Reliability

7/10

The talk presents a novel approach with practical examples and acknowledges limitations, but lacks rigorous empirical validation and detailed methodology.

Key Moments

Markers derived by PSI from the transcript: the creator did not define chapters.

Cited Sources

  • MLOps World Conference — The talk was presented at MLOps World | GenAI Summit 2025, and the link provides more information about the conference.

Concurring Sources

  • MLOps World Conference — The talk was presented at this conference, which focuses on MLOps and GenAI, aligning with the topic.

Contribution & Novelties

The talk introduces a novel paradigm for predictive analytics, ‘data-as-story’, which replaces traditional feature engineering with narrative text fed to LLMs. This approach leverages LLMs’ common sense and external knowledge, potentially reducing the need for labeled data and complex feature engineering. The speaker provides practical insights from real-world applications at Wealthsimple, such as a 40% lift in marketing conversion rates. The method is particularly useful for low-risk decisions and rapid prototyping. However, the approach is still nascent and requires careful consideration of privacy, hallucination, and cost.

Pour aller plus loin :

  • Large language model — Overview of LLMs, their capabilities, and limitations.
  • Prompt engineering — Techniques for crafting effective prompts to improve LLM outputs.
  • Feature engineering — Traditional approach to predictive modeling, contrasted with the data-as-story method.
  • MLOps — Practices for deploying and maintaining machine learning models in production, relevant to the implementation discussion.

144 words

Radar Profile

The radar profile shows moderate scores across all dimensions, with slightly higher scores in information quantity and quality, but lower in technical level and reliability. This reflects a talk that is informative and practical but lacks deep technical detail and rigorous validation.

Reliability 6/10

💬 No comments were provided for analysis.