Learning to select, evaluate, and interpret models as evidence-building tools that balance clarity, uncertainty, and real-world relevance.
Throughout this module, we continue working with our running example of investigating electrical fire incidents, now focusing on how models help us move from patterns to evidence-based claims.
After exploring and visualizing data, the next step is to select models that help answer your investigative question.
A model is not just an equation, it is a simplified representation of reality that helps you explain patterns, test claims, or make predictions while acknowledging variability and uncertainty.
In the context of electrical fire investigations, models help move from “what we see” to “what evidence supports a claim.”
In this module, models include:
Not every model you try will be useful. A key part of this phase is choosing which models to keep, and which to discard.
Modeling rarely happens in a straight line. Often, when testing a model, you realize that:
Going back to earlier phases is expected—and productive.
The following six-step workflow guides the process of selecting, evaluating, and iterating on models for data investigations.
Begin by restating your investigative question and asking what kind of evidence would meaningfully address it.
Electrical fire examples:
This distinction matters because different goals require different models.
Prompting technique: Zero-shot prompting
“Given this investigative question about electrical fires, is the goal better described as inference, explanation, or prediction? Explain why.”
Identify several possible models that could address the question, knowing you may discard some later.
Examples for electrical fire data:
Prompting technique: Few-shot prompting
“Here is my data and my research questions
Based on those, suggest appropriate models for investigating electrical fire causes.”
Test each model by asking:
Models that do not add insight are set aside, not forced into the analysis.
Prompting technique: Chain-of-thought prompting
“Step by step, evaluate whether this model provides meaningful evidence for the research question about electrical fire risk.”
No model perfectly represents reality. So, eexamine:
This step helps avoid overconfident claims.
Prompting technique: Self-consistency
“Analyze this model’s results twice [insert], focusing once on patterns and once on uncertainty.
Do both interpretations support the same conclusion?”
Decide how much interpretability matters for the investigation.
For electrical fire analysis:
Simple, interpretable models (e.g., averages, linear trends) often matter more than high-performance black-box models.
However, more complex models can sometimes reveal upper bounds or hidden structure.
Choose models that align with the purpose of the investigation, not just technical sophistication.
Prompting technique: Prompt chaining
Prompt 1:
“Explain this model in plain language for a community safety report.”
Prompt 2:
“Now explain what this model cannot tell us.”
If modeling reveals gaps or limitations, return to earlier phases:
Modeling informs what to do next, not just what to conclude.
Models are evidence-building tools, not final answers. In electrical fire investigations, choosing the right model means balancing clarity, uncertainty, and real-world relevance.
LLMs can assist by:
But deciding which models matter and why is an essential part of data sensemaking, and remains the your responsibility.
Here are some more articles you might like to read next: