# Brainstorm and prioritize features for a predictive model

A ranked list of candidate features grounded in the problem, with leakage checks, build effort, and expected signal.

- Topic: Data & Analysis
- Framework: Few-Shot
- Level: Standard
- Tags: Productivity, Write, Brainstorm, Fill-in Template
- Source: https://www.thinkwithprompt.com/prompts/brainstorm-and-prioritize-features-for-a-predictive-model

## Prompt

```text
Help me brainstorm and prioritize features for a predictive model.

**What the model predicts:** [target definition, unit, prediction time, outcome window]
**Available data sources:** [PASTE table names and key columns]
**Model type planned:** [e.g. gradient boosting, logistic regression, not decided]
**Current features (if any) and performance:** [LIST, with metric]

Steps:
1. Generate candidate features grouped by theme (e.g. recency, frequency, monetary value, engagement trends, tenure, product usage, support interactions, seasonality, external factors). For each, write the exact calculation and time window.
2. Include change-over-time features (last 7 days vs. prior 30), ratios, and counts of events, not just raw values.
3. **Leakage check:** flag any feature that uses information not available at prediction time, or that is a proxy for the target.
4. **Fairness check:** flag features that are, or proxy for, protected characteristics, and note whether they should be excluded.
5. Rate each feature: expected signal (high / medium / low, with reasoning), build effort, and data quality risk.

Return:
- A feature table: Feature | Calculation | Window | Source | Expected signal | Effort | Leakage risk | Notes.
- The top 10 to build first.
- Encoding and handling notes: missing values, high-cardinality categories, outliers, scaling if needed.
- How to test whether new features actually help (ablation, feature importance with caution, time-based validation).
```

## What to fill in

- target definition, unit, prediction time, outcome window
- PASTE table names and key columns
- e.g. gradient boosting, logistic regression, not decided
- LIST, with metric

Free to copy and use. From ThinkWithPrompt (https://www.thinkwithprompt.com/), a library of 2,663 curated AI prompts.