Asking Research Questions
State a hypothesis. Lomita finds the data, runs the analysis, and delivers a verdict.
What happens when you ask a question
When you type "Research whether ERCOT energy demand correlates with Dallas weather" — either in the chat or via your AI agent — here's what happens behind the scenes:
1. Discovery (30-60 seconds)
The Discovery Agent searches for relevant data:
- Scans the 104-source catalog for matches (ERCOT, NOAA weather, EIA energy)
- Searches the internet via Exa for APIs not in the catalog
- Connects the best sources to your hypothesis
- Documents any gaps (e.g., "Uber ride data not publicly available — using TomTom traffic as proxy")
2. Integration (1-2 minutes)
The Integration Engineer builds data pipelines:
- Reads API documentation for each source
- Writes and deploys ingestion scripts (Dagster assets)
- Handles authentication and scheduling
- Verifies data is flowing into the data lake
3. Analysis (1-2 minutes)
The Quant Analyst runs statistical tests:
- Pearson and Spearman correlations
- Granger causality (does X predict Y?)
- Lag analysis (how many days/weeks of lead time?)
- Regime analysis (does the relationship change in different conditions?)
- OLS regression with confidence intervals
4. Report (30-60 seconds)
The Research Narrator compiles findings:
- Executive summary with a clear verdict
- Key findings table with statistics
- Data sources used with coverage dates
- Methodology description
- Limitations and caveats
- Recommendation
Total time: 3-10 minutes depending on complexity.
Verdicts
| Verdict | What it means |
|---|---|
| Supported (green) | Statistical evidence supports your hypothesis |
| Inconclusive (yellow) | Some evidence, but not enough for a definitive answer |
| Refuted (red) | Data contradicts your hypothesis |
Every verdict comes with the methodology to prove it. You can audit the statistics, check the data sources, and verify the reasoning.
Follow-up research
After a hypothesis completes, you can:
- Follow up — "Break this down by season" creates a child hypothesis that branches from the original
- Deliver — Send the report to email, Slack, or any webhook
- Monitor — "Watch this weekly and alert me if the correlation shifts"
- Connect more data — "Connect my internal sales data and re-analyze"
Tips for good questions
Good questions (specific, testable):
- "Does consumer sentiment predict retail sales with a lag?"
- "Are ERCOT energy prices affected by Dallas weather extremes?"
- "Does VIX volatility correlate with S&P 500 drawdowns?"
Weaker questions (too broad):
- "Tell me about the economy" — too vague, no testable hypothesis
- "What should I invest in?" — Lomita analyzes data, it doesn't give investment advice
The system works best with questions that have a clear independent variable and dependent variable.