Back to module
Step 6-8 min

Ask the Real Analyst Question

Put the stack to work. Combine FMP fundamentals, Yahoo prices, and FRED macro to answer the kind of question analysts actually get paid for.

What you'll do

Ask your AI analyst the signature question this module was built for - one that requires all three data sources at once. Then learn the habits that get you consistently good answers.

Why this step matters

The whole point of the last five steps - the CLI, the MCP, the ingestion, the staging + reports layers, the context file - was to make this step possible: asking a question that used to take a junior analyst a week, and getting a worked answer in 30 seconds.

You built reports.ticker_quarterly in Step 4 specifically so questions like this one don't require re-deriving TTM FCF margin or aligning fundamentals-to-prices every session. Your AGENTS.md tells the agent to default to reports.* - so this question is mostly a SELECT against a pre-joined table.

The signature question

Paste this into your AI tool:

AI Prompt

Using reports.ticker_quarterly, find tickers (from the 5 I ingested) that meet all of the following over the last 4 quarters available in the table:

  1. ttm_fcf_margin has improved quarter over quarter (positive fcf_margin_qoq_change in each of the last 4 quarters, or strictly increasing ttm_fcf_margin)
  2. quarterly_return has underperformed the equal-weighted average across the 5 tickers by more than 10 percentage points cumulatively over the same period

Then split the qualifying set by macro regime using regime_at_qend:

  • 'inverted' - yield curve was inverted at the quarter end
  • 'normal' - yield curve was positive at the quarter end

For each ticker that qualifies, show: ticker, the 4 quarterly ttm_fcf_margin values, cumulative quarterly_return vs. the peer average, and which regime dominated the window.

Default to reports.ticker_quarterly. Only drop to staging.* or raw.* if you need to validate a number that looks off, and tell me when you do.

Show me the SQL first. Only execute once I approve the plan.

What to look for in the agent's plan

Before approving the SQL, check that the agent:

  • Queries reports.ticker_quarterly as the primary source - not raw.* or staging.*
  • Uses the pre-computed ttm_fcf_margin column instead of recomputing FCF / revenue from raw
  • Uses regime_at_qend for the macro regime split (already calculated in staging)
  • Computes the peer average across the 5 tickers within the same query (window function or subquery)
  • States explicitly if it does need to drop into staging.* or raw.* - and why

If the agent reaches into raw.fmp_fundamentals to recompute things, push back: "reports.ticker_quarterly already has TTM FCF margin and the regime - use that." This is the discipline that makes the layered pipeline pay off.

Follow-up questions to try

Once the first question runs clean, keep pulling threads - tap any prompt below to copy it:

AI Prompt

For the qualifying tickers, pull operating margin and R&D spend from staging.fundamentals for the same quarters. Does the trend change your interpretation? (Acceptable to drop to staging.* here - these columns aren't surfaced in reports.ticker_quarterly.)

AI Prompt

Rerun the analysis using reports.ticker_quarterly but tighten the threshold: require fcf_margin_qoq_change > 0.02 (200bps) in each of the last 4 quarters. Does the set of qualifiers narrow?

AI Prompt

Using reports.ticker_quarterly, overlay fed_funds_rate_at_qend for the qualifying tickers. Were they concentrated in periods with rising rates?

AI Prompt

For the qualifying tickers, query staging.macro to see what unemployment (unemployment_rate) was doing during their underperformance windows. Any relationship worth investigating?

Each of these takes a real analyst 30–60 minutes to produce manually. Your agent does them in the time it takes to write the prompt.

Save the good queries

When the agent writes a SQL query you like, ask it to save it as a Bruin SQL asset in the reports layer:

AI Prompt

Take that query and save it as stock-analyst-101/assets/reports/fcf_vs_price_underperformers.sql materialized as reports.fcf_vs_price_underperformers. Use materialization view (cheap, always re-derives from reports.ticker_quarterly). Add a top-level description, column descriptions, and a tier:reports tag.

Now it's part of your pipeline. Next run of bruin run, the view is rebuilt from fresh data. You're not just asking questions - you're growing your reports layer with every analysis you save.

Habits for getting consistently good answers

  • Ask for the plan before the result. Prompts that include "show me the SQL first and explain the logic" catch mistakes cheap.
  • Be specific about time windows. "Recent" is ambiguous; "last 4 fiscal quarters as of today" is not.
  • Name the comparison. "Underperformed" vs. what benchmark? State it. Otherwise the agent picks one and you don't know which.
  • When in doubt, ask for counts first. "How many rows satisfy X?" before "here are all the rows satisfying X" - saves you from chasing ghost results.
  • Iterate on AGENTS.md. Every time the agent gets something wrong for a domain reason, add a line to AGENTS.md so it doesn't repeat the mistake.

What just happened

You built a complete stock-analyst workstation - ingest, database, context, agent - on a single laptop, in one sitting, without writing SQL or Python yourself. The stack is yours to keep. Swap in different tickers, add new FRED series, plug in a second fundamentals source. Every addition follows the same pattern: prompt the agent, it runs Bruin, the data lands.

Where to go next