Yes. This is a good hackathon target, but I would not fork StocksAI into another permanent orchestration stack.

The useful pattern in StocksAI is straightforward: it has specialized agents for trend discovery, financial research, selection, and management; it produces structured JSON outputs; and it can trigger notifications. What we should steal is the operating pattern, then express it in our Factory model instead of CrewAI. The repo currently defines four agents, outputs tendencias.json, reporte.json, and decision.json, and runs a hierarchical research-to-selection workflow. StocksAI repository

I would brainstorm this as a Finance / Markets Pack family first, with a possible separate City only if we later need hard isolation, distinct credentials, separate risk policy, or very different runtime economics.

Core architecture¶

MARKET / FINANCIAL DATA
        ↓
DETECT / RESEARCH
        ↓
NORMALIZE EVIDENCE
        ↓
ANALYZE
        ↓
FORM HYPOTHESIS
        ↓
RISK / POLICY CHECK
        ↓
HUMAN REVIEW
        ↓
WATCH / SIMULATE / PAPER TRADE
        ↓
VERIFY OUTCOME
        ↓
LEARN

The critical difference from StocksAI is:

NO:
agent researches stock
→ agent says BUY
→ notification

YES:
signal detected
→ evidence gathered
→ thesis formed
→ counter-thesis required
→ risk constraints checked
→ human sees evidence
→ optional paper position
→ outcome tracked

That makes it a much better fit for Bluefly's governance model.

1. Daily Market Intelligence Pack¶

This is the easiest hackathon.

Every morning:

macro events
earnings
SEC filings
analyst changes
sector movement
large price moves
news

are gathered and turned into:

WHAT_CHANGED
WHY_IT_MATTERS
WHICH_WATCHLIST_NAMES_ARE_AFFECTED
WHAT_NEEDS_ATTENTION

Factory concepts:

Formula:
market-daily-brief

Event:
market.material_change

Beads:
bounded research questions

ContextControl:
morning dashboard

The output should not be "buy these three stocks."

It should be:

Five things materially changed overnight and these are the evidence-backed implications for the portfolio/watchlist.

2. Investment Thesis Factory¶

Pick a ticker and create a repeatable research process.

Example:

NVDA
  ↓
company fundamentals
  ↓
recent filings
  ↓
earnings / guidance
  ↓
competitive position
  ↓
valuation
  ↓
risks
  ↓
bull case
  ↓
bear case
  ↓
unknowns
  ↓
thesis receipt

The output becomes a durable thesis record:

THESIS=
EVIDENCE=
COUNTER_ARGUMENT=
KEY_ASSUMPTIONS=
INVALIDATION_CONDITIONS=
WATCH_SIGNALS=
CONFIDENCE=

Then the Factory watches whether the assumptions remain true.

That is much more interesting than a one-time stock report.

3. Thesis Drift Monitor¶

This one fits the Factory extremely well.

After a thesis is approved:

"Company X grows revenue 25%+
while maintaining margin Y"

Orders monitor:

earnings
10-Q
10-K
guidance
management changes
major competitor news
regulatory events

Event:

investment.thesis_changed

Then:

existing thesis
→ compare new evidence
→ classify:
   STRENGTHENED
   UNCHANGED
   WEAKENED
   INVALIDATED

This becomes continuous investment research instead of periodic rediscovery.

4. SEC Filing Intelligence Pack¶

Very concrete and useful.

For every watched company:

10-K
10-Q
8-K
Form 4
13D / 13G

ingest, compare, and answer:

WHAT_CHANGED FROM LAST FILING?
WHAT LANGUAGE CHANGED?
WHAT NEW RISKS APPEARED?
WHAT NUMBERS MOVED MATERIALLY?
WHAT WAS REMOVED?

Most of the work can be deterministic first.

Then use models only for interpretation.

This is exactly the same principle we use in Factory documentation maintenance:

detect cheaply, reason only where ambiguity exists.

5. Earnings Factory¶

Before earnings:

previous guidance
consensus expectations
historical surprises
key metrics
known risks

After earnings:

actual vs expected
guidance change
margin change
cash flow
segment performance
management commentary

Then produce:

PRE_EARNINGS_THESIS
POST_EARNINGS_EVIDENCE
DELTA
THESIS_STATE

That would make a strong visual demo.

6. Watchlist Operations¶

Instead of manually checking 50 stocks:

WATCHLIST
   ↓
events continuously detected
   ↓
only material exceptions surface

Possible signals:

price gap > threshold
volume anomaly
earnings date approaching
filing posted
insider purchase/sale
guidance revision
credit downgrade
sector shock
significant news

No signal:

NO AGENT
NO BEAD
NO MODEL

Material signal:

Event
→ bounded Bead
→ research Formula
→ evidence
→ dashboard

That is pure Factory logic.

7. Portfolio Risk Monitor¶

A Pack that looks across the portfolio rather than individual names.

Examples:

sector concentration
single-name exposure
factor concentration
market-cap concentration
correlated holdings
earnings clustering
macro sensitivity

ContextControl-like UI:

PORTFOLIO RISK

Technology          47%
Top 5 positions     61%
Earnings next 7d    4 holdings
High correlation    3 clusters
Material alerts     2

The system surfaces exceptions rather than making trades.

8. Contrarian / Red-Team Research Pack¶

This may be one of the coolest demonstrations of multi-agent work.

Every bullish thesis automatically requires:

BULL RESEARCH
vs
BEAR RESEARCH

Then Witness-like review asks:

Did both sides use credible evidence?
Did either side omit contrary information?
Which assumptions are unsupported?
What would falsify the thesis?

This is much better than four agents agreeing with one another.

The Factory should structurally force disagreement.

9. Paper Trading Factory¶

This is where I would stop before real capital.

Approved hypothesis:

→ simulated trade
→ entry timestamp
→ entry price
→ thesis
→ target
→ invalidation
→ risk assumptions

Then track:

1 day
1 week
1 month
3 months

and compare outcome against prediction.

Now your Factory can measure:

which signals work
which research patterns fail
which models overstate confidence
which analysts/agents add value

That is actual learning.

10. Strategy Backtesting Pack¶

A Formula defines a hypothesis:

"Buy companies after earnings
when revenue surprise > X,
guidance raised,
and valuation < Y."

Then deterministic code tests the historical data.

Agent work should be limited to:

forming hypothesis
interpreting results
identifying confounders

Python/data tooling does the math.

That is important.

Do not have an LLM "backtest" from memory.

11. Dividend / Income Factory¶

Could be more approachable for a hackathon audience.

Track:

dividend yield
payout ratio
free cash flow
dividend growth
debt
coverage
ex-dividend date
earnings stability

Output:

INCOME QUALITY
DIVIDEND SAFETY
CHANGE SINCE LAST REVIEW

Again, research and monitoring, not automatic investment direction.

12. Insider Activity Pack¶

Monitor Form 4 filings.

Classify:

routine sale
option exercise
meaningful open-market purchase
cluster buying
executive disposal

Then enrich with:

position size
historical behavior
company events
valuation

Event:

market.insider_activity_material

13. "Why Did This Move?" Factory¶

This would be a great visual product demo.

Enter:

TSLA -8.2% today

Factory gathers:

company news
sector moves
market move
options activity if available
analyst action
filings
macro events

Then returns:

LIKELY CONTRIBUTORS

1. ...
2. ...
3. ...

EVIDENCE STRENGTH=
UNEXPLAINED_COMPONENT=

Not fabricated certainty.

This could become a compelling AG-UI surface.

14. Portfolio Decision Journal¶

Every investment decision becomes a durable receipt:

DECISION=
DATE=
THESIS=
EVIDENCE=
RISKS=
ALTERNATIVES=
EXPECTED_OUTCOME=
INVALIDATION=

Months later:

OUTCOME=
WHAT_ACTUALLY_HAPPENED=
WHICH_ASSUMPTIONS_WERE_WRONG=
WHAT_SHOULD_BE_REUSED=

This is probably the most Bluefly-like concept of all.

It converts investing from:

"I thought this looked good."

into:

"Here is what we believed, why, and whether the evidence eventually supported it."

15. Investment Capability Flywheel¶

You could apply the same Factory law:

RESEARCH RUN
→ OUTCOME
→ LESSON
→ CAPABILITY CANDIDATE
→ REUSE
→ VALIDATE AGAIN

Example:

CAPABILITY:
Detect earnings-guidance divergence

RUN 1:
useful signal

RUN 2:
useful signal

RUN 25:
measured precision / false positives

Then you actually know whether a research capability is useful.

Side Pack or Separate City?¶

For a hackathon, I would absolutely start with:

ONE FACTORY
+
FINANCE PACK

Something like:

Pack:
markets-research

with:

market-daily-brief
company-thesis
filing-delta
earnings-analysis
thesis-monitor
portfolio-risk
paper-trade
outcome-review

I would only make it a separate City if we eventually need:

separate financial-data credentials
hard security isolation
different model providers
regulated workflows
broker connectivity
capital execution
independent work/evidence retention

If money can actually be moved, I would lean strongly toward a separate governed City/security boundary.

For read-only research and paper trading, a Pack is enough.

Best hackathon concept¶

I think the strongest one-day hackathon is:

Investment Thesis Factory¶

Pick 5 companies.

Build this flow:

MARKET SIGNAL
    ↓
COMPANY RESEARCH
    ↓
SEC FILING EVIDENCE
    ↓
BULL CASE
    ↓
BEAR CASE
    ↓
RISK CHECK
    ↓
THESIS
    ↓
HUMAN APPROVAL
    ↓
PAPER POSITION
    ↓
CONTINUOUS MONITORING

UI could show:

APPLE
──────────────────────────────

THESIS STATUS
● ACTIVE

Conviction
72%

Bull Case
...

Bear Case
...

Key assumptions
3

Material risks
4

Evidence
27 sources

Last changed
2 hours ago

NEXT EVENT
Earnings in 12 days

[ VIEW EVIDENCE ]
[ CHALLENGE THESIS ]
[ CREATE PAPER POSITION ]

Then another screen:

PORTFOLIO THESIS BOARD

AAPL   ACTIVE       72%
NVDA   WEAKENING    58%
MSFT   ACTIVE       81%
TSLA   CHALLENGED   39%
AMZN   REVIEW       66%

That would showcase nearly everything you care about:

Agents
Formulas
Events
Beads
Evidence
Context
Human approval
Visual UI
Continuous monitoring
Reusable capability
Economic measurement

without touching real money.

What I would borrow from StocksAI¶

From the repo, keep the conceptual sequence:

TREND DISCOVERY
→ DEEP RESEARCH
→ SELECTION / PRIORITIZATION
→ NOTIFICATION

But convert it to:

SIGNAL
→ EVIDENCE
→ RESEARCH
→ BULL / BEAR
→ POLICY / RISK
→ HUMAN DECISION
→ PAPER OUTCOME
→ LEARNING

That is a much more mature Factory experiment than cloning its CrewAI hierarchy.

And I would deliberately make the hackathon rule:

No real trades. No brokerage write access. No autonomous buy/sell. Research, evidence, simulation, and monitoring only.

That keeps the experiment technically interesting while letting you focus on what the Factory is actually good at: governed, repeatable decision support.