Built an A.I. Investment Committee. One Program Gathers the Data. Another Makes Six Experts Argue Over It.

Important Notes:

This article is a technical and educational demonstration of how a piece of software works. It is not investment advice, not a recommendation, and not a view on any security. Any company or data shown is illustrative only. See the full disclaimer at the end.

A number of my friends in the investment community keep asking me a version of the same question. And it usually starts with a complaint.

The hard part of investing today is not forming opinions. It is the data mining. The information that actually moves a stock no longer sits politely in a broker's report — it is scattered across the internet, it arrives fast, and gathering it quickly and efficiently is a genuine slog.

Talk to anyone who follows US equities and you hear the same list of things they now feel they have to keep up with:

• X (formerly Twitter) — because it is simply faster. News and reaction break here long before they reach a research note.

• Polymarket — because people increasingly read the betting odds on global and macro events, and those events move the equity market.

• YouTube — because it is full of KOLs who, at their best, explain a stock from an angle you had not considered.

• SEC filings — Form 4 — because they want the official record of who is actually trading the stock: every time a company's own insiders buy or sell their shares, they are legally required to file it.

• The professional reports — the rigorous fundamental work from the likes of CFRA (and Schwab), which serious investors still pay for and read alongside everything above.

Some of it is free and scattered, some of it is official, some of it is paid — and pulling it all together, every day, by hand, is the pain. So the question my friends keep asking comes in two parts. Can A.I. just go and gather all of this for me — the latest data and chatter — in one shot? And then can A.I. actually help me make sense of it?

https://www.youtube.com/watch?v=CNBgtgrv7C4

My answer is yes. And it comes as two programs.

Program One does the gathering. One click, one stock name, and it brings back the news and the data from all those channels at once. The heavy lifting is done by an innovative open-source project on GitHub called last30days — a genuinely clever way to sweep the social web fast. (https://github.com/mvanhorn/last30days-skill). But what i have modified is to add more data sources.

Program Two does the thinking. Here, state-of-the-art A.I. takes over to do more fact-finding, more research, and the actual analysis.

We start with ingesting the some professional analyst reports. Users can choose anything they want. A.I. then turns this document into disucssion points and fact finding basis.

https://www.youtube.com/watch?v=SoG7rDS4Cx8

But this is where most people go wrong. They take all that data, paste it into a single chatbot, ask a question, and hope. Sometimes it works. Usually it disappoints. So in Program Two I do something different: I structure the A.I. as an investment committee — several specialist actors, each an expert in one thing, who then argue and debate until the best idea survives.

Why build the thinking that way? Three reasons.

Because it plays to what an A.I. uniquely is. A single model has, in effect, read how every kind of expert thinks — the bull, the bear, the quant, the risk manager — so it can become any of them on demand.

Because of focus. Ask one A.I. to hold everything at once and answer in a single breath, and it spreads itself thin. Give each actor one job and one lens, and each stays sharp. Modern models are getting better at this — but the principle still holds.

And because it mimics us. We do not make serious decisions by asking one oracle. We put specialists in a room and let them disagree. I always want A.I. to do what good humans already do — only faster, and without getting tired.

So: one program to gather, one to deliberate — and keeping them apart is the whole trick.

https://www.youtube.com/watch?v=XYvgu4QSIbI

Why I Kept Them Separate

I could have built one giant program that does everything. I deliberately did not — and the separation is the part I am most pleased with.

Picture a newsroom. The reporters go out and bring back facts. The editors sit around a table and decide what it all means. You do not want your reporters writing the editorial, and you certainly do not want your editors inventing facts. Keep the two jobs apart, and each one gets better.

That is exactly the split here. The gathering program has no opinions; it just fetches, cleanly. The thinking program does not care where the data came from; it just reasons. And because they are separate, the whole thing stays flexible: I can add a new data source and the committee never changes — it simply has more to chew on; I can sharpen the committee without touching a single line of the data plumbing.

One program reads the world. The other argues about it. Together they behave like a research desk that never sleeps.

Program One — Gather Everything

No person can watch all of this at once. A machine can. So the first program's only job is to go and read everything, the same way, every time — including the single line in a filing a tired analyst skims past at 6pm.

Here is exactly what it pulls, and what each piece gives you.

The professional baseline. Two reports I already subscribe to — a Schwab equity rating and a CFRA analyst report. (You can use any reports you prefer from any of your favorite professional source). Here we just pick two. (Why two? Well, that is becuase you can get two different flavors which tend to focus on different things and have different styles.) What they have usually are : the grown-up scorecard: the letter grade and percentile, the fair-value estimate, the twelve-month target, earnings forecasts, the fundamentals and the named risks.

The live price. A real-time quote — current price, 52-week high and low, market cap — so nothing is reasoned on stale numbers.

What the crowd is saying. Here I did not reinvent the wheel. The real-time listening layer is powered by last30days, an open-source project on GitHub (https://github.com/mvanhorn/last30days-skill) that has quietly become one of the most popular downloads of its kind. It sweeps the public internet for what people are saying right now:

• Reddit — retail discussion, due-diligence threads, and the screens that flag which stocks are oversold or overbought.

• X — the fastest takes on the planet: traders, influencers, technical calls, the viral bull-and-bear threads.

• Hacker News — the technology and operational angle, where the product story surfaces.

• Polymarket — real money betting on outcomes, a surprisingly honest read on probability.

• YouTube — longer analysis and earnings breakdowns, read straight from the transcripts.

• The wider web — the finance press and research sites, for grounding.

  
  

What I added on top. last30days gives you the chatter. But an investor needs more than chatter — so I extended it with three more sources of my own:

• What the insiders are doing — straight from the SEC: Form 4 filings, the legally-required record every time a company's own executive buys or sells their shares. Who, how much, at what price, and which direction. (https://www.secform4.com/)

• 
  

• What the ticker crowd thinks — StockTwits: a running bull-versus-bear tally on the specific stock, and the posts people actually engaged with. (https://stocktwits.com/)

• 
  

• What the news and the analysts are doing — a timestamped news feed, plus the movement of analyst opinion (how many rate it buy, hold or sell, and whether that shifted this month) and the next earnings date.

• 
  

One command, one stock name, and all of that lands on the table together.

And notice this: almost none of it costs anything. The filings are a government website. The chatter, the sentiment, the news, the analyst flow — all free. The only things I had to bring were the two broker reports I already pay for.

Program Two — The Committee Deliberates

Now the second program takes over.

Picture the committee around the table: the relentless optimist, the gloomy sceptic, the quant who only trusts the charts, the industry veteran who knows the business cold, the risk manager who worries about everything, and a sixth voice whose whole job is to listen to the crowd.

In a real firm, that is five senior people, five calendars and a room — where the loudest voice tends to win and the junior who disagrees stays quiet. Here, every voice is equal, always prepared, and the sceptic cannot be talked over.

And here is the part that still feels like a magic trick: those six experts are not six programs. They are one model, asked to play six roles — because, as I said at the start, it has read them all and can become any of them on demand. During the debate it remembers, so everything each character said in round one is still there in round two. The bear answers the bull directly. The optimist must defend the exact point the sceptic just attacked.

You are not getting one A.I. opinion. You are getting an argument. And arguments are how weak ideas get found out.

Before they argue, they check. When the crowd insists "analysts see big upside," the committee does not take the tweet's word for it. It sends the claim back out to the live web, pulls what the analysts actually say with the sources attached, and tells you honestly what it could not confirm.

Facts first. Opinions second. A debating A.I. with no grounding is just a confident liar.

It also keeps two kinds of evidence apart — the soft stuff (chatter, sentiment) and the hard stuff (fundamentals, insider trades) — so the noise of the first never masquerades as the certainty of the second.

Watching It Work

Let me show you, with a real example.

Take McDonald's, on a day when the stock was sitting near its 52-week low. The crowd was loud and split. On X, one camp was waving the chart and shouting "falling knife — don't catch it."

Another was passing around the fact that a famous billionaire owned it and that analysts saw plenty of upside. On Reddit it kept appearing on oversold screens. Pure narrative — and pure noise, in both directions.

So the committee went looking for something harder. It reached into the insider filings the first program had gathered. And there it was: in the prior month, two senior executives had sold small amounts. Nobody had bought. Not a single share, at the lows.

That is not a scandal — executives sell for all sorts of ordinary reasons. But it is a fact the chatter simply could not give you, and it sharpens the whole argument. If this were truly the bargain the optimists claimed, you might expect at least one insider to step in at the bottom. None did.

The optimist now has to answer for that. The sceptic gets to lean on it. The committee writes down both, and you read the part that survived.

That is the magic of two programs working in sequence: the first one reads the world and lays it on the table; the second one checks it, argues over it, and decides what it is worth.

(To be clear, and I will say it again below: I ran McDonald's purely because I needed a real, recognisable name to test the plumbing. Nothing here is a view on the company.)

One Thing I Have to Say Plainly

I need to be completely clear about what this is — and what it is not.

This is a programming exercise. Full stop. It exists to demonstrate how a machine can gather data, check it, and argue over it. I am not expressing any view on McDonald's, or on any other stock.

When the program prints a rating, a price target, or a trade idea, that is the mechanical output of the process — the shape of an answer, produced to show the method works. It is not my opinion, it is not a recommendation, and it is not investment advice. I chose McDonald's only because it is a name everyone knows and a convenient way to test the wiring.

Markets are hard, and no clever pipeline changes that. Anything the system produces should be treated as a starting point for your own work — and your own professionals — never as a signal to act.

The Bigger Picture

For most of financial history, this kind of synthesis — read everything, weigh every side, watch the crowd and the insiders at the same time — required a research desk and a very expensive terminal. A team. A budget. An institution.

Almost everything the gathering program reaches for is free and public. The committee that argues over it is one A.I. wearing six hats. And because the two halves are separate, each can grow on its own: tomorrow I can teach the gatherer a new source, or give the committee a new member, and the other side never even notices.

That is the part I keep coming back to. The committee used to be a room you had to be invited into. Now it is a process anyone can run.

One program reads the world. The other argues about it. The data is mostly free. The argument is yours to read.

And the one thing it cannot do is decide for you. That part — still, thankfully — is human.

Final Output

And here is the final output. To be clear, we just want to make this as an illustration and therfore we have manually redacted the output.

McDonald's Corporation (MCD) — AI Investment Brief

Redacted demonstration copy

⚠️ Redacted for public sharing. This is a demonstration of an AI research method, not investment advice and not a view on any security. All ratings, price targets, entry and exit levels, position sizing, and trade structures have been removed and marked [REDACTED]. The structure, the data sources, and the committee's qualitative reasoning are shown so you can see how the program works — nothing here is a recommendation.

Date: June 09, 2026 · Model: claude-opus-4-5 · Rating: [REDACTED] Current Price: $277.78 (live, 2026-06-09) · 52-Week Range: $271.85 – $341.75

SECTION A — INVESTMENT BRIEF

1. Price & Target Summary

Metric	Value
Current Price	$277.78 (fetched 2026-06-09)
Bull Target (12-mo)	[REDACTED]
Base Target (12-mo)	[REDACTED]
Bear Floor	[REDACTED]
Upside / Downside	[REDACTED]
52-Week Range	$271.85 – $341.75
% Off 52-Week High	−18.7%
% Above 52-Week Low	+2.2%

Valuation narrative (figures redacted). MCD trades near its 52-week low, a meaningful de-rating from its $341.75 peak. At [Current P/E], it trades below its [Historical Average Forward Multiple] cited by CFRA. CFRA's qualitative rating diverges from its quantitative fair-value model, an interesting tension the committee picks up below. The Schwab model assigns a "C" (Marketperform) at the 33rd percentile — neutral quantitative expectations. Dividend yield is ~2.7% (vs. ~1.4% for the S&P 500), giving income support. Specific targets and fair-value figures redacted.

2. Key Growth Drivers

#	Driver	Conviction	Evidence	Horizon
1	Global comparable-sales recovery	HIGH	Q1 2026 global comps +3.8% (vs −1.0% prior year); U.S. comps inflected to +3.9% from −3.6% (Schwab/CFRA)	6–18 mo
2	"McDonald's > NEXT" strategy	MEDIUM	Announced 1 Jun 2026: AI integration (ArchIQ, ARCHY voice AI), robotics, hospitality, food quality	12–36 mo
3	Unit expansion to 50,000 by 2027	MEDIUM	~2,100 net additions in 2026 (~5% unit growth); systemwide sales +11% YoY in Q1	12–24 mo
4	Loyalty program scale	HIGH	MyMcDonald's Rewards >$9B quarterly systemwide sales across 70 markets in Q1	Ongoing
5	Category expansion (chicken, beverages)	MEDIUM	Chicken/coffee initiatives flagged as "incremental growth optionality" (CFRA)	12–24 mo
6	Franchised-model cash generation	HIGH	95%+ franchised; mid-40% operating margins; >80% FCF conversion; ~$7.2B 2025 FCF	Structural

3. Key Risks

Rank	Risk	Probability	Impact	Monitoring signal
1	Consumer spending / traffic decline	HIGH	HIGH	Monthly same-store sales; low-income traffic commentary
2	Margin compression from promotions	MEDIUM	HIGH	Operating-margin trajectory; McValue / $5 Meal Deal economics; beef costs
3	FX headwinds	HIGH	MEDIUM	Q1 op income +12% reported vs +6% constant-currency; USD vs EUR/GBP/AUD
4	Valuation multiple compression	MEDIUM	MEDIUM	Forward P/E vs 10-yr average; analyst ratings shifting (+2 holds recently, Finnhub)
5	Food safety / brand risk	LOW	HIGH	Any recurrence of a contamination event (ref. Oct 2024)
6	Geopolitical / regulatory exposure	MEDIUM	MEDIUM	Operations in 100+ countries; sanctions, labour, local tax
7	Insider selling (sentiment)	LOW	LOW	Two senior executives sold ~$862k in late May near 52-week lows; no insider buys (Form 4 data)

4. Valuation vs. History vs. Peers

Metric	MCD (current)	MCD 5Y avg	S&P 500	QSR peers
P/E (TTM)	23.1x	~24.5x	46.4x	20–25x
P/E (Fwd 2026E)	21.6x	24.0x	39.6x	18–22x
EV/EBITDA	~15.0x	15.5x	—	12–16x
P/Sales (TTM)	7.24x	7.8x	9.00x	3–5x
Dividend Yield	2.7%	2.3%	1.4%	1.5–3.0%

Reading (factual): on the multiples above, MCD trades modestly below its own historical P/E, EV/EBITDA and P/Sales, richer than QSR peers on sales-based measures and roughly in line on others, with a dividend yield above both its own history and the S&P 500. Any valuation verdict or price target is withheld.

5. Upcoming Catalysts

Catalyst	Expected	Impact	What to watch
Q2 2026 earnings	Week of Aug 4–6	HIGH	Comps trajectory (traffic vs check), margin commentary, loyalty update
"> NEXT" execution	Ongoing	MEDIUM	ArchIQ / ARCHY rollout, franchisee adoption, labour offsets
U.S. consumer data	Monthly	MEDIUM	Gas prices, confidence, fast-food traffic proxies
Dividend increase	Sep/Oct	MEDIUM	49-year streak; magnitude signals confidence
Unit-growth update	Q2/Q3	LOW–MED	Progress toward 50,000 stores

SECTION B — INVESTMENT COMMITTEE DEBATE

Six specialist actors. Three rounds. Their stated ratings, targets and trade levels are redacted; their reasoning is shown so you can see the argument.

Round 1 — Opening Arguments

🐂 Bull PM (long-only, 18-month horizon). Fundamentals are inflecting: Q1 2026 delivered +3.8% global comps with U.S. comps swinging ~750bps positive year-on-year; adjusted EPS grew ~6%. The stock trades relative to its historical multiple, and the 95%+ franchised model throws off >$7B annual FCF at 80%+ conversion, funding nearly 50 years of consecutive dividend growth. [Rating, target and total-return figures REDACTED.]

🐻 Bear Analyst (valuation sceptic). CFRA's own quantitative fair-value model flags the stock as overvalued even as their qualitative rating is positive — a notable internal contradiction. Schwab's model is neutral ("C", 33rd percentile) with "D" grades on Quality and Valuation weighted high. The comp growth is check-driven, not traffic-driven — price increases, not more customers — and management has cited reduced traffic among lower-income cohorts. The gap to its historical multiple may be warranted, not anomalous. [Fair-value and downside figures REDACTED.]

📊 Quant Trader (technicals, momentum). Price bounced off the weekly long-term moving average at multi-week lows; StockTwits sentiment is ~80% bullish with elevated options activity. But momentum is weak — material underperformance versus the S&P over the trailing six months — and a reversal needs follow-through above resistance. Cautiously constructive on a mean-reversion basis only. [Entry, target, stop and strike levels REDACTED.]

🔬 Industry Specialist (moat, supply chain). The moat is real and durable: 45,000+ restaurants across 100+ countries, ~$140B systemwide sales, real-estate control, and a capital-light franchised model. The "> NEXT" strategy signals serious AI investment (ArchIQ, ARCHY) that should help labour costs and throughput. Watch chicken-category execution against strong incumbents. [Rating/target REDACTED.]

🌍 Macro / Risk Manager (FX, rates, sizing). FX is a real headwind (Q1 op income +12% reported but only +6% constant-currency). The consumer is under pressure (gas prices, softer spending). Beta near 0.41 confirms a defensive profile — but defensiveness is not immunity. [Position-sizing guidance REDACTED.]

📡 Sentiment Analyst (crowd narrative, flow). The crowd narrative is "quality defensive on sale" — StockTwits ~80% bullish, traders celebrating the bounce. Red flag from the Form 4 data: two senior executives sold ~$862k in late May, near the exact levels retail is buying; no insider buying. The Finnhub analyst breakdown skews positive (≈8 strong-buy, 16 buy, 15 hold, 1 sell) but has shifted +2 into holds recently — valuation fatigue. No crowded short to squeeze; a long-biased, complacent setup that could be punished on a miss. [Rating/target REDACTED.]

Round 2 — Rebuttals (selected)

• Bull rebutting Bear: the quantitative fair-value model is backward-looking and misses the Q1 comp inflection; positive check growth reflects successful trade-up, and management guided to [X]% operating margins (Q1 delivered [Y]%). [Figures REDACTED.]

• Bear rebutting Bull: the comp swing is flattered by weak prior-year compares; CFRA itself revised comp estimates lower on softer macro data; paying a historical-average multiple for ~7% EPS growth in a higher-rate world is not a value opportunity. [Multiples/targets REDACTED.]

• Quant rebutting Sentiment: the insider selling is immaterial in dollar terms (<0.005% of market cap) and looks like routine diversification; what matters for price is positioning and flow. [Levels REDACTED.]

• Industry rebutting Macro: FX is understood and managed; geographic diversification is risk mitigation; MCD's value positioning is the antidote to a squeezed consumer.

• Macro rebutting Industry: "MCD gains share in a recession" is consensus — and consensus is often wrong at turning points; the "cheap burger" narrative has eroded as menu prices rose.

• Sentiment rebutting Quant: the signal is directional, not dollar-size — insiders sold into weakness, and ~80% bullish retail means the easy money is already positioned long, leaving asymmetric downside on a miss.

Round 3 — Final Positions

Each actor closed with a stated rating, price target, entry and stop. All of those specifics are redacted. Their reasoning in one line each:

• 🐂 Bull PM: comp inflection is real, franchise model is bulletproof, and the dividend record anchors the bull case. [Position/target/entry/stop REDACTED.]

• 🐻 Bear Analyst: the historical premium is unwarranted given decelerating growth and margin risk; the quant models confirm fair-to-overvalued. [Position/fair value REDACTED.]

• 📊 Quant Trader: a mean-reversion trade, not a conviction position; needs confirmation to add. [Levels REDACTED.]

• 🔬 Industry Specialist: moat intact, AI strategy differentiated, but execution risk warrants patience. [Position/target REDACTED.]

• 🌍 Macro / Risk Manager: defensive compounder, not a growth stock — size for uncertainty until Q2 traffic confirms. [Sizing/target REDACTED.]

• 📡 Sentiment Analyst: retail bullish but not crowded enough to be contrarian; insider selling and analyst hold-shifts signal valuation fatigue. [Position/target REDACTED.]

Consensus Table

Actor	Action	Target	Entry	Stop	Conviction
🐂 Bull PM	[REDACTED]	[REDACTED]	[REDACTED]	[REDACTED]	HIGH
🐻 Bear Analyst	[REDACTED]	[REDACTED]	—	—	MEDIUM
📊 Quant Trader	[REDACTED]	[REDACTED]	[REDACTED]	[REDACTED]	MEDIUM
🔬 Industry Specialist	[REDACTED]	[REDACTED]	[REDACTED]	[REDACTED]	MEDIUM
🌍 Macro / Risk Manager	[REDACTED]	[REDACTED]	[REDACTED]	[REDACTED]	MEDIUM
📡 Sentiment Analyst	[REDACTED]	[REDACTED]	—	—	LOW

Committee weighted target: [REDACTED]

Surviving Arguments

🔴 Bear argument that survived all three rounds: comp growth is check-driven, not traffic-driven — price increases are masking demand weakness among low-income cohorts; if traffic fails to inflect, margin compression follows as the company leans on value promotions.

🟢 Bull argument that survived all three rounds: MCD is a Dividend Aristocrat trading relative to its historical multiple with a ~2.7% yield, while Q1 U.S. comps inflected sharply; the franchised model generates strong, durable free cash flow. Defensiveness is a feature, not a bug, in an uncertain macro environment.

Position Sizing

[ENTIRE SUBSECTION REDACTED.] Position sizing is itself a recommendation, so both the sizing and the reasoning behind it are withheld. The committee's underlying inputs — valuation multiples versus history, the macro and FX headwinds, the Q2 earnings date, the dividend record — already appear as facts in Section A and in the attributed debate above; the appraisal and the sizing call do not.

SECTION C — ACTIONABLE TRADE IDEAS

[ENTIRE SECTION REDACTED.] The program produced specific trade ideas (entries, stops, an options structure, and sizing). These have been removed for public sharing. This document demonstrates a research method only and is not investment advice or a recommendation on any security.

Prepared as a programming/method demonstration. Not investment advice. Not a view on any security. Conduct your own due diligence and consult your own licensed professionals.

DISCLAIMER — PLEASE READ IN FULL

Nature and purpose. This article is published for general information, educational, and technical-demonstration purposes only. Its sole purpose is to illustrate how a software/AI system gathers data and structures analysis. It is not intended to analyse, evaluate, or express any opinion on the merits of any security, issuer, or investment.

Not advice; no recommendation. Nothing in this article constitutes or should be construed as financial, investment, legal, tax, accounting, or any other professional advice, or as a recommendation, opinion, offer, solicitation, or inducement to buy, sell, hold, or otherwise deal in any security, financial instrument, or product. No view on any security, issuer, sector, or market is expressed or implied.

Illustrative content. Any company name, ticker, figure, rating, price, target, or program output shown is illustrative only and may be redacted, simplified, hypothetical, fabricated, incomplete, or out of date. It is included to demonstrate software behaviour, not to inform any investment decision. Do not rely on any data, figure, or statement in this article for any purpose.

AI-generated content. Parts of the content and the system described are generated by artificial-intelligence software, which can and does produce errors, omissions, and fabricated information ("hallucinations"). The content has not been independently verified and is provided "as is" and "as available", without warranty of any kind, whether express or implied, including as to accuracy, completeness, reliability, or fitness for any particular purpose.

No regulated activity; not a licensed adviser. This article is not, and must not be construed as, the carrying on of any regulated activity — including "advising on securities" or any analogous activity — in Hong Kong or any other jurisdiction. The author is not a licensed or registered investment adviser and is not holding himself out as one. If you require advice, you must consult an appropriately licensed and independent professional.

No relationship; no reliance. Nothing in this article creates any advisory, fiduciary, client, or other relationship between the author and any reader. You should not act, or refrain from acting, in reliance on anything in this article.

Risk and forward-looking statements. Investing involves substantial risk, including the possible loss of principal. Past performance is not indicative of future results. Any forward-looking statements are inherently uncertain and may prove incorrect.

Jurisdiction. This article is not directed at, and is not intended for distribution to or use by, any person in any jurisdiction where such publication, distribution, or use would be contrary to applicable law or regulation. It is your responsibility to comply with the laws applicable to you.

Limitation of liability. To the maximum extent permitted by applicable law, the author disclaims all liability for any loss or damage of any kind arising from or in connection with any use of, or reliance on, this article or its contents.

Do your own research. Always conduct your own due diligence and seek advice from qualified, licensed, and independent financial, legal, and tax professionals before making any decision.