AEM 6991

MPS Capstone Project I

Prof. Ariel Ortiz-Bobea

1 · Introductions

Wednesday, August 26, 2026

Introductions

About me

• Ariel Ortiz-Bobea. Dyson School and the Brooks School of Public Policy. At Cornell since 2014.
• Before Cornell: Resources for the Future; PhD, University of Maryland; Ministry of the Environment of the Dominican Republic.
• Research: how people cope with environmental change — especially what climate change does to the economy, and to agriculture in particular. More at arielortizbobea.github.io.

Office 450B Warren · ao332@cornell.edu

Three minutes. Then one concrete result of your own and what its identification problem was — that is the thing they will be doing all term, and it is worth more than the CV.

Now you

Four things, briefly:

1. Your name, and where you are from
2. What you studied or did before this — training, or work
3. Your concentration
4. What you want out of this program, and after it

Go round the room. This is the block — thirty minutes budgeted, and it will use them. It is how the room learns who it is, and they hear each other’s interests before teams get formed next week.

Session A has nine minutes of slack behind this. Let it run.

Keep it moving. If someone runs long on item 4, that is the one to let breathe; items 1–3 are the ones to cut.

Write down anything that sounds like a real question rather than a topic. You form teams from the survey responses next week, and this is your first signal.

The course

Two courses, one project

AEM 6991 (fall). You produce a plan — a question, a design, and evidence that the data exists.

AEM 6992 (spring). You execute that plan.

The deliverable in December is a proposal somebody could hand to a team in January.

Everything you commit to in the fall you live with in the spring. A question that is vague in October is a disaster in March.

What this course is for

By December you should be able to:

• Find a question worth a semester, and say why it matters
• Design a study that could answer it — and name the assumption carrying it
• Prove the data exists, can be obtained, and measures what you need
• Say what you can and cannot claim, and defend that in a room
• Work in a team that delivers on time, with the work split fairly

These are the registrar’s four outcomes for 6991 plus the two this section adds — evidence sourcing, and design vocabulary. Do not read them off. Say the fifth one is graded as hard as the rest, because the course description calls teamwork a major component and every other section treats it as goodwill.

What this is not

You are not writing a paper for a journal.

• No gap in the literature to fill
• No contribution to claim
• Nobody is going to referee this for an academic journal

The single most common misunderstanding. Students hear “research design” and reach for the genre they have seen in seminars. Kill it here.

What it is

You are building the analysis a competent team would need before making a decision that costs real money.

The standard for “competent” is the same standard a referee would apply.

Rigor and genre are different axes. The rigor bar does not move. The genre does.

If the design would not survive a referee, it does not survive here. But the audience is a decision-maker, not a discipline.

This also rescues the literature review from feeling academic: you read to find out whether your question is already answered, not to find a gap.

Placement and skills

MPS AEM, class of 2025:

Job function	Share		Industry	Share
Finance	40%		Financial services	38%
Marketing / sales	21%		Technology	22%
General management	15%		Manufacturing	11%
Consulting	13%		Consumer packaged goods	5%
Data / analytics	6%		Other	24%

Goldman · JPMorgan · Morgan Stanley · Citi · HSBC · Tencent · Alibaba · P&G · PwC · KPMG

Source: Cornell SC Johnson, MPS AEM career outcomes. 82% received offers within six months.

Dyson splits what you need into “sharp skills” — statistical thinking and data analytics — and “smart skills”: communication, teamwork, project management, leadership.

GMAC asked 1,100 recruiters what they value in business graduates in 2026. Communication first. Problem-solving second.

Essentially none of these are research jobs. That is the argument for the framing on the previous slide — not my taste, the placement data.

The callout is the argument for how this course is weighted. Half the grade is a team artifact and participation is 20%, because the two things employers rank first are the two things a solo written exam cannot test.

Data analysis and interpretation came fourth in the same GMAC survey, up from tenth in one year. Worth saying if you have the time.

The course has a website home

arielortizbobea.github.io/aem6991

• Every meeting has a page — what session A covers, what session B makes you do, and what is due
• The syllabus is there in full, and it is the contract
• Handouts, resources and worked examples, all permanent

Canvas carries grades and submissions. Everything else is on the site.

Still being built. Meetings 2 onward are outlines for now, and they fill in before each meeting. The syllabus and the schedule are settled.

Put it on the projector for thirty seconds. Open meeting 1’s page, show that the slides and the manual are the same material, and show the schedule. Then close it.

Be straight about the state of it — they will click through and find stubs. What is settled is the syllabus, the dates and the arc. What is not is the teaching material for meetings 2 to 14, which lands before each meeting.

Say the one thing that matters: nothing is handed out on paper that is not also there, so a missed meeting is recoverable.

The fourteen meetings

Getting started	1 Introductions
Foundations	2 Causality · 3 Questions · 4 Ethics · 5 Tools and AI
Methods	6 Models · 7 Model failures · 8 Designs I · 9 Designs II
Communication	10 Writing · 11 Figures · 12 Speaking
Delivery	13 Midterm · 14 Pitch

Thirty seconds, then navigate to the site and show one meeting page live — what session A covers, what session B makes you do, what is due.

Do not walk all fourteen. The point is the shape: the question and the foundations first, then methods, then how to communicate it, then delivery. Say why that order — you cannot pick a design before you have a question, and you cannot get data before you know what the design needs.

Assignments you submit

Sep 1	Research interest survey	individual
Sep 2	Your team’s ground rules	team
Sep 30	Project pitch — 1 page, presented · rate your teammates	team
Nov 11	Proposal outline · rate your teammates	team
Nov 18	Midterm, in class	individual
Dec 2	Final presentation	team
Dec 11	Research proposal, ~5 pages	team

Three team milestones and one exam. That is the spine — say it that way rather than reading the table.

The pitch is one page and five minutes. The outline is where the whole argument gets fixed. The proposal is that outline written out.

Nov 18 is fixed and will not move. Dec 11 is to be confirmed.

How you are graded

Half individual, half team.

Quizzes	15%		Project pitch	5%
Midterm	15%		Proposal outline	15%
Participation	20%		Research proposal	30%

Do not read the table. Say the two things that matter: half of it is yours alone, and the team half is one paper reached in three steps.

Participation starts at full marks and you lose them — a missing statement of interest, an unfiled check-in, a late peer evaluation, arriving late, an absence without notice. The detail is on the site.

Room is Riley-Robb B15 all semester. TA to be announced.

Quizzes

Short, in class, and without warning. About six across the semester.

• Three questions, three minutes, on your own device
• On the lecture you are sitting in, not on a reading
• Tied to your NetID — it is also the attendance record
• The lowest is dropped, so an honest miss costs nothing

Say why they are unannounced: a quiz you can revise for is a reading check, and a quiz you cannot is a reason to be here and paying attention. That is the thing worth rewarding.

No quiz today. Say the mechanic and move on.

AI: permitted, disclosed, verified

Writing the code is no longer the scarce skill. Neither is drafting the prose. Knowing what to ask for is. So is noticing when the answer is wrong.

From meeting 12 you work with real data through an assistant. What is graded is the questions you ask and the checks you run, never code you wrote.

McKinsey now runs analyst interviews where you work a problem live on their AI platform, and they score how you prompt it, how you catch it being wrong, and how you turn that into a recommendation.

And the other half. In a trial with a thousand students, those who used a plain chatbot to practise scored 17% worse once it was taken away — worse than students who never had it. Those given a version that made them work did not.

The midterm is where that shows up.

Say plainly that policies differ across sections and that this one is deliberately permissive. The McKinsey fact IS the reason, so do not rush it. Employers put “skills using AI tools” FIRST on what they expect to value in five years, while ranking it 14th today.

The PNAS study is Bastani et al., ~1,000 high-school maths students. Be accurate about it: the harm came from using it as a crutch during practice, not from using it at all. The guardrailed version showed no harm. That distinction IS the policy — permitted, disclosed, verified.

Do NOT bring the grim labour-market version into week 1 — entry-level postings down 35%, a third of employers already substituting AI. True, and it will flatten the room. It belongs in meeting 14.

15 minute break

Session B · Your interests

What this hour is for

Teams do not exist yet. I form them from what you tell each other this afternoon, and from the statement you write this week.

Round one	In a circle — you say it, somebody says it back	20 min
Round two	New group — they find what stands in the way	18 min
Report back	Two or three per group, to the room	12 min
	Where to look for data, and what is due	8 min

Put this up and leave it up while you explain. They pace themselves better when they can see that round one is not the whole hour.

Groups of five, by proximity. Say out loud that these are NOT the teams, or some of them will negotiate instead of working.

Sit them in circles. It matters — a circle makes listening visible and a row does not.

If the break ran late, take it out of round two, not round one. Round one is the one that produces the sentence everything else uses.

Round one · say it, then hear it back

In your circle, one at a time:

1. You say what you would like to explore, and why it interests you
2. The person on your left says it back, in one sentence
3. You confirm it, or fix it in one sentence

Then the next person. Everyone speaks once and summarizes once.

Twenty minutes. Two each, roughly.

The summary is the exercise, not the speech. Say that before you start: your job while somebody talks is not to plan your own turn.

The correction in step 3 is where the work happens. Most summaries come back subtly wrong, and that is the moment a vague interest gets sharper. Push on it if a group is being polite: “is that actually what you meant?”

Circulate with a pen. What you hear here is better than what you get in the statements next week.

Round two · new groups, and what stands in the way

Re-form into different groups of five.

1. Describe your idea again — to people who have not heard it
2. The group names the obstacles. Not you
3. You write them down

Eighteen minutes.

The group names the obstacles, not the owner — the owner is invested and has usually been avoiding the obvious one. Somebody hearing it cold names it in ten seconds.

Expect: no data, nobody who would act on it, too big, already answered. Those are four of the six anti-patterns and half the feasibility filter, arrived at by the room rather than taught. Do not name them as such today. Meeting 3 does that, and it lands harder because they got there first.

Second time describing it to a fresh audience is also the real test of whether it is communicable. The first group already knew what they meant.

Report back

Each group picks two or three to tell the room.

For each: what the person wants to explore, and the biggest obstacle the group found.

You never report your own. Somebody else says yours.

Twelve minutes — six groups at about two minutes.

Never your own is the rule that makes the listening real, and it removes the defensiveness people bring to describing their own idea.

Write every one on the board as it comes. The room seeing them all at once is the point: overlaps become visible, and that is how people find each other before the research interest survey is due.

Collect every sheet. Together with the survey responses, this is what you form teams from.

What Cornell already pays for

Before you assume data has to be bought or scraped.

Companies & finance	WRDS · Capital IQ · Refinitiv · PitchBook · EDGAR
Markets & industries	Mintel · Passport · Statista · IBISWorld
Demographics	SimplyAnalytics · census.gov · BLS
Terminals	Bloomberg, three at Mann Library

The middle row is reports, not microdata. Good for sizing a market, useless for estimating on.

Two minutes. Do not read the list — say the point: this is tens of thousands of dollars of subscriptions you already have, and teams routinely propose buying data the College already owns.

The full list with links is on the Where to get data page. Point at it.

Bloomberg is worth an hour even if their project never uses one — it is the thing finance interviewers assume you have touched.

Dewey

app.deweydata.io — a commercial data marketplace the Johnson College subscribes to. About 84 datasets, free to you.

Property	Real estate transactions, listings, rents
Foot traffic	Weekly patterns, store visits, points of interest
Spending	Card panels, brand tracker, SKU-level transactions
Labour	Job postings with salary, layoff notices, wages
Firms & markets	Company insights, equities, bonds, ESG
Place & risk	Climate risk, weather, CDC health, census tables

Register tonight with your Cornell email. Approval is a human step and takes days.

Three minutes, and browse it live from YOUR account — they will not have access yet.

Say the two things they will ask. Yes it is free, the College pays. And no, you cannot pick a dataset and reverse-engineer a question onto it — though working backwards from what exists is legitimate if the question still has to survive on its own.

Each dataset page carries its schema, coverage and dates. Making them click that is the lesson: naming a dataset is not the same as checking it holds what you need.

Your assignment

Research interest survey, on Canvas, due Tuesday September 1 at 11:59 pm. Under an hour. Most answers are two or three sentences.

Checkboxes	Topics, methods you have run yourself, languages, and whether you would rather develop your own idea
Your question	What you want to find out, and why it matters
Where it came from	A course, a job, a paper, a place you know well
What you bring	A job, an industry, a language, access to somebody
Where this fits	What you want to be doing after this program

Do not research whether your data exist. An idea with no data plan is fine.

Changed after this meeting was taught. In class on August 26 this was a one-page memo in five sections. It is now a Canvas survey. Two days is not long enough to write about feasibility, and structured answers are what team formation actually runs on. Announce the change rather than letting them find it.

Last thing they hear, so land it. Say the deadline twice.

Two questions are how I match teams. What you bring is capability, and without it I can only match on topic, because five people with the same background is a worse team. Where this fits their plans is motivation, and it is the better predictor of who still cares in November.

Point back at the board: this is not a blank page.

How the teams get made

I read all of them, then form the teams.

I am matching on two things: what you want to work on, and what you bring. A team of five people with the same background is a worse team.

Most of these ideas will not survive team formation. That is expected, and it is not a verdict on yours.

“I don’t know yet” is useful information, not a failure. It means the idea is still a topic rather than a question, which is exactly what meeting 3 is for.

Teams are announced before meeting 2. Meeting 2 is causal language, and your team writes its own ground rules that afternoon.

Say plainly that they do not pick their teams, and say why: matching on interest and on complementary backgrounds beats self-selection, which sorts by who already knows whom.

Tell them the honest thing about “I don’t know yet” — it costs nothing, and it is better information for you than a fabricated certainty.