How AI Empowers Economists to Do Better Work
2025-11-23
AI is not here to do your thinking.
It’s here to amplify your economic reasoning.
By the end of this lecture, you will:
Most students talk to AI like it’s a person.
That’s why they get mediocre results.
AI doesn’t understand—it predicts based on patterns.
AI is a sophisticated probability engine:
Example: “The regression coefficient is ___”
AI predicts: “significant” > “statistically” > “positive” based on learned patterns
In econometrics, you specify: - Functional form - Variables - Assumptions - Identification strategy
AI needs: - Clear role/expertise level - Rich context about your problem - Explicit goals and constraints - Iterative refinement
Both require precision, but different types!
Vague prompts → Vague outputs
Precise prompts → Precise outputs
This is “Machine English”—learning to speak AI’s language
Every effective prompt has three components:
Who is the AI?
Define expertise level and role
What context exists?
Provide data, background, constraints
What’s the goal?
Specify desired output clearly
Help me with my regression analysisProblems: - No expertise specification - No context provided - Unclear what “help” means - Will get generic, unhelpful response
You are an econometrician specializing in panel data
methods and causal inference. You have published in
top journals and regularly review manuscripts.
I'm analyzing the effect of minimum wage increases on
employment using county-level panel data (2010-2020).
I have 3,200 counties, quarterly observations, and
treatment variation from state-level policy changes.
My initial fixed effects model shows unexpected
positive effects (β = 0.03, SE = 0.01).
Review my specification and identify three potential
econometric issues that could explain this result.
For each issue, suggest a specific robustness check
or alternative specification.Actor (Expertise):
You are an econometrician specializing in panel
data methods and causal inference...→ Steers AI toward technical, field-specific knowledge
Input (Context):
I'm analyzing... county-level panel data...
3,200 counties... unexpected positive effects...→ Provides the grounding AI needs to reason appropriately
Mission (Goal):
Identify three potential econometric issues...
suggest specific robustness checks...→ Clear, actionable output specification
Scenario: You need to choose between Random Forest and XGBoost for predicting housing prices.
5 minutes: Work individually or in pairs
Then: We’ll compare approaches
ACTOR: You are a machine learning researcher with
expertise in ensemble methods and predictive modeling
in economics. You understand both statistical theory
and computational tradeoffs.
INPUT: I'm building a house price prediction model
using 50,000 observations with 75 features (mix of
categorical and continuous). Features include:
location, size, age, school quality, crime rates,
economic indicators. I need to decide between Random
Forest and XGBoost.
MISSION: Compare Random Forest vs XGBoost for this
specific use case. Structure your response as:
(1) Key theoretical differences, (2) Performance
tradeoffs for my data characteristics, (3) A clear
recommendation with 2-3 deciding factors, (4) One
practical implementation tip for the chosen model.The world’s smartest AI sounds clueless without context.
Inside AI: Billions of numbers in a mathematical space
Context: The map that tells AI where to look and what matters
| Component | Meaning | Example in Economics |
|---|---|---|
| M | Memory | Previous iterations, past conversations, documented decisions |
| A | Assets | Data files, papers, code, model outputs |
| A | Actions | Tools AI can use (search, compute, visualize) |
| P | Prompt | Your instruction (using AIM) |
In our last conversation, you helped me identify
endogeneity concerns in my IV strategy. You suggested
testing for weak instruments using F-statistics.I'm attaching:
- My Stata output showing first-stage results
- The dataset description (15,000 firms, 10 years)
- Relevant literature summary from my lit reviewPlease search recent econometrics papers on weak
instrument diagnostics and compute the effective
F-statistic from my output.[Use AIM structure here for specific request]How do I clean my dataset?Result: Generic advice about missing values
MEMORY: Yesterday we discussed my wage data structure
and identified issues with outliers in the top 1%.
ASSETS: [Attach CSV with 50K observations, variables:
wage, education, experience, industry, year]
ACTIONS: Analyze the distribution of wages, identify
outliers using economic reasoning (not just statistical).
PROMPT (AIM):
Actor: You are a labor economist experienced with
microdata cleaning and wage distribution analysis.
Input: My data has wages ranging from $2/hr to
$5,000/hr (clearly problematic). Sample is
representative of US workers 2015-2023.
Mission: Propose three defensible approaches to
handle outliers that I can justify in my methods
section. For each: (1) the rule, (2) economic
rationale, (3) expected impact on sample size.You’re implementing gradient boosting but getting an error:
My code has an error, fix it.Why this fails: - No code provided - No context about what you’re trying to do - No information about your data - AI will guess randomly
ACTOR: You are an expert in Python machine learning
libraries (scikit-learn, XGBoost) and debugging data
preprocessing pipelines.
INPUT: I'm implementing gradient boosting for house
price prediction. My dataset has 50K rows, 75 features.
I've already handled missing values with median
imputation. Here's my code:
[paste your code]
Error message: "ValueError: Input contains NaN,
infinity or a value too large for dtype('float64')"
MISSION:
1. Diagnose the specific cause (walk through the logic)
2. Suggest exactly which line is problematic
3. Provide corrected code with explanation
4. Recommend a validation step to prevent this in futureScenario: You’re confused about bias-variance tradeoff in the context of regularization.
Explain bias-variance tradeoffACTOR: You are a statistics professor teaching ML
to economics graduate students. You excel at
connecting ML concepts to econometric intuition.
INPUT: I understand that regularization (like LASSO)
adds a penalty term to prevent overfitting. I'm
confused about why this is called a "bias-variance
tradeoff" rather than just "reducing variance."
MISSION: Explain the bias-variance tradeoff in the
context of LASSO, using:
1. Mathematical intuition (no heavy proofs)
2. An analogy to omitted variable bias in regression
3. A concrete example with economic data
4. When I should prefer bias over variance in applied workYou need to synthesize 20 papers on unemployment duration modeling.
Actor: Specify field expertise
Memory: Build across multiple conversations
Assets: Provide paper abstracts/summaries
Actions: Ask AI to search recent developments
Step 1: Identify Experts
List the top 10 researchers and 5 seminal papers on
unemployment duration modeling and machine learning
methods. Include why each is influential.Step 2: Synthesize with Context
ACTOR: You are an expert in labor economics and
duration analysis, familiar with both traditional
hazard models and modern ML approaches.
ASSETS: [Provide summaries of your 20 papers]
MISSION: Create a 3-paragraph synthesis organizing
these papers by: (1) Traditional approaches,
(2) ML innovations, (3) Current debates.
Identify 2 gaps in the literature.You ran a neural network and got 89% accuracy. Now what?
Don’t ask: “Is this good?”
Do ask:
ACTOR: You are an ML expert who evaluates model
performance critically in economic contexts.
INPUT: My neural network predicts loan default with
89% accuracy on test set. Dataset: 100K loans,
15% default rate (imbalanced classes). Features:
credit score, income, debt ratios, employment history.
MISSION: Evaluate this 89% accuracy critically:
1. Is it meaningful given class imbalance?
2. What other metrics should I report?
3. What does this accuracy mean economically?
4. What robustness checks should I run?If the output isn’t what you need,
debug your prompt, not the AI
Before answering, think step by step about:
1. What economic theory applies here
2. What assumptions you're making
3. What potential confounds exist
Then provide your final recommendation.Use when: You need to see AI’s reasoning process
Before solving my problem, ask me 3 clarifying
questions about:
- My data structure
- My research question
- My constraints
Ask them one at a time. Then synthesize my answers
and provide your response.Use when: Your initial prompt might be ambiguous
Before answering, propose two sharper, more specific
versions of my question. Explain why each is better.
Ask which I prefer (or suggest a hybrid).Use when: You’re not sure how to frame your question
The goal: Iterative improvement—you and AI learning together
What features should I include?I'm predicting firm bankruptcy using financial ratios.
Before suggesting features, think step by step:
1. What economic theory predicts bankruptcy?
2. What data limitations might I face?
3. What multicollinearity concerns exist?
4. What temporal dynamics matter?
Then suggest 5 feature categories with rationale.AI now provides theoretically grounded, contextually appropriate feature suggestions rather than a generic list
AI learns from millions of sources: - Some brilliant - Some average
- Some wrong
Vague prompts sample from the average.
Expert-steered prompts sample from the best.
How do I model time series data?Explain time series modeling approaches using insights
from:
- Hamilton's "Time Series Analysis" framework
- Recent advances in machine learning for macro
forecasting (Coulombe et al.)
- The Fed's approach to nowcasting (NY Fed Staff Reports)
Compare traditional ARIMA vs. modern ML methods
(LSTM, Random Forest) for my quarterly GDP forecasting
problem with 40 years of data.When you don’t know the experts:
Step 1: List the top 5 researchers, 3 key papers,
and 2 recent surveys on [your topic]. Include why
each is influential.
Step 2: Using these sources, synthesize the current
best practices for [your specific problem]. Identify
any debates or trade-offs in the literature.This turns AI from echo chamber to expert synthesizer
Can I use random forests for causal inference?ACTOR: You are an econometrician familiar with both
the potential outcomes framework and modern ML methods.
INPUT: I want to estimate treatment effects of job
training programs using observational data. I have
rich pre-treatment covariates (50+ variables).
MISSION: Explain causal forests (Athey & Imbens) vs
traditional propensity score matching. Reference:
- Athey & Imbens (2016) on recursive partitioning
- Wager & Athey (2018) on causal forests
- Recent applications in labor economics
Compare: (1) identification assumptions,
(2) practical implementation, (3) when to prefer
each method for my use case.AI sounds equally confident
when right and when wrong
As researchers, verification is your job, not AI’s
List every assumption you made in your response.
Rank each by confidence (high/medium/low).Cite two independent academic sources for each major
claim. Include: paper title, authors, year, and a
direct quote.Find one credible source that disagrees with your
recommendation. Explain the trade-off.Recompute every figure you provided. Show your
step-by-step math or code. Explain any rounding.Critical for: Statistical calculations, effect sizes, model comparisons
If models disagree → red flag → investigate deeper
AI says: “Your R² of 0.73 is excellent for social science data”
Your verification:
1. What assumptions are you making about my model type,
field, and prediction context?
2. Cite two sources that provide benchmarks for R²
in applied economics
3. Find one source suggesting when high R² might be
problematic (overfitting, data leakage)
4. Reconsider your assessment given that my model has
100 features for 500 observationsAI is a research assistant, not a research supervisor
Most students use AI like a vending machine: 1. Input generic question 2. Copy generic answer
3. Submit generic work
Professors can spot this instantly.
Make AI outputs sound like you and reflect your thinking:
| Letter | Quality | What to Ask |
|---|---|---|
| O | Original | “Give 3 non-obvious angles” |
| C | Concrete | “Back claims with real examples” |
| E | Evident | “Show your reasoning in bullets” |
| A | Assertive | “Take a stance, defend it” |
| N | Narrative | “Structure as: problem → insight → action” |
"Machine learning can improve economic forecasting
through better pattern recognition and handling of
large datasets. This is important for policy decisions."❌ Obvious, vague, no personality
O: What non-obvious insight exists here?
"Most macro forecasting focuses on prediction accuracy,
but ML's real value may be in uncertainty quantification
during regime changes—something traditional models fail at."
C: Back with concrete example?
"The 2008 crisis: DSGE models gave point forecasts,
ML ensembles would have shown widening prediction intervals."
E: Show reasoning?
"If economic relationships become unstable → parameter
uncertainty increases → ensemble disagreement rises →
this is measurable signal traditional CI's miss"✅ Original, specific, reasoned
Use it iteratively:
ACTOR: You are my critical research advisor
INPUT: [paste your draft paragraph]
MISSION: Evaluate this paragraph using OCEAN framework:
- O: Is the idea non-obvious? What's surprising?
- C: Do I use specific examples/data?
- E: Is my reasoning chain clear?
- A: Do I take a clear position?
- N: Does it flow logically?
For each weakness, propose one specific revision.The goal: AI accelerates your work without replacing your thinking
Common Question: “Which algorithm should I use?”
Better Approach with AIM + Expert Steering:
ACTOR: You are an ML researcher who specializes in
prediction problems in social science contexts, where
interpretability often matters as much as accuracy.
INPUT: I'm predicting student dropout using:
- 10,000 students, 40% dropout rate
- 50 features: demographics, grades, attendance,
financial aid, engagement metrics
- Goal: Both predict AND understand key risk factors
- Need to present to non-technical policy makers
MISSION: Compare 3 approaches (Logistic, Random Forest,
XGBoost) on two dimensions:
1. Expected performance vs interpretability trade-off
2. Practical implementation for my specific goal
Recommend one with 3 concrete reasons. Include one
limitation I should acknowledge.What hyperparameters should I tune for XGBoost?ACTOR: You are an ML practitioner experienced with
gradient boosting and computational constraints in
academic research.
INPUT: XGBoost for house price prediction
- 50K observations, 75 features
- Running on laptop (8GB RAM, no GPU)
- Need results within 2 hours for deadline
MISSION: Prioritize the top 3 hyperparameters to tune
that will give me the most performance gain. For each:
1. What it controls (in economic intuition)
2. Reasonable range to search
3. Computational cost of tuning it
Suggest a grid search strategy that fits my constraints.Challenge: You have 5 million transaction records and need to engineer features.
Effective Strategy:
ACTOR: You are a data engineer experienced with
pandas/dask for large economic datasets.
INPUT: 5M credit card transactions, need to create
customer-level features for fraud detection:
- Current approach: pandas groupby → memory error
- Feature ideas: transaction frequency, amount stats,
merchant diversity, temporal patterns
MISSION:
1. Diagnose why my current code fails (memory management)
2. Provide dask/chunking strategy with code example
3. Suggest 3 additional features I'm missing
4. Estimate runtime for my dataset sizeScenario: You’re checking regression assumptions and see heteroskedasticity.
ACTOR: You are an econometrics TA who helps students
understand diagnostic tests and remedies.
INPUT: My OLS regression shows:
- Breusch-Pagan test: p < 0.01 (reject homoskedasticity)
- White test: p = 0.03
- Residual plot: clear fan shape
- DV: log(wages), IVs: education, experience, controls
MISSION:
1. What is the practical implication (not just
"biased standard errors"—explain economically)
2. Compare 3 solutions: robust SE, WLS, transformation
3. Recommend one for my context with justification
4. Provide Stata/Python code for implementationGoing from generic to excellent:
How do I forecast GDP?ACTOR: You are a macroeconomist familiar with both
traditional time series and modern ML forecasting
methods (e.g., Coulombe et al. 2020 review).
MEMORY: We previously discussed my quarterly GDP data
(1980-2023) and identified potential structural break
in 2008.
ASSETS: [attach: ACF/PACF plots, ADF test results,
basic ARIMA output showing AR(2) model]
MISSION:
1. Evaluate if my AR(2) is appropriate given 2008 break
2. Compare AR(2) vs LSTM vs Random Forest for my
16-quarter ahead forecast
3. Recommend approach balancing: accuracy,
interpretability, uncertainty quantification
4. Suggest how to handle COVID period (2020-2021)
Reference relevant papers for each recommendation.*Unless explicitly permitted
In your methods section:
"I used ChatGPT (GPT-4, OpenAI) to assist with
debugging Python code for data preprocessing and to
clarify the mathematical intuition behind gradient
boosting. All code was reviewed, tested, and modified
based on my dataset's specific requirements."In your acknowledgments:
"I used Claude (Anthropic) to help structure my
literature review and refine explanations of technical
concepts. All interpretations and conclusions are
my own."If you’re unsure whether your AI use is appropriate,
ask your instructor before submission
Better to clarify than to face integrity violations
Use AI to:
- Synthesize recent papers
- Identify methodological gaps
- Generate research questions
- Create work plan for weekUse AI to:
- Debug code iteratively
- Explain error messages
- Optimize performance
- Understand new methodsUse AI to:
- Interpret results with economic lens
- Draft explanations (then revise heavily)
- Create visualizations
- Structure argumentsUse AI to:
- Critique your own work
- Verify calculations
- Check robustness
- Polish presentationGolden Rule: If you can’t explain the code to a classmate, you don’t understand it yet.
Goal: AI accelerates, you think and write
For complex analyses, break into steps:
Step 1: "Explore the data structure and identify
potential issues"
Step 2: [Based on output] "Given these issues,
propose 3 preprocessing strategies"
Step 3: [After your choice] "Implement strategy 2
with Python code, explaining each line"
Step 4: [After running] "Interpret these results
and suggest next steps"Why this works: Each step builds context for the next
Technique: Have AI play different expert roles
Round 1: "You are a skeptical econometrician. Critique
my identification strategy. Be harsh."
Round 2: "You are a supportive methods expert. Given
the critique above, suggest improvements."
Round 3: "You are a journal referee. Evaluate the
revised strategy. What would you still question?"Result: Better work through multiple perspectives
Original: Explain your code to a rubber duck to find bugs
AI Version:
I'm going to explain my entire analytical pipeline
step-by-step. After each step, you ask me:
1. Why did I make this choice?
2. What assumption does this depend on?
3. What could go wrong here?
This will help me find gaps in my reasoning.
[Then explain your pipeline]For model selection or methodological choices:
Create a comparison table with these columns:
| Method | Strengths | Weaknesses | Best Use Case | Implementation Difficulty |
Compare these 4 approaches for my problem: [describe]
1. [Method 1]
2. [Method 2]
3. [Method 3]
4. [Method 4]
Rank them for MY specific context and justify top choice.For checking your understanding:
I need to explain [complex ML concept] to undergraduate
economics students who only know basic statistics.
First, explain it to me at graduate level with proper
terminology.
Then, create an explanation using:
- No jargon
- An economics analogy
- A concrete example
- Why they should care
I'll use this to test if I really understand it.Bad: “What’s the right model for my data?”
Good:
"I'm deciding between models A and B. Here are the
tradeoffs I see: [list]. What am I missing? What
assumptions differentiate them? When has each failed
in practice?"AI is a discussion partner, not an authority
One-shot prompting rarely works well
Think of it as a conversation, not a search engine
AI “forgets” after too many exchanges.
Solution:
Let me summarize our conversation so far:
[2-3 sentences capturing key decisions/findings]
Given this context, [new question]Or start fresh with complete context using MAP framework.
AI hallucinates academic references
If you cite a fake paper, that’s on you, not AI
Each of these fails: - “Explain regression” - “Fix my code”
- “Is this good?” - “Help with my thesis”
Why? No Actor, Input, or clear Mission
Solution: Always use AIM framework
Week 1 vs Week 4 Comparison:
| Metric | Week 1 | Week 4 Goal |
|---|---|---|
| Prompt quality | Generic | AIM-structured |
| Iterations needed | 5-8 | 2-3 |
| Output usefulness | 40% | 85% |
| Time to good result | 30 min | 8 min |
| Understanding | Shallow | Deep |
| Independence | Low | High |
General Purpose: - ChatGPT Plus (GPT-4) - Claude (Anthropic) - Gemini Advanced
Specialized: - GitHub Copilot (coding) - Cursor (AI code editor)
Research Specific: - Elicit (literature review) - Scholarcy (paper summaries) - Connected Papers (visualization)
Data Analysis: - Julius AI (data chat) - DataCamp’s AI assistant
For this course:
Office Hours: Bring your AI conversations—we’ll debug together!
AI is not here to replace your thinking.
It’s here to amplify your economic reasoning.
Due before next class:
Submit via: [Course LMS]
Every prompt you write,
every revision you push,
every judgment you make…
You’re training yourself,
not just the AI
AI restores human worth by:
Automating the routine,
so you can focus on the creative,
the critical,
the uniquely human
Let’s discuss:
Instructor Office Hours:
[Day/Time, Location]
TA Office Hours:
[Day/Time, Location]
Course Resources: - Slack channel: #ai-help - Shared prompt library: [Link] - Example notebooks: [Link]
Remember: Using AI effectively is a skill.
Give yourself grace as you learn!
Stay curious. Stay critical. Stay human.
[Your Email]
[Course Website]
ACTOR: You are [specific expertise + context about quality]
INPUT: [Your specific situation, data, constraints,
what you've tried, what's not working]
MISSION: [Specific output requested, format,
any constraints or requirements]Use for: Every significant prompt
MEMORY: [Previous conversation context, decisions made]
ASSETS: [Files, data, papers, code you're providing]
ACTIONS: [Tools AI can use: search, compute, analyze]
PROMPT: [Your AIM-structured question]Use for: Complex, multi-part questions
After receiving AI output, push it:
Use for: Making outputs reflect your thinking
ACTOR: Expert Python/R debugger
INPUT: [code + error + data context]
MISSION: Diagnose cause, suggest fix, prevent futureACTOR: Professor teaching economists
INPUT: [concept + your confusion]
MISSION: Explain with economic intuition + exampleACTOR: ML practitioner in social science
INPUT: [problem + data + constraints]
MISSION: Compare 3 methods, recommend with rationaleML & Big Data in Economics