Economist interview questions for AI training work
AI training platforms hire people with a Economist background to evaluate AI outputs in that field, checking whether an answer is factually sound, appropriately reasoned, or safe to act on in ways a generalist reviewer couldn't judge. The screening interview is built to confirm that expertise, drawing on Economic modeling and forecasting, Policy and market analysis and Explaining findings to non-economists.
The 10 questions come from that kind of interview, in technical, scenario and behavioral rounds, each with a full written answer that shows what a strong response sounds like.
Technical (5)
How do you choose between a structural model and a reduced-form approach when you are asked to forecast a variable?
I start from the question being asked, since a reduced-form model forecasts well when the underlying relationships are stable, while a structural model is needed when a policy change will shift those relationships. If the client wants to know what happens under a new rule, I accept the extra assumptions of a structural model because the simpler one would forecast a world that no longer exists.
What steps do you take to check that a forecasting model is not overfitting the historical data it was estimated on?
I hold out the most recent period, estimate on the rest, and compare the out-of-sample errors to a naive benchmark such as a random walk or a simple average. If the model cannot beat the benchmark out of sample, the extra parameters are fitting noise, and I would rather ship a simpler model with honest error bands than a complex one that looks precise in-sample.
How do you build a sensitivity analysis around a central forecast so the ranges mean something to a decision maker?
I identify the two or three inputs that drive most of the variance, then vary each over a range grounded in historical swings rather than arbitrary percentages. I present the results as scenarios with a short story attached to each, because a decision maker can act on a scenario they can picture but not on a bare confidence interval.
When you assess the likely effect of a proposed regulation on a market, how do you separate the direct effect from behavioral responses?
I map out who the rule changes incentives for, then look for comparable past rules or natural experiments where those agents already responded. The direct mechanical effect is usually easy to estimate, while the behavioral response is where most of the uncertainty sits, so I state it separately with its own range instead of folding it into a single number.
How do you translate a statistically significant result into a statement about whether the effect matters in practice?
I convert the coefficient into units the reader cares about, such as dollars per household or percentage points of employment, and compare it to the size of other known effects. A result can be significant and still too small to change any decision, so I lead with the magnitude and keep the p-value as supporting detail.
Scenario (3)
A minister wants a one-page summary of a 60-page forecast report by tomorrow morning. How do you decide what goes on the page?
I would put the central forecast, the single biggest risk to it, and the one decision the numbers bear on, in that order, with a small chart instead of a table. Everything that supports those three points goes in a footnote with a reference to the full report, since the page exists to let them act rather than to prove I did the work.
Two credible data sources give you conflicting readings on the same indicator right before a deadline. How do you proceed?
I would check whether the gap comes from definitions, timing, or sample coverage, because that explains most such conflicts and often tells me which source fits my question. If I cannot resolve it in time, I report both readings and show how my conclusion changes under each, which is more useful to the reader than quietly picking one.
Your model's forecast from last quarter turned out to be badly wrong and a client is asking why. How do you respond?
I would decompose the miss into the part caused by inputs that came in differently than assumed and the part caused by the model's own structure, since those call for different fixes. Then I would walk the client through that split and what I am changing, because clients forgive a wrong forecast more readily than a vague explanation of one.
Behavioral (2)
Tell me about a time you had to explain an uncomfortable economic finding to an audience with no technical background.
I presented analysis showing that a popular local subsidy program was mostly benefiting households that would have made the same purchase anyway. I framed it around three example families rather than the regression output, and while the room was skeptical at first, the program was redesigned with an income cap the following year.
Describe a time you changed your own forecast after someone outside your team challenged an assumption in it.
A colleague from the operations side questioned my assumption that supplier lead times would stay flat, pointing to contracts they had seen being renegotiated. I built a scenario with longer lead times, found it moved the cost forecast by more than my entire error band, and made it the base case. The revised number ended up much closer to what happened.
Practice saying your answers out loud
Knowing an answer and saying it under pressure are different skills. The Academy has free modules and mock exams for that, and practice tasks shaped like the work itself.


