Duration
About 3.5 hours
You set every stage timing from the professor panel, so it fits one block or splits across two sessions.
AI in Financial Decision-Making Simulation
AI simulation where students set the decision threshold
The Finsimco AI in Financial Decision-Making Simulation hands each student a credit model that is right ninety-four per cent of the time and losing nineteen million dollars a year.
- About 3.5 hours, single-player
- Thresholds, error cost and oversight
- In-class, online or hybrid delivery
At a glance
Format
Single-player simulation
Each student holds the data, threshold and oversight roles in turn and submits one decision policy.
Level
Undergraduate to MBA
Adjustable: one model on a clean dataset, up to segments, a proxy variable, a regulator and model drift.
Professor tools
Timeline and session controls
Manage students, adjust stage timings, control progression and monitor individual inputs.
Prerequisites
None
No prior knowledge. Thresholds, precision and recall, error costs and selection bias are taught inside the activity.
Delivery
In class, online or hybrid
The web-based simulation supports in-person, remote, blended and homework-supported delivery.
Assessment
Optional, per student
Every submission is individual, so the threshold, the oversight policy and the review answers export per student.
Course fit
Fits finance and analytics modules
Designed for Financial Technology, Banking, Data Analytics, Risk Management and Information Systems courses.
Concepts covered
The topics an ai in financial decision-making module has to cover: the expected cost of the two error types, the decision threshold, precision and recall against accuracy, selection bias in training data, human oversight and override logs and model drift after deployment.
The decision threshold runs from zero to one along the bottom. As it rises the model approves fewer applications, so the cost of the defaults it lets through falls steeply and then flattens, while the cost of the good business it turns away rises throughout. The sum is U-shaped, with a minimum at about zero point three eight costing ten point two million dollars a year, against eighteen point nine at the threshold the model shipped with. Below, four hundred and ten human overrides: the human was wrong more often than right.
Learning outcomes
Written to drop into your module descriptor without rewriting. On completing the simulation, your students will be able to:
- Convert a model's error rates into expected cost in currency.
- Explain why accuracy is not the objective and locate the cost-minimising threshold.
- Identify selection bias in a training sample drawn from past approvals.
- Design an oversight policy that specifies when a human may override and why.
- Explain what an override log reveals about a model's blind spots.
- Evaluate a model deployment on the decisions it produced rather than on its test metrics.
What students do
Each student works the model alone, holding the data, threshold and oversight roles in turn. All three want something different from the same model, and only one policy is deployed.
The gameflow every student sees before they start. The three roles come into play while the errors are priced and through the year; briefing, the assessment and the debrief are worked as a single view. Materials arrive stage by stage: model card and role cards at briefing, then the error costs, threshold model and training sample, then the decision log, override cases and drift monitor.
Data
Students establish what the model was trained on, which is a record of applications that were approved and says nothing about the ones that were not.
Threshold
Students choose where to approve, which is a business decision about the relative cost of two mistakes and is usually left to whoever shipped the model.
Oversight
Students decide when a person may overrule the model and record what happens, because that log is the only map of where the model is blind.
How it works
Five stages. Students price the errors, set the cut-off and read the year of decisions it produced.
- 1
Stage 1 · 15 min
Briefing Docs · Model Card · Role Cards
1Briefing
Students watch the intro video, review a credit model with a ninety-four per cent accuracy figure, a shipped threshold nobody chose deliberately and a portfolio of four hundred thousand applications a year, and take on all three roles before any decision is changed.
- 2
Stage 2 · 55 min
Error Costs · Threshold Model · Training Sample · Tool Tips
2Price the Errors
The errors are priced. Working as data, the student finds that the training sample contains only approved applications; as threshold, converts default cost and lost margin into dollars and locates the minimum of the total; as oversight, writes the rule under which a person may overrule the model and what they must record.
- 3
Stage 3 · 55 min
Decision Log · Override Cases · Drift Monitor · Live Scores
3The Year
A year of decisions runs. Four hundred and ten overrides are logged and the humans are wrong more often than right, but the cases they get right cluster in a population the model never saw. The model drifts in the second half of the year, and a regulator asks why a particular applicant was declined.
- 4
Stage 4 · 40 min
Review Questions · Cost Report · Outcome Report · Review Pack
4The Assessment
The policy is assessed. Students are asked what their threshold cost against the best available, what the training sample could never have told them, which overrides were worth keeping, and what they would say to the regulator about the declined applicant.
- 5
Stage 5 · No limit
Winners · Completion Summary
5Debriefing
Students review their own decision policy, compare thresholds and costs across the cohort, revisit the overrides that separated them and reflect on the feedback. Winners and the completion summary give you a structure for the discussion.
What the experience looks like
The student interface carries the costs, the threshold and the overrides. The professor panel carries the timeline, the controls and every individual cost report.
Threshold
Students work cut-off by cut-off, with the approval rate, the cost of defaults let through and the cost of good business declined side by side, and a written reason required for the threshold set.
Year
The override column is the one worth keeping. Humans were wrong more often than right, and the cases they got right were all in a segment the model had never seen.
Professor Controls
Your view. Adjust any stage timing, start or stop the session, and compare the threshold each student chose with what their policy cost over the year.
See how this simulation fits your module
Explore the student and professor experience, and discuss timing, setup and syllabus fit.
Where it fits your syllabus
It works best as the session that puts a currency on a confusion matrix.
Teach before it
Comfort with percentages and expected values. Thresholds, precision and recall, selection bias and drift are explained inside the activity.
Where it sits
Mid module, typically weeks five to nine, after the basics of modelling.
What it sets up
Model governance, fairness testing, explainability requirements, automated underwriting, and any assessment asking students to appraise a model deployment.
The AI in finance course guide sets out where it sits in a full module.
What students get wrong
The three mistakes students make on this case, and what each one gives you in the debrief. Our simulations are designed around them: read the learning science behind it.
1
They try to improve the model
A better model is the obvious answer and is worth a fraction of what moving the threshold is worth. Students spend the first stage discussing features and the chart at the end shows that the cut-off was carrying almost all of the loss.
2
They trust the training sample
The model has only ever seen applications that were approved, so it has no evidence about the population it is being asked to judge. Students who do not notice this build a policy on a sample that is systematically missing its own counterexamples.
3
They treat overrides as noise
Humans override the model and are wrong more often than right, which makes it tempting to remove the override entirely. The cases they get right are concentrated exactly where the model is blind, and deleting the override deletes the only record of it.
How to debrief it
The debrief runs itself. Every student sees their results, the cohort comparison and the decisions that produced them, with nothing for you to prepare. If you would rather lead the discussion, these questions work straight from a slide, and they use what your cohort actually did.
- What did your threshold cost against the best available one?
- What could the training sample never have told you?
- Which overrides were worth keeping, and why those?
- Who in the room set the lowest threshold, and what did it cost them?
- What would you say to the regulator about the declined applicant?
Assess the cost reasoning, not the accuracy
Assessment is optional. Run the simulation ungraded as an applied activity, or use platform data as supporting evidence. Every student generates error costings, a threshold choice, an oversight policy, an outcome and rationales. Because the simulation is single-player, every input belongs to one named student, so there is no group-work marking to untangle. All of it exports as CSV or PDF, per student, so you can assess reasoning rather than a result.
What professors can assess
- Whether both error types were priced in currency before the threshold was chosen
- Whether the oversight policy specifies what an overriding human must record
How to interpret the score
- A high accuracy figure is not evidence of a good deployment; read the cost curve
- Use the leaderboard for cohort comparison, not for grading
On generative AI: written rationales are exactly what marking teams now worry about. Because every rationale here is tied to a costing the student recorded inside the platform before the year ran, the reasoning can be checked against the evidence rather than taken on trust.
Low risk to try with a class
Students rate it
93% of students rate Finsimco simulations Excellent, across more than 30,000 responses.
You are not on your own
Teaching notes, a setup and facilitation guide and a gameflow diagram come with it, and we support your first delivery directly.
No grading commitment
Run it ungraded the first time. Assessment is optional and can be added once you have seen how your cohort responds.
A bad session is recoverable
Pause or stop at any point and resume later without losing submitted data, so an overrun costs you a timetable slot, not the activity.
Read what lecturers say.
What AACSB and AMBA say about simulations
Ron Tuninga, Managing Director for EMEA at AACSB, and Victor Hedenberg, Membership Director at AMBA, on why simulation-based learning belongs in business education. Watch both interviews.
FAQs
Related simulations
If this one is not quite right for your module, these are the closest alternatives.
Get the course fit sheet
A one-page summary of this simulation, covering stages and timings, roles, concepts, learning outcomes and the debrief questions, plus a full sample teaching pack from one of our live simulations, so you can see exactly what you would be running.
Book a Demo
During the call, we can:
- Show the student and professor experience
- Discuss format, timing and syllabus fit
- Walk through setup, live delivery and optional assessment evidence
- Answer questions from your module team