Situations machine learning

Situations: machine learning

Modules 2 and 3, Applied Machine Learning and Deep Learning. The situations where the model is excellent and the system is useless.

The pattern in hard ML interviews is consistent: the candidate is handed a problem, reaches for a model, and never interrogates the label, the split, the operating point or the cost of being wrong. Published guidance on ML interview cases says the same thing, that candidates are tested on whether they can pick metrics, build honest validation and avoid leakage under realistic constraints, and that a common failure is claiming strong performance without validating properly (Data Interview, BuildML, checked 10 Sep 2026).

Read The Situation Bank first for the card format and the difficulty ladder.

---

The five mechanisms

flowchart TB
    Q["a model performs well offline"]
    Q --> A["<b>label</b><br/>what does the target<br/>actually mean, and when<br/>does it become known?"]
    Q --> B["<b>leakage</b><br/>does a feature know<br/>something the moment<br/>of prediction cannot?"]
    Q --> C["<b>split</b><br/>does validation resemble<br/>how it will be used?"]
    Q --> D["<b>operating point</b><br/>who acts on the output,<br/>and how many can they act on?"]
    Q --> E["<b>cost</b><br/>are the two errors<br/>the same size?"]

Loading

---

Flagship cards

S-FIN-ML-01 · Five hundred alerts a day

Unit	Kalpa Financial Services
Level	L2 Confounded, escalating to L3
Asked by	Head of Fraud Operations. She has nine analysts and no budget for a tenth.
Origin	seed. Capacity-constrained fraud triage is a standard hard interview case; see the sources above.

What they say

"The model the last team built has 99.8 percent accuracy and it is useless. My analysts are drowning and we are still paying out chargebacks. Fix it."

What is in the room

Eighteen months of transactions with a chargeback_received flag. Analyst review logs with a disposition per reviewed case. The team of nine, who between them can work about 500 alerts a day against roughly 300,000 daily transactions.

INTERNAL: the twist, and what to listen for

What is actually true

Three separate problems, and only the first is the one they think they have.

1. The metric is the wrong shape. At roughly 0.2 percent fraud, a model that says "never" is 99.8 percent accurate. Accuracy cannot see this problem at all.
2. The capacity is the product. The business question is not "is this fraud" but "which 500 of 300,000 do nine people look at today". That is precision in the top 500, not AUC.
3. The label lags. A chargeback lands 45 to 90 days after the transaction. The most recent three months of training data are labelled as clean because nobody has complained yet, so the model learns that recent patterns are safe.

Why the obvious read fails

Every one of these is invisible to a leaderboard metric. A model can win on AUC and be unusable.

The first question a strong candidate asks

"How many alerts can your team actually work in a day?" A candidate who asks about capacity before asking about features is thinking about the system. The second question should be "when do you find out a transaction was fraudulent?"

The KPI that settles it

Precision at 500, plus value recovered per analyst hour. Not AUC, not F1, and not accuracy.

Where it turns technical

Ranking rather than classifying, a threshold set from a capacity constraint rather than from 0.5, and a training set that excludes the immature label window.

How it escalates

To L3. Analysts only ever review what the model surfaced, so the disposition labels exist only for cases the model already liked. Every retrain learns from its own past decisions. Breaking that needs a random holdout that gets reviewed regardless of score, which costs capacity the Head of Fraud does not have. The candidate must argue for spending scarce analyst time on cases the model thinks are fine, to a person who is already drowning. That negotiation is the real test.

To L4. Ask what happens to the denominator when fraud is prevented. A blocked transaction never becomes a chargeback, so a successful model destroys its own future labels and looks like it is degrading.

---

S-HLT-ML-02 · The feature that already knew

Unit	Kalpa Health
Level	L2 Confounded
Asked by	Clinic operations director, who wants to phone likely no-shows the day before
Origin	built on target leakage through an operationally-populated field

What they say

"The no-show model is at 0.94 AUC in testing and it is barely better than a coin flip in the clinic. The data science team says the data must have changed."

What is in the room

Appointment records with booking time, patient history, clinic, slot, and a set of operational fields including reminder_call_outcome, slot_released, rebooked_flag. The training notebook.

INTERNAL: the twist, and what to listen for

What is actually true

slot_released is set by the front desk when a patient fails to arrive. It is in the training data, it is the single strongest feature, and it does not exist at the moment the prediction has to be made, which is the evening before. The model is reading the answer.

reminder_call_outcome is subtler and worth catching too: it is populated the day before, so it is legitimately available, but "no answer" correlates with no-show partly because the same people are unreachable. That one is a real feature with a fairness question attached, not leakage.

Why the obvious read fails

Nothing about the data changed. The offline number was never real. A validation split that shuffles rows cannot see this, because the leaked field is present on both sides of the split.

The first question a strong candidate asks

"For each feature, at what moment does its value become known?" The discipline is a point-in-time audit of every field against the moment of prediction. Candidates who talk about correlation with the target rather than availability at inference are guessing.

The KPI that settles it

Performance on a strictly time-forward split, using only fields whose values are frozen as of the evening before the appointment.

Where it turns technical

Reconstructing the feature set as of a timestamp, which is the same skill as building a point-in-time feature store. This is where a build week can go deep.

How it escalates

To L3. Once leakage is removed, the model is weak but not worthless. The director asks whether it is worth deploying. Calling 200 people to prevent 30 no-shows may or may not pay, and the answer depends on the cost of an empty slot versus a receptionist's hour. The candidate has to build that arithmetic, not a better model.

To L4. Ask what happens to the label after deployment. A patient who is called and then attends was going to be a no-show and is now recorded as an attendance. The intervention destroys the label it was trained on, and measuring the model after launch needs a holdout that is deliberately not called. Explaining why that is ethically defensible in a clinic is part of the answer.

---

S-LOG-ML-03 · The orders that have not arrived yet

Unit	Kalpa Logistics
Level	L3 Contested
Asked by	Head of Last Mile, who has promised the retail unit a delivery-time estimate on the product page
Origin	built on right-censoring in an operational prediction target

What they say

"Predict delivery time from order to doorstep. We want to show it on the product page. The data is all there, two years of it."

What is in the room

Order timestamps, dispatch timestamps, delivery timestamps where present, pincode, courier partner, weight band, and a status field. About four percent of rows have no delivery timestamp.

INTERNAL: the twist, and what to listen for

What is actually true

The four percent without a delivery timestamp are not missing data. They are orders that were never delivered: lost, returned to origin, or still in transit after months. They are the worst outcomes, and dropping them trains the model on a world where nothing ever goes badly wrong.

The prediction shown on a product page is also not a mean. A promise that is beaten half the time is a promise broken half the time.

Why the obvious read fails

dropna() looks like hygiene and is actually a decision to ignore every failure. The mean absolute error will look excellent and the customer promise will be wrong in exactly the cases that generate complaints.

The first question a strong candidate asks

"What does a missing delivery timestamp mean?" The general form: "Is this missing at random, or is missingness the outcome?"

The KPI that settles it

Not MAE. The business needs a quantile: the time by which, say, 90 percent of orders arrive, because that is what can be promised. And a separate model or rule for the probability of not arriving at all.

Where it turns technical

Quantile loss rather than squared error, and censored observations handled rather than dropped. This is where survival analysis earns its place in a curriculum that otherwise would not justify it.

How it escalates

To L3, which is where this card lives. Retail wants an aggressive number because it converts. Operations wants a conservative one because it is what gets missed. The model can produce any quantile; choosing which one to show is a commercial decision with a cost on both sides, and the candidate must present the trade-off rather than pick silently.

To L4. The displayed promise changes behaviour. A longer quoted time loses the impatient customer, so the orders that get placed are different, so the delivery mix changes, so the model drifts because of its own output. Naming that feedback loop before it happens is the signal.

---

Seeded situations, ready to expand

ID	Unit	Level	What they say	The twist underneath
S-FIN-ML-04	Financial Services	L2	"Credit model degraded after March"	Underwriting policy changed in March. The population is different; the model is fine and the world moved.
S-FIN-ML-05	Financial Services	L3	"The model is not using any protected attribute"	Pincode plus device type reconstructs it. Absence from the feature list is not absence from the model.
S-FIN-ML-06	Financial Services	L4	"Rejected applicants default less"	They were never given a loan, so they cannot default. The rejected population has no labels at all, and the model is being evaluated on survivors.
S-RET-ML-07	Retail	L2	"Recommendation CTR beat the baseline"	The baseline was shown in a different slot on the page. The layout, not the model, moved the number.
S-RET-ML-08	Retail	L3	"Demand forecast is accurate to 3 percent"	Accurate on the aggregate, useless per SKU per store, which is the only level anyone orders stock at.
S-RET-ML-09	Retail	L4	"The model predicts returns well"	It has learned the seller, and the seller is a proxy for a category the business is about to exit. The feature will vanish next quarter.
S-CON-ML-10	Connect	L2	"Churn model recall is 80 percent"	The label is "cancelled within 30 days", but retention offers go out at 45 days. It predicts churn nobody can act on in time.
S-CON-ML-11	Connect	L3	"We should target the highest churn scores"	The highest scorers have already decided. The value is in the persuadable middle, which needs uplift modelling, not propensity.
S-CON-ML-12	Connect	L4	"Retraining monthly keeps it fresh"	Retention offers driven by the model change who churns, so each retrain learns from a world the previous model created.
S-LOG-ML-13	Logistics	L2	"Route model works in testing"	Random split across time. Trained on December, tested on December, deployed in monsoon.
S-HLT-ML-14	Health	L3	"The escalation model has 92 percent precision"	The two errors are not equal. A missed escalation and a false alarm cost different things, and precision alone cannot express that.
S-HLT-ML-15	Health	L4	"Performance is stable across sites"	It is stable because one large site dominates the average. At the three smallest sites it is worse than the rule it replaced.

---

The question to teach before any model

Every card on this page is answered faster by a candidate who asks these four, in this order, before touching an algorithm:

1. What exactly is the label, and when does its value become known?
2. At the moment of prediction, which of these features actually exists?
3. Who acts on the output, and how many outputs can they act on in a day?
4. What does each of the two errors cost, in the units the business already tracks?

A candidate who asks all four has usually found the trap before the interviewer reveals it. That is the intuition this bank exists to build.