Probability Essentials for Interviews: Rules, Conditional Probability & Bayes' Theorem

Before the risk flag, a payment looks normal: a customer, a cart, a checkout button. After one signal changes - device, location, transaction velocity - the same payment becomes a probability question: “How likely is fraud, given what I now know?”

Think of probability as a disciplined way to move from the full world of uncertainty to a smaller world shaped by evidence. The essentials are not many: count correctly, avoid double-counting, condition on the right denominator, and update using Bayes when evidence arrives.

Core Explanation: The Rules That Actually Run Probability

Probability is easiest when you stop memorising formulas and start asking: “What is the universe I am counting inside?” Most candidate errors come from using the wrong universe.

1. Complement Rule: Count What Is Easier

If A is an event, its complement means “A does not happen.”

P(not A) = 1 - P(A)

Use it when direct counting is messy. If the probability of at least one customer complaint is hard, calculate “no complaints” first and subtract from 1.

2. Addition Rule: Use It for “OR” Questions

For two events A and B:

P(A or B) = P(A) + P(B) - P(A and B)

Why subtract? Because the overlap gets counted twice. If an app user is both a premium subscriber and a wallet user, adding both groups double-counts that user unless you remove the overlap.

3. Multiplication Rule: Use It for “AND” Questions

For two events A and B:

P(A and B) = P(A) × P(B|A)

Read it in English: probability that A happens, then B happens given A has happened. If the events are independent, then P(B|A) = P(B), so the formula becomes P(A) × P(B).

4. Conditional Probability: Shrink the Denominator

P(A|B) means “probability of A given B has already happened.” The denominator is no longer the whole population. It is only the B-world.

P(A|B) = P(A and B) / P(B), as long as P(B) > 0.

5. Bayes' Theorem: Reverse the Conditional

Bayes' theorem answers the question business teams ask constantly: “Given this signal, how likely is the underlying cause?”

P(H|E) = [P(E|H) × P(H)] / P(E)

Where:

• H = hypothesis, such as “transaction is fraud.”
• E = evidence, such as “risk engine flagged it.”
• P(H) = prior probability before new evidence.
• P(E|H) = likelihood of seeing evidence if hypothesis is true.
• P(H|E) = posterior probability after evidence.

Worked Example: The Fraud Flag That Looks Stronger Than It Is

A payment platform has the following risk signal:

• 1% of transactions are actually fraudulent: P(Fraud) = 0.01
• The flag catches 90% of frauds: P(Flag|Fraud) = 0.90
• The flag wrongly flags 5% of genuine transactions: P(Flag|Genuine) = 0.05

Question: If a transaction is flagged, what is the probability it is actually fraud?

Step 1: Calculate flagged frauds: 0.01 × 0.90 = 0.009

Step 2: Calculate flagged genuine transactions: 0.99 × 0.05 = 0.0495

Step 3: Total flagged transactions = 0.009 + 0.0495 = 0.0585

Step 4: P(Fraud|Flag) = 0.009 / 0.0585 = 15.4%

The flag is useful, but it does not mean “90% chance of fraud.” That is the base-rate lesson.

Definitions You Should Be Able to Say Cleanly

When Probability Becomes a Business Decision: Measures to Track

In analytics roles, probability often appears as a score: churn probability, fraud probability, default probability or conversion probability. The score is only useful if you know how it behaves when converted into decisions.

Case Study: Razorpay and Probabilistic Risk at Checkout

Situation: Indian digital payments run at high speed across UPI, cards, wallets and netbanking. For a payment gateway, the hard problem is not simply detecting suspicious transactions. It is deciding whether a specific transaction should be approved, challenged, routed differently or blocked without damaging genuine customer experience.

The move: Payment risk systems use signals such as device fingerprint, transaction velocity, merchant category, past behaviour, payment instrument, geography and anomaly patterns. In probability language, the decision is: P(fraud | signals). Bayes' theorem explains the discipline behind the logic: start with the base rate of fraud, update it using evidence, and compare the posterior risk with a business threshold.

The lesson: The primary driver of better risk decisions is not one magic fraud rule. It is calibrated probabilistic scoring. Supporting drivers include richer transaction signals, fast feedback loops from chargebacks or disputes, merchant-level risk patterns, and threshold tuning based on business cost. The strategic “so what” is clear: a risk model must optimise both protection and conversion, because excessive false positives can hurt good merchants and genuine customers.

A shallow answer says, “Razorpay uses data to detect fraud.” A placement-ready answer says, “Payment risk is a conditional probability problem: the same signal can mean different things depending on the prior, merchant context and cost of errors.”

How AI Changes Probability Essentials

1. AI makes probability visible in everyday decisions. Modern ML models output probabilities: propensity to buy, probability of default, churn risk, fraud risk, delivery delay risk. The manager's job is not to recite the algorithm; it is to ask whether the probability is calibrated and whether the threshold fits business cost.

2. AI increases the importance of conditional thinking. Segments behave differently. P(churn|new user, low engagement, high complaint history) is not the same as P(churn) for the full customer base. AI systems generate these narrower conditional worlds at scale.

3. Generative AI needs uncertainty checks. LLM answers can sound certain even when evidence is weak. A probability-trained analyst asks: What is the base rate? What evidence changed the belief? What alternative hypothesis explains the same signal?

Interview Relevance

Common Mistake

What to Revise Next

Now that you can combine events and update probabilities, move to the shapes probability takes in real data. Revise these next as a journey from uncertainty to inference: