Prepare for and practice interview questions from Oliver Wyman across topics like Statistics, SQL, Machine Learning and more.

Oliver Wyman Interview Questions

Oliver Wyman Interview Guides

Machine Learning

In which case would you use a bagging algorithm versus a boosting algorithm

Bagging vs Boosting

How would you build the recommendation algorithm for type-ahead search for Netflix

Type-ahead Search

Replaced by QUESTION 791 - Should be deleted

Bias Variance Tradeoff

Describe linear regression to various audiences with different levels of knowledge.

Explaining Linear Regression to Different Audiences

Credit Card Fraud Model

Write a query to select the top 3 departments with at least ten employees and rank them according to the percentage of their employees making over 100K in salary.

Employee Salaries

Select the 2nd highest salary in the engineering department

2nd Highest Salary

Duplicate Rows

Write a query to create a pivot table that shows total sales for each branch by year

Branch Sales Pivot

Create top_ads with the top 3 ads and return the row counts for inner, left, right, and cross joins with ads

Size of Joins

Statistics

Explain what a p-value is to someone who is not technical

P-value to a Layman

What are the assumptions of linear regression?

Assumptions of Linear Regression

How do you detect and handle correlation between variables in linear regression, and what happens if you ignore it

Correlation in Regression

What is the difference between type I and type II errors?

Type I and II Errors

How would you tackle multicollinearity in multiple linear regression

Multicollinearity in Regression

ML Ops

How would you scale training this recommender system to millions of movies and users

Scaling Up Recommender

Design a strategy for monitoring and evaluating a machine learning model's real-time performance after deployment.

Production Model Monitoring

How would you get a model ready for production use and adaptation to new datasets?

Model Deployment Preparation

Design an end-to-end retraining and production pipeline for a machine learning model without dedicated infrastructure or support staff.

Solo ML Deployment

AI & Agentic Systems

Fine-Tuning VS RAG

Design a RAG system that restricts LLM responses exclusively to ingested data, preventing use of pre-trained knowledge.

RAG Strict Source Control

How would you build and justify the components of a Transformer encoder layer in PyTorch for large-scale text data?

Transformer Encoder Layer

Suppose your default risk model has high recall but low precision. What business implications might this have for a mortgage bank?

Suppose your default risk model has high recall but low precision. What business implications might this have for a mortgage bank?

Tradeoffs Between Precision and Recall

You have been tasked with designing three classes: <code>text_editor</code>, <code>moving_text_editor</code>, and <code>smart_text_editor</code>. These classes are to be created with specific functionalities as defined below:

<ol>
<li><code>text_editor</code> class:

<ul>
<li><code>write_line(string:str)</code>: A method which appends a given string to the end of the existing string.</li>
<li><code>delete_line(char_num : int)</code>: A method which deletes <code>char_num</code> number of characters from the existing string, starting from the end. If there are no characters left, the method should do nothing.</li>
<li><code>special_operation()</code>: A method which currently does nothing.</li>
<li><code>get_notes()</code>: A method which returns the internal string.</li>
</ul></li>

<li><code>moving_text_editor</code> class:

<ul>
<li>This class extends <code>text_editor</code>. The <code>special_operation()</code> method is overridden. Initially, the cursor will be at the end of the current string. If <code>special_operation()</code> is called, it moves the cursor to the beginning of the string, any additional appends will be appended to the beginning of the string instead. Calling <code>special_operation()</code> again reverses the cursor operation.</li>
</ul></li>

<li><code>smart_text_editor</code> class:

<ul>
<li>This class extends <code>text_editor</code>. In this class, the <code>special_operation()</code> method is overridden to serve as an undo operation, allowing it to undo an infinite number of operations.</li>
</ul></li>
</ol>

<h3>Input</h3>

<pre tabindex="0" class="chroma"><code>[[&#39;special_text_editor&#39;], [&#39;write_line&#39;, &#39;special_operation&#39;, &#39;write_line&#39;], [&#39;World&#39;, &#39;Hello, &#39;]]
</code></pre>

<h3>Output</h3>

<code>&#34;Hello, World&#34;</code>

Create a text editor while displaying method overriding as a concept.

Text Editor With OOP

Design Patterns

Write a SQL query to select the 2nd highest salary in the engineering department.

Note: If more than one person shares the highest salary, the query should select the next highest salary.

Example:

Input:

<code>employees</code> table

<table>
<thead>
<tr>
<th>Column</th>
<th>Type</th>
</tr>
</thead>

<tbody>
<tr>
<td><code>id</code></td>
<td>INTEGER</td>
</tr>

<tr>
<td><code>first_name</code></td>
<td>VARCHAR</td>
</tr>

<tr>
<td><code>last_name</code></td>
<td>VARCHAR</td>
</tr>

<tr>
<td><code>salary</code></td>
<td>INTEGER</td>
</tr>

<tr>
<td><code>department_id</code></td>
<td>INTEGER</td>
</tr>
</tbody>
</table>
<code>departments</code> table

<table>
<thead>
<tr>
<th>Column</th>
<th>Type</th>
</tr>
</thead>

<tbody>
<tr>
<td><code>id</code></td>
<td>INTEGER</td>
</tr>

<tr>
<td><code>name</code></td>
<td>VARCHAR</td>
</tr>
</tbody>
</table>
Output:

<table>
<thead>
<tr>
<th>Column</th>
<th>Type</th>
</tr>
</thead>

<tbody>
<tr>
<td><code>salary</code></td>
<td>INTEGER</td>
</tr>
</tbody>
</table>

Let’s say you work at Allstate. Allstate is running <code>N</code> online ads right now. The table <code>ads</code> contains all those ads, ranked by popularity via the <code>id</code> column (e.g., the entry with <code>id = 1</code> is the most popular, etc.).

Create a subquery or common table expression named <code>top_ads</code> containing the top 3 ads (by popularity) and return the number of rows that would result from the following operations

<ol>
<li><code>ads INNER JOIN top_ads</code></li>
<li><code>ads LEFT JOIN top_ads</code></li>
<li><code>ads RIGHT JOIN top_ads</code></li>
<li><code>ads CROSS JOIN top_ads</code></li>
</ol>

Note: Please make the <code>join_type</code> column in your output have the values <code>inner_join</code>, <code>left_join</code>, etc. for each of their respective join types

Note: Please return only one query with each number in a different row

Example:

Input:

<table>
<thead>
<tr>
<th>Column</th>
<th>Type</th>
</tr>
</thead>

<tbody>
<tr>
<td><code>id</code></td>
<td>INTEGER</td>
</tr>

<tr>
<td><code>name</code></td>
<td>VARCHAR</td>
</tr>
</tbody>
</table>
Output:

<table>
<thead>
<tr>
<th>Column</th>
<th>Type</th>
</tr>
</thead>

<tbody>
<tr>
<td><code>join_type</code></td>
<td>VARCHAR</td>
</tr>

<tr>
<td><code>number_of_rows</code></td>
<td>INTEGER</td>
</tr>
</tbody>
</table>

Given a <code>employees</code> and <code>departments</code> table, select the top 3 departments with at least ten employees and rank them according to the percentage of their employees making over 100K in salary.

Example:

Input:

<code>employees</code> table

<table>
<thead>
<tr>
<th>Columns</th>
<th>Type</th>
</tr>
</thead>

<tbody>
<tr>
<td><code>id</code></td>
<td>INTEGER</td>
</tr>

<tr>
<td><code>first_name</code></td>
<td>VARCHAR</td>
</tr>

<tr>
<td><code>last_name</code></td>
<td>VARCHAR</td>
</tr>

<tr>
<td><code>salary</code></td>
<td>INTEGER</td>
</tr>

<tr>
<td><code>department_id</code></td>
<td>INTEGER</td>
</tr>
</tbody>
</table>
<code>departments</code> table

<table>
<thead>
<tr>
<th>Columns</th>
<th>Type</th>
</tr>
</thead>

<tbody>
<tr>
<td><code>id</code></td>
<td>INTEGER</td>
</tr>

<tr>
<td><code>name</code></td>
<td>VARCHAR</td>
</tr>
</tbody>
</table>
Output:

<table>
<thead>
<tr>
<th>Column</th>
<th>Type</th>
</tr>
</thead>

<tbody>
<tr>
<td><code>percentage_over_100k</code></td>
<td>FLOAT</td>
</tr>

<tr>
<td><code>department_name</code></td>
<td>VARCHAR</td>
</tr>

<tr>
<td><code>number_of_employees</code></td>
<td>INTEGER</td>
</tr>
</tbody>
</table>

Tell me about a data project that didn’t go the way you expected. What did you set out to do, what surprised you, and how did you handle it?

Describing a data project and its challenges

Hurdles In Data Projects

Behavioral

Let’s say you work as an LLM engineer at an enterprise software company. You are tasked with designing a retrieval-augmented generation (RAG) system to ensure that the language model only provides answers based on a specific set of ingested documents, and never responds using its original training data.

How would you architect this system?

Let’s say we’re comparing two machine learning algorithms. In which case would you use a bagging algorithm versus a boosting algorithm? 

Give an example of the tradeoffs between the two.

In machine learning, when would you use a bagging algorithm over a boosting algorithm?

In machine learning, when would you use a bagging algorithm over a boosting algorithm?

Bagging vs. Boosting

How would you handle the data preparation for building a machine learning model using imbalanced data?

Addressing imbalanced data in machine learning through carefully prepared techniques.

Data Preparation for Imbalanced Data

How would you explain what a p-value is to someone who is not technical?

What does a p-value in a statistical test represent?

What does a p-value in a statistical test represent?

P-value to a Layman I

In a statistical test, how does a low p-value (less than 0.05) influence our decision about the null hypothesis?

In a statistical test, how does a low p-value (less than 0.05) influence our decision about the null hypothesis?

P-value to a Layman II

How would you build the recommendation algorithm for type-ahead search for Netflix?

Three zebras are chilling in the desert. Suddenly a lion attacks.
Each zebra is sitting on a corner of a triangle with sides of equal length.
Each zebra randomly picks a direction and only runs along the outline of the triangle to either edge of the triangle.

What is the probability that none of the zebras collide?

Three zebras are chilling in the desert. Suddenly a lion attacks.

Each zebra is sitting on a corner of a triangle with side of equal length. Each zebra randomly picks a direction and only runs along the outline of the triangle to either edge of the triangle.

What is the probability that none of the zebras collide?

Three zebras are chilling in the desert. Suddenly a lion attacks.

Each zebra is sitting on a corner of a triangle with side of equal length.  Each zebra randomly picks a direction and only runs along the outline of the triangle to either edge of the triangle.

What is the probability that none of the zebras collide?



Three Zebras

Brainteasers

Given a list, write a function <code>encode</code> which one-hot encodes the items in the list. When one-hot encoding, the most significant bit (index zero) should be represented by the first unique element in the list.

Example 1:

Input:

<pre tabindex="0" class="chroma"><code>elements = [1, 3, 4, 3, 2, 1]
</code></pre>

Output:

<pre tabindex="0" class="chroma"><code>def encode(elements) -&gt; [[1, 0, 0, 0], [0, 1, 0, 0],
 [0, 0, 1, 0], [0, 1, 0, 0], [0, 0, 0, 1], [1, 0, 0, 0]]
</code></pre>

Example 2:

Input:

<pre tabindex="0" class="chroma"><code>elements = [&#39;hi&#39;, &#39;hello&#39;, &#39;amazing&#39;, &#39;amazing&#39;]
</code></pre>

Output:

<pre tabindex="0" class="chroma"><code>def encode(elements) -&gt; [[1, 0, 0], [0, 1, 0], [0, 0, 1], [0, 0, 1]]
</code></pre>

Implement one-hot encoding algorithmically.

One-Hot Encoder

Data Structures & Algorithms

Given a 2-D matrix P of predicted values and actual values, write a function precision_recall to calculate precision and recall metrics.

Return the ordered pair (precision, recall).

Example:

Input:

<pre tabindex="0" class="chroma"><code>
P = [[121, 9],
 [17, 144]]
</code></pre>

where P represents the following table

<table>
<thead>
<tr>
<th></th>
<th>Predicted =TRUE</th>
<th>Predicted =FALSE</th>
</tr>
</thead>

<tbody>
<tr>
<td>Actual =TRUE</td>
<td>121</td>
<td>9</td>
</tr>

<tr>
<td>Actual =FALSE</td>
<td>17</td>
<td>144</td>
</tr>
</tbody>
</table>
Output:

<pre tabindex="0" class="chroma"><code>def precision_recall(P) -&gt; (0.931, 0.877,)
</code></pre>

Write a function to calculate precision and recall metrics.

Precision and Recall

Let’s say you work as a machine learning engineer at a ride-hailing company. You have recently deployed a new ETA prediction model to production.

How would you monitor and evaluate the real-time performance of this model after deployment?

Given an array and a target integer, write a function <code>sum_pair_indices</code> that returns the indices of two integers in the array that add up to the target integer. If not found, just return an empty list.

Note: Can you do it on \(O(n)\) time?

Note: Even though there could be many solutions, only one needs to be returned.

Example 1:

Input:

<pre tabindex="0" class="chroma"><code>array = [1 2 3 4] 
target = 5 
</code></pre>

Output:

<pre tabindex="0" class="chroma"><code>def sum_pair_indices(array, target) -&gt; [0 3] or [1 2]
</code></pre>

Example 2:

Input:

<pre tabindex="0" class="chroma"><code>array = [3]
target = 6 
</code></pre>

Output:

<pre tabindex="0" class="chroma"><code>Do NOT return [0 0] as you can&#39;t use an index twice.
</code></pre>

Given an array and a target integer, write a function that returns the indices of two integers in the array that add up to the target integer.

Target Indices

Let’s say that you are leading the development of an internal chatbot for operations support at Thomson Reuters, an internal tooling in their news division for referencing articles.

How would you design a system given a choice between fine-tuning the LLM or using Retrieval-Augmented Generation (RAG)?

Let’s say we’re designing a deep learning pipeline to process millions of customer support chat logs, and we want to use a Transformer encoder layer built in PyTorch to extract useful representations for downstream tasks like sentiment classification.

How would you approach building this Transformer encoder layer from scratch—what are the essential components, how would you implement them in PyTorch, and why are elements like residual connections and layer normalization critical for stable and effective training?

You work as a data scientist at Cars.com. The company is launching a new personalized car recommendation engine that must support both real-time and batch inference, adapt to changing user preferences, and enable frequent model updates via CI/CD pipelines.

How would you design the end-to-end architecture for this system, ensuring it can robustly detect and respond to data distribution shifts while maintaining seamless CI/CD for machine learning models?

Design an end-to-end ML system for personalized car recommendations at Cars.com, supporting real-time and batch inference, data distribution shift detection, and CI/CD for frequent model updates.

Car Recommendation Architecture

What are the assumptions of linear regression?

Which assumption of the residuals of the standard linear regression model can not be overcome by increasing the sample size?

Regression assumptions

Given a <code>users</code> table, write a query to return only its duplicate rows.

Example:

Input:

<code>users</code> table

<table>
<thead>
<tr>
<th>Column</th>
<th>Type</th>
</tr>
</thead>

<tbody>
<tr>
<td><code>id</code></td>
<td>INTEGER</td>
</tr>

<tr>
<td><code>name</code></td>
<td>VARCHAR</td>
</tr>

<tr>
<td><code>created_at</code></td>
<td>DATETIME</td>
</tr>
</tbody>
</table>

In the context of hypothesis testing, what are type I errors (type one errors) and type II errors (type two errors)? What is the difference between the two?

Bonus: Describe the probability of making each type of error mathematically.

You are given a dictionary with two keys <code>a</code> and <code>b</code> that hold integers as their values.

Without declaring any other variable, swap the value of <code>a</code> with the value of <code>b</code> and vice versa.

Note: Return the dictionary after editing it.

Example:

Input:

<pre tabindex="0" class="chroma"><code>numbers = {
 &#39;a&#39;:3,
 &#39;b&#39;:4
}
</code></pre>

Output:

<pre tabindex="0" class="chroma"><code>def swap_values(numbers) -&gt; {&#39;a&#39;:4,&#39;b&#39;:3}
</code></pre>

Swap the values of `a` and `b` in the dictionary without declaring any other variable

Swap Variables

Let’s say we’ve built a churn prediction model for a subscription service using user behavior data from the past 12 months, and now we want to deploy it to production and also apply it to a new user segment that wasn’t in our original training data.

How would you approach preparing the model for deployment and ensuring it generalizes well to the new dataset?

Let’s say that you’re training a classification model.

How would you combat overfitting when building tree-based models?

Let's say that you're training a classification model.   How would you combat overfitting when building tree-based models?

Overfit Avoidance

You were asked to forecast sales for a retail chain with strong seasonal patterns using seasonal ARIMA (SARIMA).

How would you justify using SARIMA in this context, and how does it compare to other time series forecasting techniques?

Compare SARIMA to other time series models for seasonal sales forecasting.

SARIMA in Retail Forecasting

Forecasting & Time Series

You work as the sole machine learning engineer at a mid-sized e-commerce company. You have developed a predictive model for customer purchase likelihood that needs to be retrained periodically and reliably served in production. There is no dedicated data engineering or ML-Ops support.

How would you automate the retraining and deployment of this model while ensuring reliability and minimal downtime?

How would you tackle multicollinearity in multiple linear regression?

Imagine you work at a major credit card company and are given a dataset of 600,000 credit card transactions to build a fraud detection model.

How would you approach this task? Write your answer in the comments.

How would you explain the concept of linear regression to three different audiences: a child, a first-year college student, and a seasoned mathematician? Ensure your explanations are tailored to the understanding level of each audience.

Calculate Moving Average	SQL	Easy
Predict Customer Churn	Machine Learning	Medium
A/B Test Significance	Statistics	Medium
Optimize Query Performance	SQL	Hard
Feature Importance Analysis	Machine Learning	Medium
Clean Missing Data	Python	Easy
Neural Network Architecture	Deep Learning	Hard
Calculate Cohort Retention	SQL	Medium
Bayesian Probability	Statistics	Easy
Recommend Similar Products	Machine Learning	Hard

Oliver Wyman Interview Questions

Oliver Wyman Interview Guides

Oliver Wyman Interview Questions

Challenge

Oliver Wyman Salaries by Position

Discussion & Interview Experiences

Discussion & Interview Experiences