---
title: "Applications"
book: "The Interview Book"
subject: quant
language: en
chapter: 3
exercises: 0
source: https://one-course.com/books/quant/18/en/chapter/3-applications
---

# Chapter 3 — Applications

Four hundred CVs arrive for twenty first-round slots, and the reader has one afternoon: thirty-six seconds a CV. Whatever is not on the first half of the first page is, for practical purposes, not there. A line that says what was built and what number it moved is read; a line that lists technologies is skimmed; a line that claims a Sharpe ratio of 3 on a student project is read twice, for the wrong reason. An application is the first measurement in the funnel of [Chapter 1](https://one-course.com/books/quant/18/en/chapter/1-what-each-interview-tests-by-role#ch-iv-what-each-interview-tests-by-role), and the cheapest one to improve.

## 3.1 The one-page CV

**Definition 3.1 (CV screen).**

The *CV screen* is the stage at which a person or a program reads an application and decides whether it goes on to an assessment or an interview. Its inputs are the CV, any form answers and, where the firm has them, a referral and the candidate’s earlier applications.

The reader of a screen has a short list of questions: is the candidate eligible (degree, date, right to work), is there evidence of the role’s core abilities, and is there anything exceptional. The first half-page has to answer all three.

**Method 3.2 (Writing a CV line).**

1. Start with a verb for what you did (built, measured, proved, reduced), not with the tools.
2. Name the object: a pricer, a feed parser, a study of an effect.
3. Give a number with its baseline: “from 40 to 6 milliseconds”, “out-of-sample $R^2$ of 0.8% against 0.1% for the benchmark”, “1 200 tests”.
4. Name the tools last, and only those you can be interviewed on for thirty minutes.
5. Cut every line that fails steps 1 to 3. One page; the most relevant evidence first.

**Example 3.3 (One line, before and after).**

Before: “Used Python, pandas and scikit-learn to build machine-learning models for stock prediction.” After: “Tested whether overnight order imbalance predicts next-day returns in 500 US stocks, 2015–2024; walk-forward $R^2$ 0.3% ($t = 2.1$), gone after costs; Python.” The second line tells a research interviewer what to ask about, and its honest last clause is worth more than the result: it shows the candidate knows what costs and multiple testing do ([Chapter 14](https://one-course.com/books/quant/18/en/chapter/14-statistics#ch-iv-statistics)).

A CV is also the first list of interview questions. Every line invites “tell me about this”, and the answer should be ready in two lengths: ten seconds, and five minutes with the numbers and what went wrong.

## 3.2 Internships, projects and competitions

Evidence of the role’s abilities comes in three forms, each read differently. An internship at a trading firm or bank is evidence that another firm’s screen was passed and that the candidate worked in the environment; its value is in what was delivered, which the candidate must be able to describe without disclosing the firm’s confidential information. A personal project is evidence of initiative, and its value is in the honesty of its measurement: a baseline, an out-of-sample test, a statement of what failed. A competition (mathematics olympiads, programming contests, data-science and trading competitions; the pipelines are in One Quant Book 17, chapter 29) is evidence of ability under time pressure, read through its ranking.

Rankings in data competitions need care, because a public leaderboard scored on a small sample rewards noise.

**Proposition 3.4 (The best of kkk equal scores).**

If $k$ teams of equal true skill are scored with independent noise of standard deviation $\sigma$, the expected best score exceeds the common true score by $\sigma\,m_k$, where $m_k = \E[\max_{i\le k} Z_i]$ for independent standard normal $Z_i$: $m_{10} \approx 1.54$, $m_{60} \approx 2.32$, $m_{100} \approx 2.51$, $m_{400} \approx 2.97$. The asymptotic $\sqrt{2\ln k}$ overstates at these sizes ($3.46$ for $k = 400$).

**Proof.** The maximum has density $k\,\varphi(x)\,\Phi(x)^{k-1}$; the values are its mean by numerical integration, confirmed by simulation in the chapter’s tests. ∎

The same proposition explains why a strong result from many attempts is weaker evidence than it looks: the best of sixty backtests of worthless variants over five years has an expected annualised Sharpe ratio of about $2.32/\sqrt{5} \approx 1.04$ (One Quant Book 4, chapter 12, on multiple testing and the deflated Sharpe ratio). A candidate who reports how many variants were tried is believed more, not less.

## 3.3 Referrals and recruiters

**Definition 3.5 (Employee referral, agency recruiter).**

An *employee referral* is an application submitted or endorsed by a current employee of the firm, who vouches that the candidate is worth screening. An *agency recruiter* is an intermediary retained by the hiring firm, usually paid a fee on a successful hire, who finds candidates and manages their process on the firm’s behalf.

A referral changes the screen’s prior. In a study of nine large firms in three industries, referred applicants were more likely to be hired and to accept offers, with similar skills on paper, and referred workers were substantially less likely to quit (Burks, Cowgill, Hoffman and Housman, 2015). The mechanism is information: the referrer knows the candidate and the firm. A referral from someone who has not worked with the candidate carries little of it, and a good referrer says so.

A recruiter is paid by the firm, so the candidate is the product being placed, not the client. That does not make the recruiter an adversary: a placement requires the candidate to accept, and a recruiter who places people badly loses clients. It does mean that the recruiter’s incentives are known and can be computed ([Interview question 3.5](#iq-iv-applications-5)). In the United Kingdom an employment agency may not charge a work-seeker a fee for finding them work, except in a few occupations listed in regulations ([Box 3.1](#dat-iv-applications-law)); a recruiter who asks a candidate for money is not a recruiter.

## 3.4 Automated screening

Large firms receive more applications than people can read, and some screens are partly automated: a program ranks CVs, or scores a recorded video interview, or filters by answers to form questions. Two consequences follow. The CV must be readable by a program as well as by a person (plain text, standard section names, no information only in images). And the rules are changing: automated screening tools are regulated in some places, with audits and notices to candidates.

**As of September 2026 — Rules on automated screening.**

**New York City**: Local Law 144 of 2021 prohibits employers and employment agencies from using an automated employment decision tool unless it has had a bias audit within one year of its use, information about the audit is publicly available, and the required notices have been given to candidates; the city began enforcing it on 5 July 2023. **European Union**: the AI Act (Regulation (EU) 2024/1689), Annex III, point 4(a), classifies as high-risk “AI systems intended to be used for the recruitment or selection of natural persons, in particular to place targeted job advertisements, to analyse and filter job applications, and to evaluate candidates”; the Digital Omnibus on AI (Regulation (EU) 2026/1744, in force since 27 July 2026) moved the application of the high-risk requirements to such systems from 2 August 2026 to 2 December 2027. **United Kingdom**: section 6(1)(a) of the Employment Agencies Act 1973 prohibits an employment agency from charging a fee to a person for finding them employment, with exceptions prescribed by the Conduct of Employment Agencies and Employment Businesses Regulations 2003 (regulation 26 and Schedule 3, mostly entertainment and modelling).

## 3.5 Worked answers

**Example 3.6 (A project line that survives the first follow-up).**

*“Your CV says: ‘Developed a profitable mean-reversion strategy on ETFs.’ Talk me through it.”* The line invites the question that sinks it: profitable after what? The candidate who has done the arithmetic before the interview says: “Fifty sector and country ETFs, 2010 to 2020, daily data. Before costs it had a Sharpe ratio of 0.9 at 10% volatility, so about 9% a year. It traded its book about a hundred times a year; at 5 basis points a trade that is 5% a year, which leaves 4% and a Sharpe ratio of 0.4. So: a real effect, most of it eaten by turnover, and my next step was to trade less often.” The rewritten line follows from the answer: “Backtested a daily mean-reversion strategy on 50 ETFs (2010–2020): Sharpe ratio 0.9 before costs, 0.4 after 5 bp costs; halving turnover kept 0.6.” The last clause must be true and tested before it is written; here the candidate’s rerun at half the turnover had a gross return of 8.5% and costs of 2.5%, which is 6% at 10% volatility, and the line says exactly that.

**Example 3.7 (Stating a ranking so that it can be checked).**

*“You write that you were ‘top 1% in a forecasting competition’. What exactly was it?”* Rankings invite two checks: the denominator and the leaderboard. “Twelfth of 2 000 teams on the private leaderboard, so the top 0.6%; on the public leaderboard I was fifth, and I dropped because my last submissions were tuned to the public split” is a complete answer: it gives the number, which board, and the lesson about overfitting to a validation set that [Chapter 19](https://one-course.com/books/quant/18/en/chapter/19-machine-learning#ch-iv-machine-learning) will test again. “Top 1%” alone reads as rounding in the candidate’s favour, and interviewers ask. [Figure 3.1](#fig-iv-applications-bestof) shows how far the best of many equally good entries is expected to sit above its true level.

![Expected best of k equally skilled entries whose scores carry independent standard normal noise, in noise standard deviations above the common true score (solid), and the approximation √2 k (dashed), which overstates it at every size shown. With 400 entries the leader is expected to be 2.97 standard deviations above the truth, which is why public leaderboards and best-of-many backtests shrink when retested. Data: fig_iv_bestof.py.](https://one-course.com/images/onecourse/chapters/quant-18/iv-applications/fig-2c1ee71bb633.svg)

***Figure 3.1.** Expected best of $k$ equally skilled entries whose scores carry independent standard normal noise, in noise standard deviations above the common true score (solid), and the approximation $\sqrt{2\ln k}$ (dashed), which overstates it at every size shown. With 400 entries the leader is expected to be 2.97 standard deviations above the truth, which is why public leaderboards and best-of-many backtests shrink when retested. Data: `fig_iv_bestof.py`.*

**Example 3.8 (An unusual path).**

*“You spent two years as a secondary-school teacher between your degree and your master’s. Why?”* The interviewer is checking that the CV’s story is coherent and that the candidate owns it. A scoring answer is short, true and connects forward: “I wanted to know whether I could explain mathematics to people who did not want to hear it. I could, mostly, and I learned to check what someone understood rather than what I had said. I missed working on hard problems myself, which is why I went back for the master’s.” It avoids apology and does not over-explain; the follow-up, if there is one, is about the master’s.

## 3.6 Question bank

**Interview question 3.1 ★ researcher, mle • any.**

Rewrite this CV line so that a research interviewer would want to ask about it: “Used Python and scikit-learn to build machine-learning models for cryptocurrency price prediction.”

**Solution of Interview question 3.1.**

Say what was tested, on what, with what result against what baseline, and what failed: “Tested whether one-hour order-flow imbalance on two exchanges predicts the next hour’s return in 20 crypto pairs, 2022–2025; walk-forward hit rate 51.2% against 50.0% (t = 1.4), not significant after costs; Python.” The rewrite gives the interviewer three things to ask about (the data, the validation, the costs) and shows the candidate reports a null result honestly.

*What the interviewer is looking for: a verb, an object, a number with its baseline, and honesty about the outcome.*

**Interview question 3.2 ★ trader • market maker.**

“Walk me through your CV in one minute.” What do you say, in what order, and what do you leave out?

**Solution of Interview question 3.2.**

Four sentences, in the order of relevance to the role: what you are now (degree, year, subject); the one piece of evidence closest to trading (an internship, a competition, a project with numbers); what that taught you about decisions under uncertainty; why this role now. Leave out the chronology, every item the interviewer can read, and anything you cannot discuss for five minutes. Stop at sixty seconds and let the interviewer choose what to open.

*What the interviewer is looking for: selection and order by relevance, brevity, and a clear reason for the application.*

**Interview question 3.3 ★ developer • proprietary firm.**

Your CV lists C++, Rust, Python, Java, Go, Haskell and CUDA. The interviewer says: “Let’s do this one in Haskell.” You have written two hundred lines of Haskell in your life. What should the CV have said, and what do you say now?

**Solution of Interview question 3.3.**

The CV should have listed only languages the candidate can be interviewed in for thirty minutes, with the others under “also used” or not at all. Now: say plainly that Haskell was a short course project, that the strong languages are C++ and Python, and offer to write the solution in either while explaining how it would look in a functional style (recursion instead of loops, immutable data, a fold). Pretending costs more than the admission; the interviewer chose Haskell to find out.

*What the interviewer is looking for: honest calibration of one’s own skills, and a constructive alternative.*

**Interview question 3.4 ★★ researcher, mle • systematic fund.**

You finished third of 400 teams on a data competition’s public leaderboard and fortieth on the private one. If every team had the same true skill and public scores carried independent noise with standard deviation 0.01, by how much would you expect the best public score to exceed the true score? What do you say when an interviewer asks about the drop?

**Solution of Interview question 3.4.**

With 400 equal teams and noise 0.01, the expected best public score exceeds the true score by $0.01 \times m_{400} \approx 0.01 \times 2.97 \approx 0.030$ ([Proposition 3.4](#prop-iv-applications-max)). A drop from third to fortieth is what noise alone produces when the public sample is small; the private leaderboard, scored on more data, is the better measurement. In the interview: say so, say what you did to avoid fitting the public board (cross-validation that mimicked the private split, a limit on submissions used for selection), and what you would do differently.

*What the interviewer is looking for: the expected maximum of noisy scores, and a mature reading of one’s own result.*

**Interview question 3.5 ★★ trader, researcher • any.**

Suppose a recruiter is paid 20% of your first-year base salary if you accept, and your offer’s base is 150 000. What is the fee? You ask for 160 000, and the recruiter judges that asking puts a 10% chance on the placement failing. How does the recruiter’s expected fee change, and what does that tell you about the advice you will get?

**Solution of Interview question 3.5.**

The fee is $0.20 \times 150\,000 = 30\,000$. At 160 000 it would be 32 000, a gain of 2 000, but with a 90% chance of closing the expected fee is $0.9 \times 32\,000 = 28\,800$, which is 1 200 below the certain 30 000. On these numbers the recruiter prefers that you do not ask. The advice is not dishonest for that; it is predictable. Ask the recruiter for facts (the band, what the firm has paid similar hires), not for a recommendation, and negotiate the number yourself ([Chapter 7](https://one-course.com/books/quant/18/en/chapter/7-offers-compensation-and-non-competes#ch-iv-offers-compensation-and-non-competes)).

*What the interviewer is looking for: an expected-value reading of an intermediary’s incentives.*

**Interview question 3.6 ★★ developer, mle • bank.**

A large bank tells you that its first screen is partly automated. What can you do about it, what should you not do, and what rights do candidates have where you live?

**Solution of Interview question 3.6.**

Make the CV machine-readable: plain text in a standard layout, conventional section names, the role’s vocabulary where it is true, dates in one format, no content only in images or tables. Answer form questions precisely. Do not stuff keywords or hide text: automated screens are followed by people, who reject it. On rights: in New York City such tools must have had a bias audit and candidates must be notified; in the EU recruitment tools are high-risk systems under the AI Act, with requirements applying from December 2027 ([Box 3.1](#dat-iv-applications-law)); elsewhere, ask the firm what the tool does and whether an alternative process exists, which some firms offer as an accommodation.

*What the interviewer is looking for: practical legibility, no gaming, and awareness that candidates have rights that vary by place.*

**Interview question 3.7 ★★★ researcher • systematic fund.**

Your CV says a strategy you designed had a backtested Sharpe ratio of 2.4 over five years. In the interview you admit you tried about sixty variants. What would the best of sixty worthless variants be expected to show, and how do you defend, or retract, the line?

**Solution of Interview question 3.7.**

Over five years an annualised Sharpe ratio estimate has standard deviation about $1/\sqrt5 \approx 0.447$, so the best of sixty worthless variants is expected near $2.32 \times 0.447 \approx 1.04$. The reported 2.4 is about $(2.4 - 1.04)/0.447 \approx 3.0$ standard deviations of one estimate above that, which is suggestive but not conclusive, since the variants are correlated and the costs may be understated. Defend it by saying what the sixty variants were, how the final one was chosen, what the out-of-sample or later-period result was, and the deflated Sharpe ratio (One Quant Book 4, chapter 12). If none of that exists, retract the number and keep the line as a description of the study.

*What the interviewer is looking for: the multiple-testing correction, done in the head, and intellectual honesty about one’s own CV.*

**Interview question 3.8 ★★★ trader, researcher, developer • any.**

A former classmate works at the firm you most want and offers to refer you. What do you ask her for, what do you not ask her for, and how does a referral change what the firm does with your application?

**Solution of Interview question 3.8.**

Ask her what the team and the role are like, what the firm values that its website does not say, and whether she knows your work well enough to refer you; if she does, ask her to submit or endorse the application and to say what she actually knows about you. Do not ask for interview questions: it puts her in breach of her employer’s trust and the questions will change anyway. A referral usually moves the application past the first screen or to a person’s desk, and it raises the prior the screen starts from; it does not change the later stages, which are the same for everyone (Burks et al., 2015, on why firms value referrals).

*What the interviewer is looking for: using a referral for information and endorsement without compromising the referrer.*

Sources and further reading

- S. V. Burks, B. Cowgill, M. Hoffman and M. Housman, “The value of hiring through employee referrals”, *Quarterly Journal of Economics* 130(2), 2015, 805–839.
- New York City Local Law 144 of 2021 and the Department of Consumer and Worker Protection’s page on automated employment decision tools.
- Regulation (EU) 2024/1689 (the AI Act), Annex III; Regulation (EU) 2026/1744 (Digital Omnibus on AI), Official Journal L, 24 July 2026.
- Employment Agencies Act 1973, section 6; the Conduct of Employment Agencies and Employment Businesses Regulations 2003, SI 2003/3319, regulation 26.
- One Quant Book 4, chapter 12 (multiple testing); One Quant Book 17, chapters 28–29 (entry routes, education pipelines).