Skip to content
Hari KharelStudy

Trading firms

All three firms say officially that the path matters more than the finish: one problem that grows under new constraints, solved in real code in your best language, while you think aloud and take hints well. Jane Street is the most collaborative and implementation-heavy. Citadel Securities runs one problem through design, code, review and extension. HRT adds systems depth (memory, OS, C++) and short, high-bar screens.

What they share

Section titled “What they share”
Jane Street Citadel Securities HRT
Problem shape Open-ended, grows in stages One problem, four stages Varied; short, exact screens
Language Any real language. “Please don’t use OCaml just because you think it will make us happy” official Mostly any. Some roles require C++, flagged in advance official C++ or Python, depending on team official
Systems depth Awareness of real-machine effects such as caching official Deep for C++ teams Deep: memory, OS, networking reported
Behavioral No separate round; a project deep-dive “Why Citadel”, project discussion A project or behavioral round
AI No published policy; final rounds are in person No published policy Reportedly not allowed unless instructed thin

Recurring problem shapes across all three reported:

Jane Street

Section titled “Jane Street”

Loop. A technical interview over Zoom, then final rounds in person, mostly collaborating on a series of coding problems in CoderPad. official Reported for mid/senior: a 1-hour screen with one two-part problem, then about 5 hours onsite (three 75-min coding or design rounds plus a project deep-dive). Loops can end early if the first rounds go badly. reported

What it tests. Problems with several reasonable approaches, not “aha” puzzles. No mental math or probability for SWE. official The published example grows in stages: memoise a function, bound the cache with O(1) FIFO eviction, then move to LRU, improving the data structure as you go. official Reported families are implementation-heavy: game simulations (Connect-K, Tetris-like), tree structures (Merkle-tree style), domain classes, and memoisation. One senior candidate said nearly every question was “best answered in an object-oriented way.” reported

How you’re judged. “Are they someone we want to work with?” Courtesy, clarity, saying “I don’t know”, language fluency, fundamentals. “We aren’t looking for you to finish.” A clean but slower candidate can beat a fast, messy one. official Prep: watch the mock interview. If you recognise a question, say so.

Citadel Securities

Section titled “Citadel Securities”

Loop (experienced hires, official). Recruiter, then a technical coding screen, then a 4-5 hour onsite: one problem in four stages with the same engineer: design, programming, code review, approach. “You define the problem, design a system, build it and then defend and extend it.” An early design choice comes back in the code review. official A 2-problem HackerRank assessment (about 70 min) is often reported first. reported

What it tests. The stated criteria: system design under ambiguity, code quality, judgment on edge cases and failure modes, stating assumptions, separation of concerns, and explaining trade-offs as you build. official Reported 2026 families:

  • producer-consumer, progressing from a mutex to a spinlock to lock-free
  • an order book (add, cancel, aggregate price levels)
  • a conflating queue (keep the latest price per symbol)
  • currency arbitrage as a graph
  • external merge sort with a heap
  • design: a metrics and alerting system

For C++ teams, expect move semantics, memory ordering and std-container internals. reported

Failure modes. Rushing to finish (the firm warns this does worse), and not writing down assumptions when requirements shift. reported

Hudson River Trading

Section titled “Hudson River Trading”

Loop. Official, but from 2021 (“we are constantly iterating”): a timed take-home (HackerRank/Codility, language reference allowed), about two phone interviews (a systems/data-structures discussion plus a programming round), then a full-day onsite. official Reported for 2025-26: the online assessment is “the biggest cut”, then 1-2 CoderPad screens, then an onsite with harder coding, an architecture discussion and a project round. reported

What it tests. Idiomatic, readable, well-encapsulated, resource-efficient code. Systems knowledge (memory, I/O, processes). Methodical problem solving and being teachable. official Reports describe a broad spread: floating point, bit manipulation, union-find, virtual memory, networking, C++ pointers and references. Screens are short and exact: one candidate was rejected after a single unconfirmed assumption in a 30-min round. reported

Their own advice: “Test your code! The sample test cases provided aren’t comprehensive.” Over-communicate when you think, backtrack or need a hint. Hints and mistakes are OK. official

Your plan

Section titled “Your plan”

Use the Sandbox filters for Jane Street, Citadel and Hudson River Trading.

2 weeks 4 weeks
Week 1 Function Result Cache, Connect K, Tiny Stack Machine, Merging Sorted Feeds, Conflating Quote Queue. Narrate every one aloud The 2-week Week 1, plus Ring Buffer, Hash Map from Scratch and Streaming Median
Week 2 Order Matching Engine, Currency Arbitrage, Blocking Queue Simulator, then Exchange Simulator as a timed mock. HRT: revise memory layout, virtual memory and processes. Prepare one project deep-dive Order Matching Engine, Consolidated Quote Book, Currency Arbitrage, Event Loop Timers
Week 3 Blocking Queue Simulator, Merkle Sync, Range Queries, Trade Bars. Systems revision (C++ teams: move semantics, memory ordering)
Week 4 Exchange Simulator timed mock, one four-stage mock (design, build, self-review, extend), deep-dive rehearsal

For Citadel’s format, after each Sandbox problem spend 10 minutes reviewing your own code out loud: what would you change, and which early decision limited you?

Sources

Section titled “Sources”