Citadel is very different from big manufacturers. Don’t ask about the original LeetCode questions, and don’t take the stereotyped essay test. The whole process is scenario-driven and focuses on the trading system. The interviewer is an engineer who writes a trading system. He will ask in-depth questions until you don't know how to do it. I successfully passed OA and three rounds of Onsite. The following is a detailed interview experience, including Citadel interview questions, my answer ideas, interviewer feedback and preparation suggestions.
Round 1: OA + Tech Screen
OA (90 minutes) Two algorithms + one system design, time is very tight.
Algorithm 1: Sliding window extreme value
Real-time price stream, fixed window size, sliding one frame at a time, requires O(n) to return max and min within the window.
Solution: Two monotonic queues (double-ended queues) maintain maximum and minimum candidates respectively.
- When joining the queue, elements that do not satisfy monotonicity pop up;
- When dequeuing, determine whether the leader of the team slides out of the window. Focus on window initialization and queue empty boundaries. After writing the code, I added a time/space complexity explanation (why it is O(n) instead of O(nk)). The interviewer paid great attention to the coding style and comment habits.
Algorithm 2: Concurrency-safe order book
Supports multi-threaded insertion/deletion of orders, price inquiry, and matching. The read-write ratio is about 7:3, ensuring consistency under high throughput.
Plan:
- Add read-write locks (std::shared_mutex) segmented by price range to reduce lock competition;
- When matching across price levels, the locks are locked in a fixed order to prevent deadlock;
- Use lock-free queues as write buffers, and background threads consume in batches to smooth latency. The reasons for not using global locks (which will become a bottleneck) and being completely lock-free (high complexity and error-prone) are explained. Atomic operations + memory order (release/acquire) are used.
System design: low-latency logging system
The trading system generates massive logs and cannot block the main path. It also needs to support quick query of time range/keywords.
Tiered storage solution:
- Hot data: Memory Ring Buffer + pre-allocated memory pool, writing O(1);
- Temperature data: mmap file, using OS page cache, asynchronous batch flushing, zero copy;
- Cold data: LSM-tree compressed storage, suitable for query; additionally mentions sendfile zero copy. Explain the reasons why traditional relational databases are not used (B+ tree random IO latency is high).
Tech Screen (45 minutes)
Self-introduction project: Distributed message queue (Go) developed by the school. Core points: Topic sharding + Raft consensus, sendfile zero copy, token bucket back pressure, WAL persistence.
Ask:
- How to double throughput and scale horizontally? → Consistent hashing + dynamic rebalancing, double writing during transition + idempotent deduplication.
- How to prevent node downtime without loss or damage? → WAL guarantees no loss, message unique ID + Bloom filter + Redis consumer deduplication, achieving Exactly-Once effect.
Special session: Distributed locks explained in non-technical language.
Simile: Just like the shared power bank cabinet at the airport, each person scans the code to get a unique cabinet door number to prevent conflicts. The interviewer agreed that this expression is important when communicating with Trader/Quant.
Small probability question: With a strategy winning rate of 55%, a profit-loss ratio of 1:1, and 100 trades, what is the probability of making at least 60 profits? Using the binomial distribution + normal approximation (μ=55, σ≈4.97), we get about 15%-18%. It emphasizes that real trading pays more attention to risk control and survivability.
Round 2: Onsite (60 minutes) - full market scenario questions
Market maker scenario The current price of the stock is US$100. Suddenly a large sell order hits the market. How to adjust the quotation?
Answer frame:
- Normal quotation: theoretical middle price pending order, dynamic spread + volume, inventory neutralization;
- Order flow monitoring: widen spread and reduce volume when buy/sell order arrival rates are out of balance;
- Extreme responses: immediately cancel buy orders, trigger circuit breakers, limit positions, and become more conservative after volatility rises.
Fermi estimate:How many cups of coffee are consumed in Manhattan every day? Assuming a total population of 3 million during the day, the average daily consumption is estimated by group of people (white-collar workers, service industry, tourists), and the final result is about 4.2 million cups. Emphasis on assuming transparency and logical chain.
Underlying implementation: low-latency order book
- Data structure: std::map (red-black tree) maintains the price hierarchy, and each price corresponds to the FIFO order queue;
- Optimization: lock-free skip table, memory pool, CPU affinity, cache line alignment (alignas(64) anti-counterfeit sharing).
Hand Tear: Time Wheel Timer(Order timeout and cancellation) Hierarchical time wheel (milliseconds→seconds→minutes), add/deletion O(1), slot-level spin lock + object pool.
Round 3: Final Round (60 minutes) - Tech Lead
Tech Lead side, half technical and half behavioral side.
Why Citadel
Citadel Securities' engineering team is small but has high output per person, and the engineer's code directly affects the profit and loss of the day. Modules written by big manufacturers may only be part of a large system, far from the final result. It's close to the problem domain, and it's clear why you do what you do.
Behavioral
When I was doing a backtesting framework during my internship, the quant researcher thought that as long as the functions were correct, I didn’t care about the performance. I ran profiling to locate hotspot functions and used data to communicate: after optimization, the backtesting time was greatly shortened and the strategy iteration was faster. After the other party agreed, the core computing module was rewritten using SIMD instructions. The backtest speed increased by 40% and the numerical results were consistent. Subsequently, other researchers also used this code. When speaking, structure: context, conflict, action, result, highlighting data-driven communication.
Distributed trading system design
Demand: Global cross-time zone trading, orders are accessed from New York, London, and Tokyo and routed to the corresponding exchange. Millisecond-level latency and 99.999% availability.
Architecture:
- Multi-active data centers: Complete services are deployed in New York, London, Tokyo, and Singapore, and can be accessed nearby
- DNS intelligent resolution plus Anycast routing automatically leads to the nearest computer room
- Use dedicated lines or optimal network paths across regions to avoid public network delays
- Inside each center: the access layer performs front-end risk control, the matching engine is completed in memory, and the trading gateway is connected to the exchange.
Consistency:
- Global data such as account balances and positions are replicated using multi-master replication
- Cross-region conflicts are resolved using version vectors, written with version numbers, and conflicts are merged according to rules.
- Most scenarios accept eventual consistency, and transient inconsistencies are allowed within the delay below the threshold.
- The transaction execution path pursues strong consistency and uses distributed transactions or Saga mode.
Availability:
- Redundancy at each level, automatic standby machine failure
- Heartbeat detection, fault detection within seconds
- The critical path is stateless or the state can be quickly rebuilt
- Circuit breaker downgrade: When an exchange fails to connect, cache orders and try again later, without affecting other exchanges.
Asking about the cross-region delay and consistency tradeoff: The trading system prioritizes low latency and availability. In extreme cases of consistency, it can be temporarily sacrificed but eventually converges. If multiple orders from the same user are routed to different regions, inconsistencies may be seen in the short term, but the execution results are subject to confirmation by the exchange and will eventually be synchronized to the global status.
Rhetorical questions
Ask graduating students for growth advice. The interviewer said that if you spend less time learning the framework, the framework will become outdated. Have a solid understanding of operating systems, networks, and compilation principles, and understand how computers work. The core ability of the trading system is to make quick decisions and weigh risks under uncertain conditions. It cannot be learned by reading a book, but can be learned through real problems.
From unclear direction to winning Citadel Offer
What I want to say the most about successfully passing the Citadel interview this time is: it’s really difficult to do it alone. Later I found Programhelp, and their seniors provided me with full OA practical assistance and in-depth interview coaching, including real question prediction, code optimization, system design deduction and high-intensity mock interviews, which made me significantly more confident and organized in real interviews.
If you are also preparing for high-intensity interviews at Citadel, Jane Street, NVIDIA, Point72, etc. And feel that reviewing alone is inefficient, I strongly recommend you to learn about Programhelp. They focus on providing OA practical assistance and interview guidance. Seniors will communicate with you directly and develop targeted plans based on your situation.
Students in need can contact them directly Programhelp Discuss in detail.
