HFT Cockpit

Why it works: PE is the hinge, not a résumé item — "look through the storytelling to put a number on real value" is a market maker's job description. You're not switching careers, you're speeding one up. The drone line shows trader reasoning instead of claiming it.

45-second cut: drop the assumptions sentence, end on "closed to seconds."

The first draft ran ~200 words against an 80-word target because each revision was appended rather than swapped in. Revise by deletion.

The only test that matters: if you could say it to Optiver, IMC or Jane Street, it isn't an answer. The raw ingredients:

1. Flat structure. 11–50 people; trader, developer and operator all report to the Managing Director. Flipped to the work's side: when a model prices something wrong, you walk to the person who wrote it and have it fixed that afternoon rather than filing a ticket.
2. Breadth. Futures, equities, ETFs, FX and crypto from one seat. Flipped: a trader who sees four asset classes together spots relationships a siloed one can't — which is literally where stat arb and basis trades come from at a firm doing exactly that. And it continues a pattern you've already demonstrated: you deliberately sampled PE, M&A and fund analysis because you wanted to understand the whole machine before committing.
3. Founder pedigree. Sjak Kuipers and Reymond Tse — Optiver → Flow Traders → their own firm. Market making learned from the Amsterdam lineage at source.

The honest version is the better story. Drafted, three weeks from zero, placed third — that's learning velocity, and "quick with new complex concepts" is written into their JD. Years of practice would prove less. The closing line is deliberate: it's disarming and it permanently closes the follow-up risk.

Where it maps: in the McDonald's round you kept buying on your own estimate while being sold to six times. Same failure mode — acting on your own assumption instead of reading what the other side's behaviour was telling you. If a "greatest weakness" question comes, that's your answer, and it's strong because you can name the fix.

Why this holds up: R is a real analytical language (data frames, regression, vectorised ops), VBA is real programming (loops, functions, conditionals), Power Query is ETL. You have every concept Python would ask of you — the gap is syntax, which is days not months. Credible because true, and it ends forward-looking rather than on the deficit.

Excel specifics to have ready: INDEX/MATCH and why it beats VLOOKUP (the lookup column's position can shift and silently break it) · SUMIFS / COUNTIFS · pivot tables · Power Query for anything that should refresh rather than be rebuilt.

40 minutes on Day 3 — highest remaining ROI. Converts "I did a module" into "I've been working through pandas this week," which is true and shows initiative. Be able to write these six from memory:

1. A list comprehension — [x * 2 for x in nums if x > 0]
2. A dict comprehension
3. df = pd.read_csv('file.csv')
4. A moving average — df['ma'] = df['price'].rolling(20).mean()
5. A loop computing running P&L
6. Filter, groupby, sort_values on a DataFrame

You're not trying to pass as an engineer — you're trying not to freeze if someone says "write me something simple."

Lead with the work one if they ask for a professional example, or if the room feels formal. Lead with Inner Mongolia if you want to be the answer they remember at the end of a long day. Both are 60 seconds.

A · The LP call — the sharper trading parallel

Why it lands: two possible states of the world, no way to resolve which inside the call — so you found the action that was correct under both. Not a compromise, a hedge. When you can't resolve the uncertainty in the time available, stop picking the right state and find the action robust across both.

B · Inner Mongolia — the memorable one

The U-turn is the best part — it's the test that separates coincidence from intent. Sharing a route for 30km is ambiguous; following through a reversal is not. That's how you confirm a signal.

The close is the trading answer: asymmetric payoff — a cheap hedge against an unbounded tail. You don't need certainty to justify a cheap precaution. Held in reserve: the local newspaper placed under the windshield before the trip, i.e. hedging in advance rather than only reacting.

Rebuild the estimate — don't retreat

When flow keeps hitting you, go back into the model and find the broken assumption. Retreating without re-estimating leaves you nowhere to stop — that's how a quote ends up below fair value.

McDonald's: retreated blindly, capitulated under fair. HDB: found the 200-households-per-block error, re-derived, landed within 10%. Same person, one round apart.

Anchored vs unanchored

Anchored (dice): you know the whole distribution — mean, floor, ceiling. Flow adjusts you; it doesn't panic you. Below fair value you should want the trade.

Unanchored (hawker centres): their flow is the only information in the room, so it moves you hard and fast.

Centre the market on your estimate

Estimated 450, quoted 450 / 500 — the estimate sat at the bottom, so the bid had no cushion and got taken. The right 50-wide market was 425 / 475.

Cutting a loss is not a price

It's a decision to stop being long. You still have to sell at a fair price — dumping below fair converts a bad position into a worse realised loss.

Selling a touch under fair is paying for liquidity: ~2 a unit to drain 25 units of drift risk is a good trade — the McDonald's dump paid ~50 a unit for the same service. An earlier version of this card offered at 155, above fair — that's how you quote when you're happy to hold, not when you want out. You caught it (4 Aug): Rule 5, applied back at the coach.

Count the constraint, not the demand

McDonald's was anchored off 7-Eleven (~400) and adjusted up for demand. But the binding constraint is viable sites: a McDonald's needs a kitchen and seating, so it lives in ~170 malls plus standalone lots. A 7-Eleven fits in void decks, petrol stations, MRT stations and office lobbies — five to ten times the location set.

When they say "tighten it"

No new information has arrived — so demand the other dial back.

Know your breakeven, exactly

Sold 5 at 8,500, bought 2 at 12,000 → +18,500 cash against 3 short → breakeven 6,166.7. That's the price where you'd finish flat, and it's what risk limits key off.

Compute it. Don't estimate it.

No trade is a result

When your market properly straddles fair value, the informed counterparty has nothing to take. Getting no fill is often exactly what good market making looks like.

Deriving more than a price

• Bracket the truth from their flow. Sold at 450, passed at 350 → centre near 400. Saying that aloud is a very strong look.
• P&L = spread × volume − adverse selection − inventory cost. After each round, ask which term hurt.
• Break-even hit rate. 50 wide earns 25 per balanced trade, so one 50-loss needs two clean round-trips to repay.
• Think price × size, never price alone.

Why this shape: "Are you familiar with the role" is a knowledge question wearing an experience question's clothes. They want to know you understand what you're signing up for. So answer the knowledge part first and at length, then place the experience gap in one flat clause.

If they push on the gap: don't fill the silence with enthusiasm. "That's the part I'd learn on a desk and can't fake beforehand" — then stop.

Why it works: it refuses the premise without arguing with it. Internships are time-boxed by design — moving between them isn't the flaw, and the only real question is whether there was a thesis behind the choices. You had one, so state it as a strategy rather than defending a pattern.

Have ready: "Which did you like least, and why?" — answer it honestly and without complaint; refusing to rank reads as evasive. And "why not stay at one of them?" → none offered a return that matched what the sampling taught you to want.

Why it holds: it names the pattern before they do, which removes most of its force, then re-diagnoses it as a consistent preference rather than restlessness. The close inverts the whole worry: the reason I left the others is the reason this one sticks. That's a real argument. "I enjoy learning" is an argument for any hard job.

Do not promise a number of years. An unprompted "I'd stay at least five" is unverifiable and sounds bought. The argument is the commitment.

Never under-sell an exonerating fact

• "Didn't break into the industry" → referred in by two employees, reached the final round, an acquisition froze hiring. Same event, completely different candidate.
• The ESG stint → "that was helping my mum with her auditing work." Eight words that kill the scatter read.

Compress the corporate drama to one sentence — it's someone else's story and invites tangents you can't control. Drop "I can prove it"; unprompted evidence reads as defensive.

Why you'll stay — now written up as set-piece 8

The diagnosis that got it there: you keep choosing domains with slow feedback and leaving them — PE pays out in five years, ESG audit in months, the drone bet in years. Trading scores you today.

That's a reason. "I enjoy learning" is an argument for any hard job.

The competing option · don't volunteer, never deny

Your old boss will build an options-pricing desk around you and train you into it. That is leverage, and leverage spent at the interview stage buys nothing — nobody hires you because someone else might. The same fact at the offer stage, when they're setting terms, is worth real money.

Volunteering it also opens a question you'd then have to defend — "if someone's building you a desk, why are you here?" — spending minutes you need for your own case.

Why that shape works: it converts the competing offer into evidence for your thesis rather than a competing pull — and it stays consistent with set-pieces 1 and 8 instead of denting them.

The one place it's a weapon: if they ask what a former manager thought of you, or about references. "My old boss offered to build a desk around me and train me into it" is third-party validation from someone who watched you work — far stronger than any claim you can make yourself. Cold and unprompted it's bragging; as the answer to that question it's devastating. Hold it for that.

Timeline: mention it only if the 2pm produces a concrete dated offer — then it's logistics, not a threat ("I have a decision to make by the 15th"). If it stays exploratory, say nothing. Don't manufacture urgency you don't have.

Market making in one breath

Quote a bid and an offer; earn the spread on balanced flow.

Two enemies. Adverse selection — informed traders pick you off; the people most eager to trade with you know something you don't. Inventory risk — you accumulate a position the market then moves against.

Defences: widen, skew, hedge, size limits.

1 · Adverse selection — the structural one

Someone knows more than you: real news, a large order about to land, or simply a faster machine.

• Stale quotes — fair value moved, your price hasn't, so you're lifted at yesterday's number.
• Queue position — late in the queue means you only get filled when the price is about to run away from you.
• Toxic counterparties — one client whose flow is systematically informed.

This is why Fenix One buys FPGAs. Latency isn't a vanity metric — it's the price of not being last to know.

2 · Inventory risk

You accumulate a position and the market moves before you can offload. Worst in the tails: overnight gaps, trading halts, limit moves.

Crowded exits — everyone widens at once under stress, so precisely when you most need a bid, nobody is there. Carrying inventory also costs real money in financing and borrow.

3 · Your fair value was simply wrong

Bad reference data, a mishandled corporate action, a stale dividend or FX assumption — and you quote confidently around a wrong centre, which is worse than quoting nervously around a right one.

Personal version: the 450 / 500 centring error, and the McDonald's estimate.

4 · The hedge fails

You quote the ETF and hedge with futures, but the hedge is never perfect — basis risk, timing gaps, correlation breaking down exactly when it's stressed.

In stat arb this is the entire risk: the pair stops converging.

5 · The spread is too thin for the risk

Competition compresses spreads, or a fee-and-rebate change flips the economics. You can be right about value and still lose because you weren't paid enough to take the other side.

6 · Operational blowups

Fat fingers, a runaway algo, connectivity loss leaving live quotes in a moving market.

Knight Capital, 2012 — about $440 million in 45 minutes on a bad deployment. Worth knowing by name; it's the canonical example.

Two sentences that sound senior

Shape of the business: many small gains, occasional large losses — negative skew. That's why risk management isn't a side activity at a prop firm, and why Kelly says bet half.

The 15-second answer

Resting limit orders — what you're actually doing

Market order: "buy 100 now, whatever the best price is." Executes immediately against what's already there.

Limit order: "buy 100, but pay no more than $10." If nobody's selling at $10 it doesn't execute — it goes into the book and waits, visible to everyone, until someone takes it or you cancel. That waiting is resting.

Maker and taker

Every trade has exactly one of each. The maker posted the resting order and waited. The taker sent the order that reached out and hit it.

In the game you were always the maker; the interviewer was always the taker.

Why exchanges pay you to be a maker

An exchange with an empty book is worthless — liquidity is the product it sells — so it bribes people into leaving orders there.

Like a marketplace that needs stalls: the operator pays stallholders to show up and display goods, charges shoppers an entry fee, and keeps the difference.

Maker-taker pricing, US equities scale: taker pays ≈$0.0030/share, maker receives ≈$0.0020/share, exchange keeps ≈$0.0010. Fractions of a cent — but on millions of shares it's real, and some strategies are roughly breakeven on the spread and profitable purely on rebates.

The part that sounds sharp

The rebate is compensation, not free money. As the maker you've handed the other side an option — they choose when to trade, and they'll choose the moment your price is wrong. You carry the adverse selection; the rebate partially pays for it.

Which is why queue position matters. At the back of the queue you're only filled after everyone ahead of you — usually meaning the price is about to move through your level. You get the bad fills and miss the good ones.

Not all venues work this way. Some run inverted (taker-maker) pricing, paying the aggressor to attract flow — so where you post is itself a strategic decision.

It's also genuinely contested: maker-taker creates a conflict for brokers routing client orders, since the best-paying venue isn't necessarily the best-filling one. Regulators have poked at it repeatedly.

The Fenix One connection

Per-share rebates are the retail-scale version. What Fenix has with JPX is the contractual form: a market-making incentive scheme where you take on quoting obligations — maximum spread, minimum size, minimum share of the day present at the touch — in exchange for fee discounts or rebates. A contract with the exchange, not just a pricing tier, and why they're on the published market-maker list.

Why the job exists — two problems

The anticipation problem. Buyers and sellers read each other's flow: the moment demand shows, sellers back away upward, and vice versa. Left alone, the spread gapes.

The depth problem. Natural two-way interest is thin, so medium orders cause outsized moves.

An MM's product is a tighter spread plus a deeper book — stability, sold for the spread.

The full revenue stack — five, not one

• Spread capture — primary.
• Maker rebates — above.
• The hedge itself — delta-neutralising in a correlated instrument at good prices is a second edge. (Also why correlated markets stay correlated: everyone hedges in each other's products.)
• Order-flow information — full depth-of-book, market-by-order data. Not front-running (illegal) — better short-term probability models.
• Payment for order flow — brokers sell retail flow to MMs (US retail equities/options); MMs pay because it's uninformed, and price-improve it in return.

Toxic vs non-toxic flow

Desks grade counterparties, not just prices. Non-toxic = uninformed, statistically predictable — retail. Toxic = flow that's systematically right just before the market moves.

The winner's curse, the stranger's coin, and being repeatedly lifted in the game are all the same fact: eagerness to trade with you IS information.

Quote skewing — Rule 5's professional name

Long and want out? Shade both quotes down: your offer becomes the attractive side, flow tilts toward buying from you, inventory drains — sometimes at a small negative edge per fill, which is fine, because it beats carrying the drift risk.

From the outside the same move reads as a nudge: an asymmetric quote pulls flow toward the attractive side. You skew to rebalance; the tape sees a push. Try it live in the Sim tab.

Liquidity engineering — depth as a dial

Internal liquidity sits inside the spread; external is resting orders further out. Stacking depth on one side raises resistance there; thinning it opens a path of least resistance — flow that's already 50/50 walks price toward the thin side.

Skew is the incentive, depth is the resistance.

Second-order: MMs read each other, so one firm re-shaping its book reads as information to the rest — the cohort leans one way with nobody coordinating.

ETF arbitrage — their bread and butter

An ETF has a NAV, the value of its basket. Authorised participants can create or redeem ETF shares against that basket, which is the mechanism pinning price to NAV.

Market makers quote the ETF and hedge instantly with futures or the underlying basket; when price drifts from fair value — the live estimate is the iNAV — they trade the gap.

Japan specifics: TSE/JPX runs a market-making incentive scheme (fee rebates against quoting obligations — that's the list Fenix One is on), and the Bank of Japan historically bought enormous ETF volume, making Japan one of Asia's biggest ETF markets.

Futures

Basis = futures − spot, driven by cost of carry (rates minus dividends or yield). Upward-sloping curve is contango; downward is backwardation.

Asia staples: SGX Nikkei, FTSE China A50, MSCI Taiwan, iron ore.

FX

Majors and pips. Forward points come from the interest-rate differential (covered interest parity).

Prop firms trade FX mostly as fast price-taking and arbitrage, not macro punting.

Crypto

24/7 and fragmented across exchanges, so cross-exchange arbitrage. Perpetual futures track spot via a funding rate paid between longs and shorts; cash-and-carry basis trades are standard.

The real risks are exchange, custody and settlement — not just price.

The latency race

Colocation in the exchange data centre; FPGAs react in nanoseconds where software takes microseconds.

Why it matters: queue priority at the front of the order book, and being first to trade against stale quotes when news moves fair value.

Statistical arbitrage

Find instruments that move together — index versus constituents, dual listings, related futures — trade the divergence, and size by your confidence that it converges.

Regulation

Conditional probability — the one formula under everything

P(A|B) = P(A and B) / P(B)

In words: throw away every world where B didn't happen, then ask what fraction of the surviving worlds also have A. Conditioning = shrinking the universe. Every mistake in this family is someone shrinking to the wrong universe.

Worked: P(two dice sum ≥ 10, given the first die is 5)? Universe shrinks to "first die is 5" (6 worlds of 36). Of those, the second die must be 5 or 6 → 2 worlds. (2/36) / (6/36) = 1/3.

Rearranged, the same formula gives you the other two tools:

• Multiplication rule: P(A and B) = P(A|B)·P(B) = P(B|A)·P(A) — a joint probability is "chance of the condition × chance given the condition", from either direction.
• Total probability: P(E) = P(E|H)·P(H) + P(E|¬H)·P(¬H) — the chance of E is the sum over every route to E. This is the Bayes denominator, and the thing you've now wobbled on three times: it's the whole room, both branches.
• Independence test: A and B are independent exactly when P(A|B) = P(A) — learning B moved nothing.

Bayes — derive it in one line, run it three ways

Never memorise it. Write the multiplication rule from both directions and divide:

P(H|E)·P(E) = P(H and E) = P(E|H)·P(H)  →  P(H|E) = P(E|H)·P(H) / P(E)

Vocabulary, because interviewers use it: prior P(H) — belief before evidence · likelihood P(E|H) — how well the hypothesis explains the evidence · posterior P(H|E) — belief after · base rate — the prior when H is rare.

Way 1 — count people (most robust under pressure): turn probabilities into a population, count who's in the evidence room. Disease: 1,000 people → 1 sick tests positive, ~10 healthy do too → 1 of 11 ≈ 9%.

Way 2 — odds form (fastest, and how desks actually think):

posterior odds = prior odds × likelihood ratio,  LR = P(E|H) / P(E|¬H)

Disease: prior odds 1:999 · LR = 0.99/0.01 = 99 → posterior odds 99:999 ≈ 1:10 → ≈9%. One multiplication, no denominator to fumble. The LR is the honest measure of evidence strength: "this signal is 99× more likely when sick" — and it still couldn't overcome a 1:999 prior.

Way 3 — the formula with the expanded denominator, for writing down and checking:

P(H|E) = P(E|H)·P(H) / [ P(E|H)·P(H) + P(E|¬H)·P(¬H) ]

Kelly — where the formula comes from

The question it answers: a bet is in your favour and repeatable — what fraction f of your bankroll do you stake each round to maximise long-run growth?

Why not bet everything? Single-bet EV says stake 100% — and one loss ends you. Wealth compounds multiplicatively, so what matters is the growth rate of the log of wealth, not the EV of one round.

Even-money derivation, one line: growth per round g(f) = p·ln(1+f) + q·ln(1−f). Set g′(f) = p/(1+f) − q/(1−f) = 0 → f* = p − q (= 2p − 1). At 60/40: f* = 20%.

General form (odds not even): win b per 1 staked → f* = (bp − q)/b = p − q/b. Sanity checks: p = ½ at even money → bet 0 ✓ · p = 1 → bet everything ✓ · longer odds (bigger b) → same p justifies a bigger stake ✓.

• Betting 2× Kelly has ≈ zero growth — all the volatility, none of the drift. Beyond that you're compounding downward with positive EV per bet. That's the sharpest sentence in the room.
• The penalty is asymmetric: under-betting costs you a little growth; over-betting can cost you everything. Your p is an estimate → bet half-Kelly.
• Markets version: optimal size ≈ edge / variance. Same object — size with edge, shrink with uncertainty, then halve for humility.

Monty Hall — it's about how the evidence was made

Why switching wins 2/3: your first pick was right 1/3 of the time, wrong 2/3. The host opening an empty door never changes those numbers — he can always open one, whatever you hold. In every wrong-pick world (2/3 of them), the door he leaves shut is the prize.

If it still itches, scale it: 100 doors, you pick one, the host knowingly opens 98 empty ones. Your door is still the 1-in-100 guess; the one he pointedly left shut carries the other 99%.

The real lesson: condition on the process, not the fact. What something means depends on how it came to be shown to you.

Fractions and decimals

×5 = ×10 ÷2 · ×25 = ×100 ÷4 · ×125 = ×1000 ÷8 · numbers straddling a round one → difference of squares (48 × 52 = 50² − 2²)

Dice and coins

The option to retry always adds value, at a decreasing rate — say that sentence and they know you get it.

Three that collapse under pressure

Each stated in full so you can self-quiz: cover the answer, say the question back, work it aloud.

Bayes. A disease affects 1 in 1,000 people. The test is 99% accurate both ways — 99% of the sick test positive, 99% of the healthy test negative. You test positive. What's the probability you actually have it?

≈9%. Count people: of 1,000, the 1 sick person tests positive, and ~10 of the 999 healthy do too — you're one of ~11 positives and only 1 is sick → 1/11. Formula: (0.001 × 0.99) / (0.001 × 0.99 + 0.999 × 0.01) = 0.00099 / 0.01098 ≈ 9%. Base rates dominate: a rarely-wrong signal still fires wrongly most of the time when the event itself is rare.

Kelly. A coin comes up heads 60% of the time and pays even money. You have $1,000. How much do you bet per flip?

f* = 2p − 1 = 20% of bankroll — $200 on the first flip, always 20% of whatever the bankroll is now. In practice bet half-Kelly (~10%): over-betting punishes far worse than under-betting, and your edge estimate is itself uncertain.

Monty Hall. Three doors, one prize. You pick a door; the host — who knows where the prize is — opens one of the other two, always revealing nothing, and offers you a switch. Do you switch?

Yes — switching wins 2/3, staying wins 1/3. "My first pick was wrong two-thirds of the time, and in every one of those worlds switching wins." The host's knowledge is the whole trick: his door was never a random reveal.

The universal EV script

When any game is posed: 1 restate the payoff · 2 list outcomes and probabilities · 3 compute EV out loud · 4 mention variance or option value · 5 only then state your price.

Fermi in four steps

Decompose → round every input to an easy number → multiply → sanity-check the order of magnitude. Always end with a range, never a point.

How to use this bank

Pick two or three per interviewer, matched to who's in the room: traders get the machine questions, management gets the firm questions, HR gets the training path. Asking a trader about office structure wastes your best access.

Lead with the desk question — it signals you think in specifics rather than job titles.

A question is also a move: the good ones show how you think, so each one below notes what it signals. The Stories tab and the day-on-the-desk sheet point here — this list is the only one to maintain.

The desk — lead with this

1 · “Which desk or market would I sit on first — Japan ETFs, crypto, futures — and what does the training path look like?”

Specific and personal: it makes them describe your actual first year. If they can't, that's information too.

The machine — for traders

2 · “How much of the quoting is automated versus discretionary on the desk I'd sit on?”

3 · “When the model does the pricing, what does the junior trader actually own — parameters, exceptions, hedging? And how is my contribution measured?”

4 · “How do traders and developers split strategy work — who owns a model's parameters day to day?”

All three signal the same thing: you understand the machine does the trading and the trader does something else. Q3 is the sharpest — it's really “what is my P&L attribution when I don't click the buttons”, a question desks genuinely argue about internally.

AI — worth asking, phrased carefully

5 · “Where has the LLM wave actually changed how you work — research, coding, operations — as opposed to the machine learning that's presumably been inside the models for years?”

6 · “If models take more of the pricing and AI takes more of the research grunt work, what does the human trader's edge here look like in five years?”

Why this phrasing. A bare “how do you use AI?” risks sounding like a buzzword question at a firm that has used statistical learning for decades. The ML-vs-LLM split IS the sophistication: ML fits models offline and the fitted parameters ride the fast path; an LLM can't live in a nanosecond loop, so its real impact is the research / coding / ops workflow around the trading. Some firms (XTX most famously) are built around ML forecasting — knowing the difference between that and “we use ChatGPT” is the point.

Q6 is really a career question — “what does the seat I'm applying for become?” — which is why it lands better than the abstract “role of AI in HFT long-term”: that version is an essay prompt that invites a lecture and reveals nothing about either of you.

The firm — for management

7 · “How is trader compensation structured — formula-linked to desk P&L, or discretionary?”

8 · “What's the relationship between the Singapore HQ and the Austin office?”

Q7 is about incentives, not greed — ask the structure, never amounts, and save it for a management round; it's a strange first question to a desk trader.

The people — the closer

9 · “What distinguishes the traders who thrive here in their first year?”

Good last question of any round: whatever they answer, your reply can quietly be “that's me” — and it ends the conversation on their favourite subject, their own good people.