Director · Global Markets AI · APAC

AI for Quant. One page.

A bank-agnostic prep blotter for Markets Quant AI/ML interviews. Work the tickets on the left, in order. Tick a ticket when you finish it. Dotted words are live jargon — hover for the meaning.

I will be the senior AI/ML authority for APAC Markets Quant: hands-on on the first models, competent in coding and modern GenAI, working with engineering — not as a software-engineering hire, and not as a latency specialist unless the desk needs one later.

How to use this page

1. Open 1 · Words if you are not from markets. Read the opening trade story, then search terms as you go.
2. Read 2 · The seat then 3 · Briefing. Say the two-minute opening out loud. 90 days is how you would start the job — not a technical mock.
3. Follow 4 · Study plan. Sit Drills closed, then mocks 5–8 timed. Mock 8 is a live desk scenario. Do not peek at model answers first.
4. If you are new to Python, sit Python first — twenty short sessions in this page, from a bid/ask to a leak-safe pipeline. Sandbox is a blank file: type, Run, read the console.
5. Coding is the lab on 7 · Coding — implement candidate.py and run tests in the page. Do not open the answer key first.
6. Morning of is for interview day only.

They are not looking for software engineers. Competent coding plus fundamentals of modern GenAI is the bar.

Jargon used in this pack

Read this before Briefing if you are not from markets. Every term below appears later in this page.

How to use it: read section 0 once (the picture). Then skim the section that matches the round you are sitting. Search the A–Z list at the end when a mock drops a word you do not know.

0. The picture in one trade

A Global Markets desk at a bank does two jobs at once:

1. Show a price so a client can buy or sell (spot dollars for offshore yuan, a stock, a bond…).
2. Manage the leftover risk after the client trades — the bank is now long or short something.

A simplified FX example:

• The bank quotes USDCNH 7.7800 bid / 7.7804 ask.
• Bid = price at which the bank will buy. Ask = price at which it will sell. The gap is the spread (here 4 pips).
• Mid is the midpoint, 7.7802. That is the “fair” number people argue about; it is not what the client pays.
• A corporate treasurer lifts the ask (buys CNH from the bank). The bank is now short CNH — that leftover is inventory.
• One second later, the market mid is 7.7810. The bank is short something that just got more expensive. That one-second move against the fill is a markout. Repeated markouts from the same clients are toxic flow / adverse selection.
• The trader widens (makes the spread larger) or skews (makes one side worse) so the next informed client is less likely to pick them off. They may also hedge (buy CNH back in another market).
• If this happens electronically on the e-trading platform, it is eFX. If a salesperson emails a request for quote (RFQ), a human or a model still has to answer with a price.

Almost every jargon word in this pack is a name for a piece of that loop: the price, the leftover risk, the client, the machine that shows the price, or the test that says the model is not cheating.

Your job in the interview is not to be the trader. It is to build ML that this loop will turn on: a number they trust, a kill switch, and a metric in PnL (profit and loss), not in Kaggle AUC.

1. A leading IB, the franchise, the people

2. Prices, books, and how the desk gets paid

3. Products (what the client trades)

4. FX in Asia (the local dialect)

5. Electronic trading and the e-trading platform

6. Microstructure words the coding round uses

7. Machine learning as used on a desk

8. Deep learning and GenAI

9. Production, MLOps, controls

10. Papers and names dropped as if everyone had read them

A–Z jump list

Adverse selection · Agentic · Algo · Almgren–Chriss · Ask · ATM · Attention · AUC · Avellaneda–Stoikov · Bandit · e-trading platform · pricing engine · desk assistant · Batch-norm · Bedrock · Bid · bp · Brier · Calibration · Canary · Centralised Quant · Child order · CNH / CNY · Co-lo · Conduct · CPCV · CQL / IQL · CTR · CVaR · Deep hedging · Desk · Distil · DMA · Dual-run · eFX · eSales · Embargo · eTrading · EWMA · Exotic · Fallback · Feature store · FICC · Fill rate · FIX · Fixing · Franchise · G10 · GBDT · Greeks · Greenfield · Hallucination · Hawkes · Hedge · HITL · Horizon · Hot path · HKEX · HKMA · Information set · Internalization · Inventory · IPS · IS · kdb+ · Kill switch · Label · Last look · Latency · Leakage · People-leader craft · LLM · LOB · Lopez de Prado · Macro · Markout · MDP · Microprice · Mid · MLflow · MLOps · MRM · NDF · Next-best action · OFI · Offline RL · Override · PIT · Platt · PnL · PoC · POV · own-book fill · PPO · PRA · Prime · Purge · QPS · QR / QD · Quote · RACI · RAG · RFQ · Regime · Receptive field · Reservation price · RL · RoTE · SageMaker · Session · SHAP · Shadow book · Sharpe · Senior technical authority · Skew · Slippage · SMCR · SOR · Spot · Spread · SS1/23 · Streaming · TCA · Teacher forcing · Tick · Tiering · Tokyo lunch · Toxicity · Transformer · Triple-barrier · TTC · Two-tower · Vanilla · Vector DB · Vol surface · VWAP · Walk-forward · WMR · Widen

If a mock uses a word that is still opaque, pause and redefine it in the interview. Directors who translate for the trader outperform Directors who only perform fluency.

The seat this pack trains

A typical Director – Global Markets AI seat in APAC: Hong Kong (or relocate), hybrid team of 4–6, sitting in a centralised Quant group that exists for Markets. Front-office aligned. Direct influence on trading outcomes. Not a software-engineering hire parked in CIB Tech.

The bar that should change how you sit every round:

Everyone needs to be competent at coding and understand the fundamentals of modern GenAI. You do not need to be an expert at all topics. They are not looking for software engineers.

Electronic / high-frequency trading exposure is advantageous, not essential. Typical profile: already Director for ~2–3 years, now taking a broader leadership remit (team + APAC + strategy), still the senior technical authority.

What to prove

Story bank

Prepare one 90-second STAR per line.

1. Hands-on AI/ML in a stakeholder-heavy shop (desk, risk, or product — not only a lab).
2. A programme you owned from first meeting through deployment and MRM/governance.
3. Markets or IB (APAC colour if you have it; if not, say how you will get it in 90 days).
4. A time you engaged front office / MD and changed the plan.
5. Already operating at Director: what broadens here (APAC, hybrid 4–6, Quant-central, GenAI + modelling).
6. A team you built or scaled, including an LLM/agentic initiative even if small.
7. If you do not have HFT: that is allowed. Show you know when latency matters and when it does not — then partner with eFX engineering. Do not fake 50µs war stories.

How this changes the four rounds

• ML — still the core “technical excellence” exam. Forecasting + recommenders + decision-under-uncertainty are on the syllabus. RL is listed; treat it as offline / simulated unless they push.
• DL / GenAI — agentic is a mandate word, not a side quest. The desk assistant is the live example. You need fundamentals, not every topic.
• Coding — competence, not a software-engineer screen. Python, information sets, a leaky backtest. Skip kdb/Java/latency-system design unless they ask.
• Project — fourth technical: a messy live scenario (leakage, metrics, serving, conduct, GenAI as a constraint). Not a 90-day org plan. How you would start the job lives on 90 days.

Memorise this sentence:

I will be the senior AI/ML authority for APAC Markets Quant: hands-on on the first models, competent in coding and modern GenAI, working with engineering — not as a software-engineering hire, and not as a latency specialist unless the desk needs one later.

Industry briefing: AI for Global Markets

Research snapshot for a generic APAC GM AI/ML Director seat. Map every mock answer back to The seat.

If bid/ask, markout, eFX, purged CV, pricing engine, SS1/23 are not already fluent, read Words first.

1. What this job actually is

You are a Director – Global Markets AI for APAC, sitting in Hong Kong, inside a centralised Quant business of the Investment Bank that exists for Markets. Front-office aligned, direct influence on trading outcomes. You are not a lone alpha researcher, not a CIB Tech engineering lead, and not a desk HFT engineer.

The mandate, in the usual words:

• Build and lead a hybrid AI/ML team of 4–6.
• Own programmes ideation → production and governance.
• Partner with trading desks and APAC business leadership on use cases with measurable PnL.
• Work across stakeholders, engineering teams, and governance — you do not replace engineering.
• Apply price forecasting, recommenders, reinforcement learning, decision-making under uncertainty.
• Initial use cases in FX, proof of concept, then scale across asset classes.
• Responsible AI and model governance.
• Mandate described as extremely broad: quantitative modelling, GenAI, and agentic applications.
• You are the senior technical authority and stay hands-on where required.
• Typical profile: already Director ~2–3 years, now a broader remit (APAC + team + strategy).
• AI/LLM initiatives and team-building are listed requirements.
• eTrading / HFT: advantageous, not essential.
• Recruiter note, typical: they are not looking for software engineers. Coding competence + fundamentals of modern GenAI; you do not need to be an expert at all topics.

A parallel IC/VP posting (AI/ML Modeller – Global Markets, London / SG / HK) is the bench you will hire: Python/PyTorch, SageMaker, eFX/alpha a plus, NLP/LLMs. Your interview mixes Director judgement with enough hands-on technical depth to be that authority — not a staff-engineer algorithms screen.

How this should change your voice: lead with Quant + GenAI/agentic + FX PoC + governance. Do not lead with “I will own the 50µs stack.” That is engineering’s job; your job is to bring models they can put on it.

2. The franchise, in one page

A leading IB. Markets at a house like this already often has a production FX desk assistant used to help colleagues initiate trades. GenAI is usually: pilots with measurable outcomes, then scale, inside a control framework.

Recent Markets mix at a leading IB (the pattern, not a ticker):

• Global Markets is still “the house that Fixed Income built.”
• Equities is the growth lean — often approaching ~30% of Markets revenue in a strong print, with a sharp quarterly jump when cash and prime fire.
• There is a regional tilt toward Asia, helping global clients access Asian markets, including through Prime. A large share of that Prime growth is usually attributed to the top 100 Markets clients.
• Financing vs intermediation mix is a stated strategic guardrail.

Do not walk in talking only about G10 spot eFX. The franchise that is growing APAC wallet is Equities + Prime + financing, while FX remains the most natural first AI production surface.

3. Global Markets and the e-trading platform

GM sells execution, liquidity, risk solutions, and financing to institutions and corporates. Asset-class map:

• Macro: G10 and major EM rates and FX.
• Equities: cash, electronic, equity derivatives, systematic, prime.
• Credit, securitised products, prime services, fixed-income financing.

The e-trading platform is the cross-asset electronic surface (Equities, FICC, futures, FX). For FX specifically:

• Products: spot, forwards, swaps, NDFs/NDS, vanilla and exotic options.
• Execution: streaming, RFQ, own-book fill (principal liquidity), mixed-venue execution (internal + external venues), Manual (trader / benchmarks).
• Access: proprietary UI, FIX API, MDPs (Bloomberg, FXall, FX Connect, 360T).
• Analytics: pre-trade volume, post-trade TCA.
• A desk assistant: client/colleague-facing GenAI on the FX platform.
• Follow-the-sun eSales from Asia open to NY close; FX is marketed as 24/5.
• Cloud-based architecture is now part of the public FX story (continuous enhancement).

APAC electronic footprint (important geographically):

• eFX hubs: New York, London, Tokyo, plus a Singapore pricing/trading engine at a leading IB (MAS FX hub strategy) — often a fourth global FX hub. Local co-lo, lower latency, NDF algos.
• Equities algos and Smart Order Router: hosted in Hong Kong and Japan private DCs; HK is co-located with the primary exchange. Direct feeds for HK and Japan; vendor feeds for other APAC exchanges.

So a HK AI Director “starting with FX” is not sitting on top of the Singapore matching engine. You sit in centralised Markets Quant in HK: the APAC technical authority who partners with London GM AI/ML, Singapore eFX engineering, and HK/Tokyo desks. Say that — then stop short of pretending you are the engine owner. The seat is explicit: not a software-engineer hire; HFT is optional.

4. Org map (roles, not names)

Reporting lines for an HK AI seat are rarely published. Treat these as seats, not people to name-drop.

Implication: first allies are eFX + eSales (PnL, desk assistant already in production at several houses) and HK Equities/Prime quants (APAC growth). Sequence: FX proof of concept, then scale. Do not pick a fight between FX and Equities. Do not staff this as an engineering org.

Behavioural language that plays at a leading IB without sounding like a poster:

• Risk and controls, change and transformation, business acumen, strategic thinking, technology.
• People-leader craft: listen, energise, align across the enterprise, develop others.

5. Industry: what “AI for Markets” means in 2026

Sell-side Markets AI has split into four stacks. A Director who conflates them fails the technical case.

Peer pattern (JPM, GS, Citi, MS, BBVA, flow specialists): tree models and online linear still dominate alpha and pricing; DL wins on limit-order-book images, options surfaces, and hedging policies; LLMs / agents win on interface and workflow, not on mid. Several banks have already productised an FX GenAI interface. The seat wants someone who can run both rails: quantitative modelling and GenAI/agentic.

APAC-specific market facts the desk will assume you know:

• CNH vs CNY, USDCNH as the offshore deliverable, NDFs for KRW, TWD, INR, IDR, PHP.
• Fixings (WMR, local fixes) as benchmark risk, not a toy.
• Tokyo/HK/SG session structure; JPY and AUD as G10 in Asia hours; CNH/KRW vol around US-China and local events.
• Last look, internalization, markouts (50ms / 1s / 10s), and the optics of those at a leading IB.
• HK as relationship and equities/prime hub; Singapore as eFX engine hub.

6. What the three quant tribes need from you

This is the job. Memorise it. Every technical answer should end with who consumes it.

Researchers (QR)

They need a research contract, not a science fair.

• Point-in-time data, as-of joins, corporate-action and fixing calendars.
• Feature store with embargo and purge (no label leakage).
• Experiment tracking (London engineering stories already use MLflow).
• Honest metrics: PnL after costs, capacity, drawdown, not only AUC.
• A path from notebook → library → shadow → production, with them remaining intellectually in the loop.
• GPU/SageMaker access without turning every idea into a six-month cloud programme.
• Permission to publish internally; some of the global heads come from a paper-writing culture.

Your failure mode with researchers: over-engineering MLOps before a first signal, or killing research curiosity with governance theatre.

Developers (QD / eTrading / data engineers)

They need interfaces, SLAs, and boring reliability. You partner; you do not become them. The seat is not hiring a software engineer.

• Stable feature schemas and model contracts.
• Latency budgets when a model does sit near the pricing engine: a research ensemble does not belong on the hot path; a distilled linear or small net might. HFT is optional on your CV — still know when to hand the last mile to Singapore/HK engine owners.
• Deterministic replay, golden datasets, CI that fails on metric drift.
• Kill switch, canary, dual-run, rollback. Observability: prediction, feature, and PnL dashboards.
• Clear ownership: who gets paged at 3am HK / 3am London.

Your failure mode with developers: shadow IT notebooks in production, or promising to rewrite their engine.

Traders (and eSales)

They need control, narrative, and money.

• A number they can argue with: mid, skew, edge, confidence, inventory implication.
• Override. If the model cannot be overridden, it will not be switched on.
• Regime flags (“Asia lunch, low volume, widen”).
• Hedging implication, not just a predicted return.
• Client-colour without leaking other clients’ flow (conduct).
• Speed: a 70% model this week beats a transformer next year.

Your failure mode with traders: black-box theatre, or optimising a loss that is not their P&L.

How a 4–6 person hybrid team should look (a defensible opening)

The seat wants AI/ML technical excellence, not an SE shop.

1. You — senior technical authority; hands-on on model #1 (FX forecast / skew).
2. Applied scientist — price forecasting / recommenders / decision under uncertainty.
3. Applied scientist (GenAI/agentic) — desk-assistant-class tools, RAG, HITL; fundamentals, not a research-LLM tourist.
4. Production-minded quant — Python, PIT features, working with London/SG engineering (registry, shadow). Not “Head of MLOps.”

5–6. Rotators from existing HK QR (Equities algo, Prime) as design partners.

Hire less than six until a desk is using an artefact. Empty seats beat a tourist team. Matrix in London MLOps rather than cloning it.

7. Controls: you will be asked

A leading IB is PRA-regulated. HK is an HKMA authorised presence. Markets models sit in model-risk even when they are “just a pricer”.

PRA SS1/23 (model risk as a risk in its own right): identification and tiering; governance and SMF accountability; development/implementation/use; independent validation; ongoing monitoring. Explicitly includes AI/ML.

HKMA: 2019 high-level AI principles (governance, design/development, ongoing monitoring). 2019 BDAI consumer-protection principles (explainability: no “black-box excuse”). 2024 GenAI consumer-protection circular: human-in-the-loop at early deployment, defined scope, hallucination risk, validation. PCPD AI guidance on privacy.

Conduct: UK SMCR / Individual Accountability; sales-practice and best-execution for algos; information barriers; client confidentiality in recommenders.

A Director phrase that plays well: We will not put a model in a higher tier than the decision it influences. A desk assistant that drafts a ticket is not a pricing model. An eFX mid that auto-quotes is.

8. Intellectual fingerprint of a typical GM AI head

Expect questions near this published taste:

• Recommenders with multiple stakeholders (client, salesperson, desk risk, inventory) — not YouTube CTR.
• Bandits / contextual bandits for execution, pricing exploration, or next-best action.
• Deep hedging and friction-aware hedging policies vs Greek matching.
• Rough vol / non-Markovian kernels and when ML “lifts” a model to something computable.
• Fairness / partial debiasing — in Markets this translates to conduct, client segmentation, and not systematically disadvantaging a client class.
• End-to-end: what changed for the stakeholder?

Do not name-drop papers unless you can discuss them. Do read at least:

• Buehler, Gonon, Teichmann, Wood — Deep Hedging.
• Lopez de Prado — Advances in Financial Machine Learning (purged CV, embargo, CPCV).
• Almgren–Chriss; Avellaneda–Stoikov; Glosten–Milgrom / Kyle.
• Gu, Kelly, Xiu — empirical asset pricing with ML (trees still win a lot of tabular prediction).
• Recent eFX markout / last-look / internalization literature at a conceptual level.

9. How to talk about the role in the opening two minutes

A tight script you can adapt:

A leading IB is investing in a centralised Markets Quant AI/ML function. I would be the APAC Director and senior technical authority in Hong Kong: a hybrid team of 4–6, hands-on on the first models, competent at coding and modern GenAI — not a software-engineering hire, and HFT only if a desk later needs it. The sequence is FX proof of concept first, then scale, with a mandate that is deliberately broad: quantitative modelling and agentic/GenAI. I would land one FX slice the desk can kill (forecasting or skew, measured in markout and fill rate) and an APAC desk-assistant tool path that never invents a price, under PRA/HKMA tiering. Then I clone the rails for Equities/Prime so APAC Markets has a local Quant AI engine, not a London slideshow.

10. Sources (public)

• Typical GM AI/ML Director and modeller postings (London / SG / HK), 2025–2026.
• H1 / Q2 2026 results pattern at a leading IB (GM mix, Equities, Asia, Prime) — read the house you are sitting.
• Public FX platform pages: streaming, RFQ, own-book vs mixed-venue, desk assistants.
• APAC electronic footprint: Singapore FX engines (MAS hub), HK/Japan equities co-lo.
• PRA SS1/23; HKMA 2019 AI principles and 2024 GenAI consumer-protection circular.

First 90 days — how you would start

Not one of the four technical rounds. This is how you would start the job. The fourth technical is Project.

Use this when they ask how you would land, in a fit conversation, or in the opening two minutes. Do not walk this into the project-based technical as if it were the case.

What “started well” means

A leading IB is building GM AI/ML as a greenfield inside centralised Markets Quant. The HK Director is the APAC senior technical authority: FX PoC first, then scale, with a mandate that is quant modelling + GenAI/agentic. You translate desk pain into tiered models existing QR/QD/traders will use. You work with engineering. You do not replace the pricing engine or HK Equities Java, and you do not staff an SE org.

In 90 days the Global Head of GM AI/ML and the APAC Head of Markets will ask: what did we turn on, who uses it, and what did it do to PnL and risk?

Two rails: (A) FX quantitative PoC (B) GenAI/agentic.

Funnel

Kill now

• Transformer on raw mid returns.
• Live RL on the pricing engine.
• Unconstrained LLM prices.
• Firm-wide “AI platform” with no user.

Spike (4 weeks, two people)

1. FX statistical PoC: toxicity / markout or RFQ win-rate for USDCNH + USDJPY Asia sessions.
2. GenAI/agentic PoC: desk-assistant tools for APAC hours (get_mid, get_skew, draft ticket) — never invented numbers; HKMA HITL.
3. Logging: if RFQ propensities are missing, the spike is the log.

Production (one statistical + one workflow, different tiers)

• Skew assistant (trader or rules commit). Metric: 1s/10s markout and fill ratio.
• Agentic ticket draft that can only call the pricer. Separate, lower model tier.

Platform (must ship with it)

• Point-in-time feature spec + golden ticks + MLflow registry + shadow harness that Equities can clone.

Architecture

Desks (HK) ── pain, override, colour
   │
Research (HK Python, SageMaker) ── train, CPCV, notebooks
   │  feature spec + golden tests
Feature / model registry (London MLOps patterns)
   │
Shadow book ── dual run
   │
Hot path (SG pricing engine / HK algo box) ── distilled model or rules
   │
Desk assistant (workflow) ── tools to mid/skew; HITL

Team (first 90 days)

• You — senior technical authority, hands-on on FX model #1.
• 1 applied scientist (forecasting / recommenders / uncertainty).
• 1 applied scientist (GenAI/agentic fundamentals) once Compliance has scoped tools — can be week 3, not day 90.
• 1 production-minded quant working with London/SG engineering.
• 0.5 London MLOps (matrix), 0.5 HK QR design partner.

Do not hire a software-engineering org. The seat forbade that identity.

Controls

• Inventory all artefacts on day 30.
• Tier 0 workflow (assistant) vs Tier 1+ if auto-skew.
• Dual run, kill switch, session-level fallback.
• CNH: train in-region if required; no raw ticks in an unconstrained London notebook.
• GenAI: scope written; no RAG over other clients.

90-day calendar

18 months

Same feature + shadow + registry pattern for HKEX algo fill probability and a Prime client-flow / utilisation tool. Do not start those until FX canary is boring.

Asks

• Tick + RFQ + markout access with a named data owner.
• Intro to SG pricing-engine engineering and HK Equities quant.
• MRM fast-track for Tier 1 skew assistant.
• Air cover to say no to a firm-wide LLM demo week 3.

Fit questions that belong here, not in the technical case

Practise these out loud after you can walk the 90-day page. They are programme / identity questions.

1. London already has a USDJPY model. Why do we need you?
2. Why not hire six PhDs and publish?
3. How do you staff India (offshore tech centre) vs HK?
4. Equities already has kdb and Java algos. Do you replace them?
5. We are not looking for software engineers. Why did you talk about gRPC and 50µs?
6. HFT is not essential. Can you still be the technical authority?
7. You have been Director 2–3 years. What is broader about this remit?
8. Give me a RACI for a production incident at 02:00 London / 09:00 HK.

Behavioural extras

Prepare STAR stories mapped to challenge, drive, align, develop, stewardship:

• Challenge: you stopped a launch; the number looked good.
• Drive: you landed a model through MRM.
• Align: you sat between two desks who wanted opposite loss functions.
• Develop: someone you hired or grew.
• Stewardship: client confidentiality or a conduct near-miss.

Keep each story 90 seconds, with a metric and a control.

Study plan and round design

Assume 10–14 days. If you have less, do the bold items only.

What the four rounds are testing (Director edition)

They will also sample risk and controls and APAC stakeholder craft (HK desks vs Singapore eFX vs London AI/ML).

14-day plan

Days 1–2 — franchise and story

• If you are not from markets or quant research, read Words once end-to-end.
• Read The seat and map seven STAR stories to the requirement list.
• Read Briefing twice. Write your 2-minute opening from the typical seat (centralised Quant, not SE, FX PoC, GenAI + modelling).
• Read 90 days. Say the 90-day page out loud. That is how you would start — it is not technical round 4.
• Map three named internal clients: eFX trader / eSales, Equities algo quant (equities-algo-quant), Prime quant (prime-quant). For each, one pain, one model, one metric, one control.
• Prepare two war stories: (1) a model you killed, (2) a model you productionised through governance. Use numbers.

Days 3–5 — ML

• Lopez de Prado: chapter on purged CV, embargo, CPCV, triple-barrier, fractionally differentiated features (conceptual).
• Calibration (reliability diagrams, Platt/isotonic — and why isotonic can overfit small books).
• Trees vs linear for tabular tick features; SHAP without pretending it is causal.
• Recommenders: implicit feedback, two-tower vs FM/kernel FM, multi-objective (client utility vs desk risk).
• Bandits: UCB, Thompson, contextual; exploration on a priced stream is real money.
• Sit the ML mock timed. Redo any miss as a whiteboard.

Days 6–8 — DL

• Why LSTMs fail on financial ticks (non-stationarity, label horizon vs receptive field, teacher forcing ≠ live).
• Temporal convolutions, TCN, and transformers on returns — usually lose to GBDT unless the representation is rich (LOB images, surfaces).
• Deep hedging: convex risk measures, friction, CVaR; what “Greek-free” means and what Risk will still ask for.
• RL: offline / batch-constrained, sim-to-real gap, conservative Q; never “PPO on live FX”.
• LLM: RAG vs fine-tune vs tools; agentic loop; hallucination of prices; HITL. The typical seat wants fundamentals, not every paper.
• Sit the DL mock.

Days 9–11 — coding

The typical seat: competent at coding, not a software-engineer interview.

• If you are new to Python, finish Python first (20 short sessions in the page). Scratch ideas in Sandbox — type, Run, read the console. Then sit the lab on 7 · Coding until the tests are green. Stubs skip; implemented functions are scored against the answer key.
• Then delete your solutions from working memory and re-implement from the problem statements in 90 minutes.
• Extra: write a 20-line critique of a messy pandas backtest (the debug problem).
• Do not grind leetcode system design, kdb, or Java unless a later interviewer asks.

Days 12–13 — project scenario (technical)

• Sit the project scenario with a timer. Record yourself. It is a live toxicity/last-look case, not a strategy deck.
• Whiteboard: decision, information set, why AUC-on-fills is a trap, costed markout vs fill, what actually runs at 80µs, desk assistant as tools not a pricer.
• If they pivot, the mini-cases (notebook in prod, Prime LLM, vendor TCA, Tokyo-open incident) are still technical.
• Fit / hiring / “why you vs London” stay on 90 days. Do not mix them into this mock.

Day 14 — integration

• Full mock: 30 min ML rapid-fire, 30 min DL, 45 min coding, 45 min project.
• Sleep. Do not cram new papers the night before.

Whiteboard openers to keep on a card

1. Validation: “Walk-forward, purge the embargo by the label horizon, report a Purged Combinatorial CV distribution, not a single Sharpe.”
2. Metric: “I optimise a cost-adjusted markout (or RFQ win-rate at target spread), and I monitor logloss/Brier. I do not ship on AUC.”
3. DL: “I need a representation DL uniquely provides (LOB tensor, surface, language). Otherwise GBDT + online calibration.”
4. GenAI: “No naked prices from an LLM. Tools for mid/skew; LLM for intent and workflow. Human-in-the-loop until the tier says otherwise.”
5. Team: “I am the senior AI/ML authority for APAC Markets Quant. I work with engineering. I am not staffing an SE org.”

Question bank you should have a 90-second answer for

• Why a leading IB, why GM, why HK, why now?
• Why FX first if Equities is the growth story?
• How do you measure success in six months? In eighteen?
• Walk me through a pricing model you would not put on the pricing engine.
• How do you work with a trader who has been burned by ML?
• SS1/23: is a gradient-boosted mid a “model”? Who is the SMF?
• How do London, Singapore, and HK split data, IP, and on-call?
• CNH flow is not EURUSD flow. What changes in the model?
• How would you use the desk assistant without creating a conduct incident?
• Describe how you develop others and align across the enterprise.

Same facts as the long mocks · shorter form

Drills

Ten closed-book questions. Pick in your head, then open the row. The long written mocks still come after this on the blotter.

ML · five closed-book

DL / GenAI · five closed-book

Walk-in ticks (project scenario)

Coverage checks for the live toxicity case — not the 90-day landing page.

Round 1 — Machine Learning (Director mock)

Time: 90 minutes Format: 20 min rapid-fire, 40 min deep dives, 30 min mini-case Rules: No laptop except for a blank sheet. Say assumptions. Tie every method to a desk consumer.

Grade yourself 0–2 per item (0 miss, 1 partial, 2 production-grade). Pass ≥ 28/40. Director bar is not “I have heard of XGBoost”.

A. Rapid-fire (20 min)

Answer in four sentences or fewer. Then check the key.

A1. Leakage

A researcher wants to predict 5-minute USDCNH mid returns with features that include the same-window realised volatility computed from the same mid series, plus a session VWAP that includes the current bar. What is wrong, and what do you do instead?

A2. Why trees still win

Why do gradient-boosted trees still beat deep nets on tabular eFX features (spreads, OFI, inventory, time-of-day, markout lags)? When would you override that default?

A3. The metric trap

A VP shows you AUC 0.62 on predicting “client will lift our offer in the next 50ms”. Why might this model lose money, and which metric would you put on the dashboard for the trader?

A4. Non-stationarity

Name three distinct non-stationarities in APAC FX and a practical ML response to each (not “retrain often”).

A5. Calibration vs ranking

A toxicity classifier is well-ranked (AUC 0.71) but uncalibrated. Why does that break last-look or spread-widen policies, and how do you fix it online?

A6. Regularisation as risk

L2 vs dropout vs early stopping vs feature bagging: which one is a risk control in a markets setting, and which is merely a training trick?

A7. Causal vs predictive

Sales wants a model of “what if we showed this client a tighter spread?” You have observational RFQ data. What identification problem is this, and what is an honest ML substitute if you cannot randomise?

A8. Multiple testing

A QR tests 400 features on two years of EURUSD and keeps those with t>2. What do you require before this set enters a production candidate?

A9. Class imbalance

Toxic flow is 2% of trades. Accuracy is 98%. What do you do to the loss, the sampler, and the decision threshold — three different knobs?

A10. Ensemble of experts

Would you train one global FX model or per-pair specialists (USDCNH, USDJPY, USDKRW NDF)? Give a hierarchical alternative.

B. Deep dives (40 min)

Speak as if a GM QR is at the whiteboard. Draw.

B1. Validation design (the job-defining question)

Design a validation protocol for a USDCNH 1-second mid-move classifier used to skew quotes.

Cover: unit of split, embargo, purge, combinatorial CV, overlapping labels, session effects (Tokyo fix, CNH London hours), costs, and a go-live gate.

They will interrupt: “Just use last 20% as test.” Fight back.

B2. Feature platform

Sketch a point-in-time feature store for eFX that researchers and the Java pricing stack can share. Include: as-of joins, clock (exchange vs local), corporate/fixing calendars, embargo, feature versioning, and what must not live in SageMaker if it is on the hot path.

B3. Recommender for eSales (multi-stakeholder)

the e-trading platform / eSales wants next-best action: which product (spot, NDF, vanilla, hedge) to show which client, subject to desk inventory and risk appetite.

• Why is this a multi-stakeholder recommender, not CTR?
• Implicit vs explicit feedback in RFQ/chat.
• Two-tower vs factorisation machine vs contextual bandit.
• How you prevent systematically worse prices for a client segment (conduct).
• Offline evaluation: IPS / SNIPS, not accuracy.

B4. Online learning

Quotes must track a changing toxicity regime. Compare: nightly batch XGBoost, exponentially weighted logistic, Bayesian linear with forgetting, and a contextual bandit on residual spread.

What can be updated inside the trading day in a bank, and what cannot, given model-risk?

C. Mini-case (30 min)

Prompt. An HK eFX trader says: “Your London USDJPY model is garbage in the Tokyo lunch and on USDCNH. Build me something I will actually turn on.”

Deliver, on one page:

1. Problem statement in PnL language.
2. Labels and horizon.
3. Features (≤12) with economic reason.
4. Model class and why not DL.
5. Validation and shadow-book design.
6. Kill switch and override.
7. Who you staff (from a 4–6 person team) and what you ask London engineering for.
8. 30/60/90 outcomes.

Model answers (fold after you attempt)

A1

Same-window realised vol uses the return you are predicting if the window includes t. Session VWAP including the current bar is direct leakage. Use strictly lagged realised measures, decaying kernels that stop at t−1, and VWAP as-of the previous bar. If you need contemporaneous vol, it must come from other instruments or a leading microstructure channel that will exist at decision time (e.g. lagged futures). Document the information set I_{t−}.

A2

Ticks → sparse, mixed-type, heavy-tailed, missing, strong known interactions (tod × pair × inventory) that trees partition cheaply. Nets need more data, careful normalisation, and still lose on shift. Override when the raw input is a tensor (LOB snapshots, vol surface, text) or you need a policy (hedging) not a score. Even then: DL encoder + tree/linear head is often the grown-up design.

A3

AUC ignores asymmetric cost, size, inventory, and adverse selection after the fill. A high-AUC model can flag retail flow as “easy” and tighten, then get picked off by a last-look predator that is rare but large. Dashboard: markout at 50ms/1s/10s by bucket, fill ratio, internalization quality, and PnL per million vs control. Use a cost-sensitive threshold.

A4

Examples: (1) Session liquidity (Tokyo lunch) → time-of-day features + stateful spread multiplier, not a global residual. (2) Regime/vol (PBOC, US-China headlines) → regime detector, widen or fall back to a robust linear mid. (3) Fixing/event (WMR, CNY fix) → calendar features and model off around events. (4) Balance-sheet / quarter-end → inventory penalty schedule. Retraining is not a response; fallback policies are.

A5

Policies need P(toxic | x), not a rank. If 0.2 means 0.8 in the tail, you will under-widen on the dangerous 2%. Fix: Platt or isotonic on a purged holdout, then online temperature / logistic intercept with a slow learning rate, monitored against realised markouts. Recalibrate per pair and session. Never isotonic-fit on the trading day itself without regularisation.

A6

Early stopping on a purged validation set and feature/model freeze with dual-run are risk controls. Dropout is a training trick that disappears at eval (unless MC-dropout, which you probably will not run at 50µs). L2 is a prior; useful, not a control. Bagging / CPCV distribution of Sharpe is a control because it estimates selection bias.

A7

Confounding: you showed tight spreads to clients you already believed were uninformed, or when inventory wanted the trade. That is not the causal effect of tightening. Honest substitutes: (1) logged propensity of historical quoting policy and IPS; (2) discontinuity at automated spread grids; (3) shadow quotes / small sanctioned exploration with a bandit cap; (4) predict outcome conditional on spread as a forecast, labelled as such, not a treatment effect. Do not say “we used SHAP so it is causal”.

A8

Require: pre-registered hypothesis or train-only discovery; purged multiple-testing (e.g. deflated Sharpe, CPCV); stability across years/pairs; economic cap on feature count; a hold-out pair or hold-out year never used for selection. 400 tests with t>2 is the definition of a false Sharpe.

A9

Loss: weighted logistic or focal, weights = cost of miss vs false widen. Sampler: do not naively downsample the easy majority if those trades pay the bills; keep them but down-weight. Threshold: set on expected markout, not F1. Three knobs, three meetings (QR, trader, MRM).

A10

Hierarchical: global backbone (tod, session, cross-asset vol, USD index) + pair-specific residual (CNH fixing, KRW NDF hours). Complete pooling fails on CNH; no pooling fails on data. Hierarchical Bayesian or a shared GBDT with pair ID + pair×feature interactions. Specialists for top pairs, pooled tail pairs.

B1 — sketch they want

• Decision time t, label = signed move / toxic fill over horizon h (e.g. 1s, 10s). Overlap ⇒ purge any train sample whose time is within h of a test sample (Lopez de Prado embargo).
• Splits: walk-forward by week, not random rows. Embargo around events.
• CPCV: many combinatorial holdouts to get a distribution of PnL, not one lucky year.
• Session: evaluate Tokyo, HK afternoon, London CNH, NY separately. A model that only works in London is a London model.
• Costs: apply typical half-spread + expected widen in the backtest. A classifier that always says “toxic” has great recall and zero franchise.
• Go-live: shadow for 4–6 weeks, dual-run vs current skew, kill if markout or fill-ratio breach. Model tier = high if it auto-skews.

B2 — sketch they want

• Clocks: exchange timestamps, not notebook datetime.now().
• Point-in-time: as-of join; no “last” without asof.
• Two compute paths: slow (SageMaker, research, minutes) vs hot (C++/Java/kdb on box next to the pricing engine — Singapore). HK team owns spec and research; engine owners own hot implementation. Same unit tests, golden ticks.
• Version features as hashes; model binary pins a feature hash.
• Do not put a Python GBDT on the 50µs path; distil to a linear/spline/small net or a lookup.

B3 — sketch they want

• Stakeholders: client (execution quality, fairness), salesperson (wallet, coverage), trader (inventory, franchise, conduct). CTR maximisation will dump risk into the book.
• Implicit feedback: RFQs, chat intents, the e-trading platform clicks; position bias.
• Start with contextual bandit / two-tower on client embedding × product embedding, constrained by a risk layer (hard inventory limits). Kernel FM is a fair adjacent answer if you can explain it.
• Conduct: monitor price quality and reject rates by client cluster; no protected-class proxies; audit logs.
• Offline: inverse-propensity on historical display policy; if logging is missing, you cannot evaluate — that is the first engineering ticket.

B4 — sketch they want

• Intraday: intercept / temperature / EWMA residual of a frozen approved model is often acceptable; full tree retraining is a model change.
• Agree with MRM what is a parameter update vs a new model.
• Bandits: exploration is a client outcome; cap ε, exclude sensitive client classes, log everything.
• Fallback: if live Brier/markout blows up, revert to the approved frozen model.

C — a strong 30-minute answer (outline)

PnL language: reduce adverse markouts on USDCNH and USDJPY Asia lunch without dropping good fill rate more than X bps of franchise.

Label: fill-level markout at 1s and 10s, plus a binary toxic flag at a quantile; horizon matches how fast the desk can hedge.

Features (example): lagged OFI, spread vs session quantile, inventory, time-to-fix, CNH-CNY basis, USDJPY futures lead, residual vs global mid, last N own-fills’ markouts, venue mix, realised vol lagged.

Model: LightGBM or L2 logistic on standardised features; not a transformer. Distil for hot path.

Validation: purged walk-forward, CPCV, session slices, costed PnL.

Shadow: identical quotes in a paper book; trader override logged; weekly review.

Staff: you + 1 modeller + 1 data/quant dev; London MLOps for registry; Singapore eFX for hot path.

30/60/90: feature PIT audit; shadow live; go-live on one pair one session with kill switch.

Self-test MCQ (untimed, after the mock)

1. Purged CV mainly fixes (a) class imbalance (b) overlap of labels with train/test (c) covariate shift (d) vanishing gradients.
2. A reliability diagram plots (a) ROC (b) predicted prob vs empirical freq (c) SHAP (d) ACF.
3. IPS in recommenders needs (a) GPU (b) logged propensities (c) BERT (d) PCA.
4. Deflated Sharpe exists because (a) vol is stochastic (b) selection / multiple testing (c) fat tails only (d) FX swaps.
5. For last-look, you need (a) rank-AUC only (b) calibrated P(toxic) and a costed threshold (c) BLEU (d) FID.

Answers: 1-b, 2-b, 3-b, 4-b, 5-b.

Round 2 — Deep Learning and GenAI (Director mock)

Time: 90 minutes Format: 15 min conceptual, 45 min design, 35 min the desk assistant / hedging case Bar: You must be able to kill a DL project and to defend one. Both. The typical seat lists GenAI and agentic applications as part of an extremely broad mandate, and says you need fundamentals of modern GenAI, not expertise in every topic. C1 (the desk assistant) is that exam. Deep hedging / LOB nets are “when would I use DL” — not a requirement to have shipped HFT.

Pass ≥ 26/36.

A. Conceptual (15 min)

A1. When DL loses

State four concrete reasons a 12-layer Transformer on lagged FX returns will underperform LightGBM on the same tabular features.

A2. Receptive field vs horizon

You predict a 10-second return with a causal TCN. How do kernel size, dilation, and depth set the lookback? What goes wrong if lookback includes a fixing that will not have printed at t in live trading?

A3. Normalisation

Why is global batch-norm dangerous on a stream of ticks? What would you use instead for a live eFX net?

A4. Attention is not causality

A trader sees attention weights on “USDJPY” ticks and concludes those ticks caused the CNH move. Your response, in one minute.

A5. Overfitting the microstructure

A LOB CNN gets 0.57 accuracy on next-event sign in-sample and 0.50 out-of-sample after a matching-engine upgrade. Diagnose, in production language.

A6. Deep hedging in one minute

What problem does deep hedging solve that delta-hedging a Black–Scholes book does not? Name one risk-measure you would train, and one reason Risk will still want Greeks.

B. Design problems (45 min)

B1. LOB or not LOB

HK Equities algo (equities-algo-quant) wants a fill-probability / queue-position model for HKEX. A VP wants a deep LOB image model “like the papers”.

Design two options: (i) classical, (ii) DL. Compare data, latency, explainability, MRM, and a 90-day feasibility on a 4–6 person team. Make a recommendation.

B2. Sequence model for NDF algos

Singapore eFX wants a child-order policy for 1M KRW NDF that reacts to sign of flow and local vol. Compare: supervised imitation of historical algo, offline RL (CQL/IQL), and a simple hawkes + inventory policy. Include sim-to-real and conduct.

B3. Vol surface / FX options

engineering's world includes FX options backtesters. A researcher proposes an operator network / DeepBSDE for USDJPY 25d RR. When is that research, and when is it a desk tool? What would you productionise first instead?

B4. Representation learning

You want a client embedding for the e-trading platform (recommenders, two-towers). Contrast: count-based, matrix factorisation, two-tower, and LLM-encoded chat. Data-leakage and conduct constraints?

C. Cases (35 min)

C1. the desk assistant (already in production)

Group CIO has said Markets is in production with an FX tool that helps initiate trades. You are the APAC Director. eSales wants to let the bot suggest a price and a size.

Write:

1. System diagram: intent → tools → models → ticket.
2. What the LLM is allowed to do vs what it must call.
3. Retrieval (RAG) vs fine-tune vs prompts.
4. Hallucination, stale mid, last-look, and client confidentiality (RAG over other clients’ tickets is a fire).
5. HKMA human-in-the-loop vs a fully straight-through ticket.
6. Telemetry: what you log for MRM and conduct.
7. Why this is a different model tier from eFX mid.

C2. Deep hedging for an FX vanilla book

A QR wants a PyTorch hedging policy for USDJPY 1w ATM, trained on simulated paths with spreads and overnight gaps (APAC).

Cover: environment, friction, objective (MSE vs CVaR vs exponential utility), model risk of the simulator, inventory and cut-off, and how you would shadow vs the desk’s current delta engine. When do you say no?

Model answers

A1

(1) Small effective N: years of ticks are not i.i.d. samples; a net’s capacity fits seasonality and one-off events. (2) Shift: the test month is not the train month; trees + a few economically named features degrade more gracefully. (3) Loss vs PnL: CE on sign ≠ markout; DL will exploit label artefacts. (4) Latency and ops: 12-layer attention is the wrong serving shape; distillation adds another failure mode. (5) Tuning surface: DL has more knobs, so more selection bias. (6) Missingness and regime: trees split on “feature missing”; nets need imputed sentinels you will get wrong.

A2

Causal TCN lookback ≈ 1 + (k−1) * Σ dilation, for typical exponential dilations ~ 2^{depth}. If depth/dilation reaches through the unprinted fix, you trained on a feature that is future information relative to live. Also: if the label horizon is 10s but the net sees 30 minutes of overlapping path, it memorizes local trajectories. Match receptive field to economically available information, and cut sequences on events.

A3

Batch-norm uses batch statistics that (a) differ between train mini-batches of mixed sessions and (b) are not available the same way online (one tick is not a batch). It also leaks via population stats if you update them carelessly. Prefer: frozen per-feature running moments computed point-in-time, LayerNorm / RMSNorm on the channel dim, or online standardisation with slow EMA that is itself a controlled parameter. Never update BN population stats from live in an unapproved way.

A4

Attention is a weighted average in a fitted basis, not a causal effect. It is unstable under multicollinearity (USDJPY and Nikkei futures). Use it as debugging, not as an explanation to a client or to Compliance. If you need attribution, prefer interventions (ablate the feed) or simple ablations, and still speak in predictive language.

A5

The CNN learned engine artefacts (queue priority, lot size, matching rules, feed snapshots) that vanished in the upgrade — classic overfit to venue microstructure. Also possible: label leakage from snapshots that include the event. Production response: freeze, kill, rebuild features from spec of the new engine, shorter memory, simpler model, and a canary on a tiny order-flow slice. This is why DL on LOB needs venue-change tests in CI.

A6

Deep hedging trains a policy that minimises a risk measure of terminal P&L under frictions (spread, impact, discrete time, constraints) where continuous delta in a complete market is a fantasy. Typical: CVaR, expected shortfall, or exponential utility. Risk will still want Greeks, scenario ladders, and limits because the policy is only as good as the simulated measure; they need a language for limit breaches that is not “the net wanted to”.

B1 — recommendation

90-day: classical. Queue position, time-in-force, spread, imbalance at top K, time-of-day, own-order flags — logistic or GBDT, served in the algo box in HK co-lo. LOB images: research track with 1 scientist, not on the live router. DL becomes interesting when you have stable, reconstructed book and a latency budget in milliseconds not microseconds. MRM: image nets are a harder explain story for best-execution. Tell the VP: papers ≠ HKEX production.

B2

Imitation: copies old mistakes and yesterday’s risk appetite; fine as a baseline. Offline RL: need logged policies and propensities; KRW NDF is thin — sim-to-real will dominate. Hawkes + inventory: few parameters, trader-interpretable, easier SS1/23. Path: Hawkes/inventory in prod, imitation as shadow, RL only in simulator until the logging stack exists. Conduct: algos that detect client type from flow can become predatory; keep the objective implementation shortfall + risk, not “pick off the client”.

B3

DeepBSDE / operator nets: research for exotic/RR calibration speed, not a Q3 desk widget. Productionise first: arb-free surface interpolators, PCA/autoencoder on tenors with hard calendar/butterfly constraints, and a fast pricer. A net that violates no-arb is a non-starter. Use DL as a warm start to a classical calibrator.

B4

Start with count / FM on products × clients using the e-trading platform and RFQ logs. Two-tower when you have enough events. LLM on chat: useful for intent, dangerous as the only embedding (PII, off-channel, jokes). Never train on other clients’ confidential flow to predict this client. Separate research embeddings from production embeddings. Reset embeddings on coverage-change. multi-stakeholder bonus: multi-objective ranking (client utility, desk risk, coverage fairness).

C1 — the desk assistant

Diagram: user utterance → intent classifier / LLM with grammar → tools: get_mid(pair), get_skew, get_limits, draft_ticket → confirmation UI → trader or client ack → EMS.

LLM must not invent a mid. It may phrase a mid returned by the pricer. RAG: product docs, playbooks, client-specific permitted static data — not other clients’ tickets, not unscoped chat.

Fine-tune: only if logs are clean and consented; otherwise RAG + tools + small intent model.

HKMA: customer-facing GenAI → human-in-the-loop at least at this stage; define scope in policy (initiate vs price vs commit).

Telemetry: prompt, retrieved docs IDs, tool I/O, user, client, timestamp, model version, whether price was overridden. Retention per policy.

Tier: workflow / assistant, not a pricing model — unless you let it commit risk, in which case it becomes a trading model. That sentence is the interview.

C2 — deep hedging

Simulator must include spread, discrete hedge times, APAC overnight jump, funding. Objective: CVaR_0.05 of discounted P&L plus a transaction penalty. Train off policy in PyTorch; never live.

Say no if: (i) they cannot replicate current desk hedges as a baseline, (ii) no limit layer, (iii) they want to replace the engine before shadowing a single tenor, (iv) the book is client-facing vanilla where explaining a net hedge to a corporate is a conduct problem.

Shadow: same fills, compare hedge error and cost vs delta engine for 1w ATM only. Promote per tenor.

Rapid MCQ

1. Teacher forcing gap is (a) BN (b) train-on-truth vs live own-outputs (c) AdamW (d) dropout.
2. CQL is used because (a) faster GPU (b) offline / overestimation on OOD actions (c) better BLEU (d) stationarity.
3. RAG’s main Markets failure is (a) vanishing gradient (b) wrong or leaked retrieved context, plus stale numbers (c) ReLU (d) PCA.
4. A VAE on a vol surface must still (a) maximise AUC (b) respect no-arb constraints or project onto them (c) use BERT (d) ignore butterflies.
5. MC-dropout at quote time is (a) free (b) usually too slow / unapproved as uncertainty (c) unbiased for PnL (d) required by HKMA.

Answers: 1-b, 2-b, 3-b, 4-b, 5-b.

Round 3 — Coding (Director mock)

Time: 90–120 minutes Language: Python (numpy/pandas). The lab on this page is the mock. Bar: they are not looking for software engineers. You need to be competent at coding: correct information sets, no leakage, code a desk would trust. You do not need kdb, Java, or a 50µs résumé. Style still matters: type hints, no iterrows.

If you are new to Python, finish Python first (20 short sessions in the page).

Scroll to the lab below. Implement candidate.py, then Run tests. Stubs skip until you replace raise NotImplementedError. Your work stays in this browser. The answer key is a separate tab — do not open it first.

broken_backtest.py is in the lab too (read-only). First test run downloads a Python runtime into the browser (cached after that).

If they give you a HackerRank-style screen, still practise these: they are closer to a GM AI Director loop (leakage, microstructure, inventory) than reversing a linked list. Still warm up with one medium array/hash problem the night before so you are not rusty.

Problem 1 — Purged walk-forward (25 min)

Implement purged_walk_forward_splits(t, horizon, n_folds, embargo).

• t is a 1-D sorted array of decision timestamps (int64 ns or int seconds — treat as comparable numbers).
• Each sample i has a label that uses data in (t[i], t[i] + horizon].
• Folds are contiguous in time (walk-forward: fold k’s test is a later block than fold k’s train).
• Purge: drop from train any index whose interval [t, t+horizon] overlaps the test interval expanded by embargo on both sides.
• Return a list of (train_idx, test_idx) numpy int64 arrays. All indices unique within a split; test blocks cover the series as evenly as possible without shuffling.

This is Lopez de Prado, not sklearn TimeSeriesSplit.

Follow-up they will ask: why embargo ≥ horizon; what if labels are triple-barrier and hitting time < horizon; how CPCV differs.

Problem 2 — EWMA fair value and OFI (25 min)

Implement:

• ewma(x, alpha) — causal EWMA, y[0] = x[0], y[t] = alpha x[t] + (1-alpha) y[t-1]. No look-ahead.
• order_flow_imbalance(bid_sz, ask_sz, bid_px, ask_px) — a simple level-1 OFI:

For each t > 0:

• bid increment: if bid_px up, +bid_sz[t]; if unchanged, bid_sz[t]−bid_sz[t-1]; if down, −bid_sz[t-1]
• ask increment: if ask_px down, +ask_sz[t]; if unchanged, ask_sz[t]−ask_sz[t-1]; if up, −ask_sz[t-1]
• OFI = bid_increment − ask_increment

Index 0 is 0.

• microprice(bid_px, ask_px, bid_sz, ask_sz) = (ask_px bid_sz + bid_px ask_sz) / (bid_sz + ask_sz) with safe denominator.

Follow-up: why microprice ≠ mid; how you would standardise OFI online; what happens at a tick-size change.

Problem 3 — Inventory-aware quotes (20 min)

Implement reservation_quote(mid, inventory, q_max, gamma, sigma, horizon, kappa, fee).

A simplified Avellaneda–Stoikov style:

• reservation r = mid - inventory gamma sigma**2 * horizon
• half-spread d = (1/gamma) log(1 + gamma/kappa) + 0.5 gamma * sigma**2 * horizon + fee (if gamma≈0, use 1/kappa + fee as the finite limit)
• bid = r − d, ask = r + d
• Clip inventory effect so quotes do not cross: if bid > ask, set both to mid.

Return dict with keys reservation, bid, ask, half_spread (all ndarray).

Follow-up: map gamma to a desk risk appetite; what you monitor live (inventory, realised spread vs model, fill asymmetry).

Problem 4 — Debug the backtest (20–30 min)

broken_backtest.py contains a deliberately leaky strategy simulator. Do not rewrite from scratch. List the bugs (there are at least five), then implement fixed_backtest(df) in candidate.py that:

• trades only on information available at bar close of t for a position held over t+1
• applies spread cost
• reports net PnL mean and a purged Sharpe that does not use future vol

Write the bug list in comments above fixed_backtest.

What interviewers listen for while you type

• You define time before you write a loop.
• You ask whether timestamps are decision time or event time.
• You refuse shift(-1) “to align the label” without an embargo.
• You vectorise, but you can write the O(n) loop correctly first.
• You mention tests: a unit test where leakage would make PnL explode.
• You mention integer cents / tick size rather than float fantasy, even if the toy uses floats.

If they switch to “systems” instead of numpy

Only if they go there. The typical seat is not a software-engineer screen. Keep it to 10 minutes:

• You specify the model; Singapore/HK engineering owns the hot path.
• Registry, canary, kill switch, information set.
• You do not need to design exactly-once fill streams unless asked.

Round 4 — Project / scenario (technical)

Time: 90 minutes (45 min working + 45 min interrogation) Deliverable: a diagnosis and a technical recommendation you could defend at a whiteboard: decision, information set, metrics, model, serving, what you would not do, kill criteria.

This is still a technical round. It is not a 90-day org plan, a hiring deck, or “how I would start as Director.” That material lives on 90 days. Here they put a messy live problem in front of you and listen for whether you can think like the senior technical authority.

Case: the toxicity score that “worked”

Prompt (read slowly, then start the clock).

APAC eFX streams USDCNH on the e-trading platform during the Asia session. Six weeks ago a GBDT toxicity score went live: if P(toxic) > 0.6 the engine last-looks or widens. A researcher reported AUC 0.71 on a random 70/30 split. The model retrains every night on yesterday’s fills. The score is fit in Python on SageMaker and pushed toward the pricing engine; eFX says the engine budget is about 80µs.

Since go-live:
- 1-second markout improved 0.2 bp
- fill rate fell 8%
- CNH corporate flow in the London/Asia overlap collapsed
- a salesperson: “the model hates our franchise clients”
- Compliance asks whether last-look is now worse for one client class

eSales also wants the desk assistant to tell the client why we widened, live.

Diagnose. Then propose what you would measure, what you would change (label, split, model, threshold, serving), what you would not do, and how you would know to kill it. Do not design a 90-day programme or a 4–6 person org chart.

Suggested structure (use it)

1. Restate the decision (5 min) — what the model is allowed to change (last-look / widen / skew), whose PnL, what “good” cannot be.
2. Diagnose (15 min) — why the reported AUC can be real and the desk still be losing; leakage; label construction; feedback loops; client mix.
3. Technical design (15 min) — information set, validation, metric, model class, calibration, threshold, where inference actually runs.
4. The desk-assistant ask (5 min) — tools vs invented prices; tiering; what must not go in the prompt.
5. Ship / kill (5 min) — shadow vs canary, fallback quote, signed kill criteria, CNH residency if it bites this pair.

Interrogation will try to break the diagnosis. Invite it.

Interrogation bank (practise out loud)

1. AUC 0.71 — why is that the wrong headline?
2. What is wrong with training only on fills?
3. Walk the information set at quote time vs the label window.
4. How does last-look poison the next day’s labels?
5. Markout +0.2 bp, fill −8%. Did you succeed?
6. How would you test “we are disadvantaging a client class”?
7. SageMaker cannot hit 80µs. So what actually ships, and what is SageMaker for?
8. How do you distil, and what drift would make you roll back the small model?
9. CNH ticks cannot leave HK. How do you train and validate?
10. The desk assistant explains the widen without inventing a mid or RAG-ing other clients. How?
11. Would you turn last-look off and only widen? Why?
12. London has a USDJPY toxicity model. What transfers, and what breaks on USDCNH?
13. A VP wants GPT-4 on the quote stream. You say?
14. What is the fallback quote when the kill switch fires?
15. Is the toxicity score a “model” under SS1/23? Is the assistant sentence?
16. How do you A/B when the policy changes who fills?
17. Inventory was ignored. How does it enter the decision without turning this into a second pricer you do not own?
18. The desk wants it “fixed this week.” What is the one-week patch vs the real fix?

Scoring rubric (grade yourself honestly)

Pass: average ≥ 4 and no 1s on diagnosis, metric, or information set.

A strong reference (do not memorise; internalise)

Restate

The model is not “predicting toxicity.” It is a policy: last-look or widen on USDCNH streaming. Success is franchise PnL — markout and fill, by client segment and session — not AUC. Last-look is a conduct surface, not a free option.

Diagnose (the actual bugs)

• Random 70/30 on tick data with a forward markout label leaks. Overlapping windows, no purge, no embargo.
• Train on fills only. Rejects, last-look declines, and “widened so they left” never enter the label. The model is scored on the survivors of yesterday’s policy. Nightly retrain locks that in.
• AUC 0.71 can be true and useless. The desk uses a threshold. They need calibration (reliability by session/size) and a costed threshold: expected markout saved vs franchise fill lost.
• +0.2 bp / −8% fill is not a win by default. Corporate CNH in overlap is exactly the franchise they cannot scare. Segment the metric: HF vs corporate, overlap vs Tokyo morning, size buckets.
• Last-look as a function of a score changes who gets filled, which changes tomorrow’s training mix — a feedback loop. It can also look like systematically worse treatment of a client class.
• SageMaker → 80µs is a category error. The engine cannot wait on a Python GBDT in the cloud. Offline research vs distilled linear / small net / rules on the pricing engine.

Technical design

• Label: 1s and 10s markout and fill, with size; not a binary “toxic” cooked to maximise AUC. If you keep a toxic flag, define it from post-trade markout given fill, not from “the last-look fired.”
• Split: walk-forward, purge by horizon, embargo. Report a CPCV distribution, not one lucky week.
• Features: lagged OFI, spread vs session, inventory, time-to-fix, CNH-CNY basis, own last-N markouts, venue mix — all as-of decision time. No same-window realised vol of the mid you are defending.
• Model: GBDT or penalised linear offline. Distil to whatever the engine will take. No transformer on mids. No live RL. No LLM mid.
• Threshold: chosen on costed shadow PnL, by session, with a cap on fill-rate damage. Override stays with the trader.
• Serving: feature contract + golden ticks; inference next to the pricing engine or a pre-computed skew table; SageMaker for fit and batch eval only.

Desk assistant

Different artefact, lower tier. The bot calls get_mid, get_skew, get_reason_code (approved vocabulary: “wider because inventory / session / last-look policy”). It does not compute the score, does not RAG other clients, does not invent a number. HITL while client-facing. SS1/23: a drafted sentence is not a pricing model unless you let it commit risk.

Ship / kill

• Shadow the new label/threshold against the live policy before touching last-look.
• Canary one session (not all CNH corporates first).
• Kill switch: revert to pre-model widen rules / last-look policy. Fallback must be a quote, not “error.”
• Off-criteria signed: fill-rate drop vs a named client segment, or markout worse than baseline, or a client-class reject gap.
• CNH: train in-region if required.

What you say no to this week

Bigger net. Lower the 0.6 because AUC looked good. GPT on the stream. Replacing the pricing engine. A firm-wide AI platform. Hiring a team to “own the problem” instead of changing the label.

Mini-cases (if they pivot)

Still technical scenarios. Sit these in 20 minutes each after the main case.

M1. The researcher’s notebook is in production. Overnight job on a desktop, no tests, making money. What do you do in week 1 vs week 8? (Do not kill PnL on day one; wrap, test, dual-run, then retire the desktop.)

M2. Prime wants an LLM over client inventory. Conduct, information barriers, who the user is, why a constrained tool beats a chatbot.

M3. Build vs buy. Vendor claims 20 bp TCA improvement. How you trial (A/B, capacity, last-look policy alignment, data egress, kill).

M4. Incident. Model widened everyone at Tokyo open, franchise volume collapsed. Walk the bridge: detect, kill, fallback, client comms, post-mortem — technically, who inspects which log.

Morning-of one-pager

If a word here is still opaque, look it up in Words — do not walk in hoping the panel will translate.

Opening (≈90 seconds)

Centralised Markets Quant AI in APAC. I am the HK Director and senior technical authority: hybrid 4–6, hands-on, not a software engineer, HFT optional. Mandate is broad — quantitative modelling and GenAI/agentic. Sequence: FX PoC, then scale. 90 days: one FX slice the desk can kill and a desk-assistant tool that never invents a mid. Then Equities/Prime on the same rails.

Seats, not names

Head of GM · CEO APAC · Head of Markets APAC · Global Head of AI/ML GM (London) · AI/ML & Quant Engineering · Equities algo quant product (HK) · Prime quant (HK).

Behavioural: risk and controls, change, business acumen. People-leader: listen, energise, align, develop.

Numbers

Recent prints at a leading IB: GM still FICC-heavy; Equities the growth lean (often ~30% of Markets in a strong half, sharp quarterly jumps). Asia tilt via Prime, top-100 clients.

Geography

SG = eFX engine (you partner, you do not own it). HK/Tokyo = Equities algos/SOR. Distil if a model must sit near the engine; do not lead with 50µs — the seat is not hiring HFT.

Four round one-liners

• ML: purged walk-forward + embargo; costed markouts not AUC; trees first; recommenders are multi-stakeholder (forecasting, recommenders, uncertainty).
• DL: DL only for tensors/policies; desk assistant tools for prices; agentic is on the mandate — fundamentals, not every topic; no live RL.
• Coding: competent Python and information sets. They are not hiring a software engineer.
• Project: live toxicity scenario — diagnose leakage / fill-only labels / AUC trap; costed markout vs fill; distil to the engine; desk assistant tools ≠ pricing tier. Not a 90-day org plan.

Kill list

Transformer on mids · PPO live · LLM-invented mid · RAG over other clients · firm-wide platform with no user · replacing HK Equities Java algos on day one.

First team

(This is the landing crib, not the project-round answer.)

You (hands-on authority) + applied scientist (forecast/recommender) + GenAI/agentic scientist + production-minded quant working with engineering. Matrix London MLOps. Do not hire six SEs.

Controls

SS1/23 inventory and tiering. HKMA HITL on customer-facing GenAI. CNH residency. Dual-run. Signed kill criteria.