Practical Fundamentals of
High-Frequency Trading
Concepts, Workflows, and Realistic Examples for Work & Interviews
Emphasis: limit order books, latency, fill dynamics, execution, production risk
Audience: HFT researchers, traders, quant devs, and interview candidates
Compiled on January 30, 2026
Practical HFT Fundamentals Microstructure, Execution, Production
Contents
1 What HFT Actually Is (and What It Is Not) 3
1.1 The HFT objective function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Roles and boundaries in a real HFT team . . . . . . . . . . . . . . . . . . . . . . 3
2 LOB Microstructure Essentials (the world your strategy lives in) 3
2.1 Basic primitives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.2 Tick size and price-time priority . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.3 Microprice and imbalance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.4 Adverse selection and toxicity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
3 Data and Time: Timestamps, Feeds, and Latency Budget 4
3.1 Market data types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
3.2 Timestamp taxonomy (critical in HFT) . . . . . . . . . . . . . . . . . . . . . . . 4
3.3 Latency budget . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
4 PnL in HFT: Accounting that Matches Reality 5
4.1 Per-fill PnL decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
4.2 Implementation shortfall (HFT version) . . . . . . . . . . . . . . . . . . . . . . . 5
4.3 Inventory risk is not optional . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
5 HFT Alpha: What Can Be Predicted at Millisecond Horizons 5
5.1 Order flow imbalance as a short-horizon predictor . . . . . . . . . . . . . . . . . . 6
5.2 Microprice reversion and “fade” signals . . . . . . . . . . . . . . . . . . . . . . . . 6
6 Fill Probability and Queue Position (the hidden core of HFT) 6
6.1 Queue position model (practical approximation) . . . . . . . . . . . . . . . . . . 6
6.2 Queue priority vs. adverse selection . . . . . . . . . . . . . . . . . . . . . . . . . 7
7 Market Making: Quoting, Inventory, and Adverse Selection 7
7.1 Core market making problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
7.2 Inventory skew . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
7.3 Spread setting under toxicity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
8 Execution and Order Types: IOC/FOK, Pegs, Requotes, and Throttles 8
8.1 Order types that matter in HFT . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
8.2 Cancel/replace and message throttles . . . . . . . . . . . . . . . . . . . . . . . . . 8
8.3 Quote life-cycle and state machine . . . . . . . . . . . . . . . . . . . . . . . . . . 9
9 Backtesting HFT: Event-Driven Simulation and Fill Modeling 9
9.1 Why bar-based backtests fail for HFT . . . . . . . . . . . . . . . . . . . . . . . . 9
9.2 Event-driven backtest loop . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
9.3 Fill simulation: conservative and realistic . . . . . . . . . . . . . . . . . . . . . . 10
9.4 Latency in backtests (must-have) . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1
Practical HFT Fundamentals Microstructure, Execution, Production
10 Overfitting and Robustness: HFT-Specific Failure Modes 10
10.1 Common traps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
10.2 Robust validation practices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
11 Risk Management in HFT: Real-time Limits and Kill Switches 11
11.1 Risk dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
11.2 Real-time controls (practical) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
11.3 Markout-based toxicity circuit breaker . . . . . . . . . . . . . . . . . . . . . . . . 12
12 Production Systems: Feed Handlers, OMS, Observability 12
12.1 Architecture (high level) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
12.2 State consistency and reconciliation . . . . . . . . . . . . . . . . . . . . . . . . . 12
12.3 Observability metrics that matter . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
13 End-to-End Case Study: Large-Tick Market Making with Queue Modeling 13
13.1 Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
13.2 Signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
13.3 Decision logic (realistic) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
13.4 Backtest requirements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
13.5 Failure modes and mitigations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
14 Interview Playbook: Questions, Drill-downs, and Strong Answer Templates 14
14.1 How interviewers probe HFT candidates . . . . . . . . . . . . . . . . . . . . . . . 14
14.2 Answer framework that works . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
14.3 Representative interview questions (with desk-oriented hints) . . . . . . . . . . . 14
15 Appendix: Practical Checklists 14
15.1 HFT research checklist . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
15.2 Production readiness checklist . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2
Practical HFT Fundamentals Microstructure, Execution, Production
1 What HFT Actually Is (and What It Is Not)
High-frequency trading (HFT) is market microstructure engineering under uncertainty.
A successful HFT strategy is rarely about predicting “daily returns”. Instead, it is about
extracting edge from:
• short-horizon price formation (milliseconds to seconds),
• order flow and queue dynamics,
• latency advantages and information asymmetry,
• execution quality (fills, adverse selection, cancellations),
• robust production systems (data integrity, fail-safes).
1.1 The HFT objective function
At very short horizons, expected gross edge per trade is small. Therefore, any realistic objective
must subtract micro-costs precisely:
max E[PnL]−E[spread paid] − E[adverse selection] − E[impact] − fees/rebates − operational losses .
| {z } | {z } | {z } | {z } | {z }
crossing toxicity your footprint maker-taker stale data/outages
Interviewers want to see that you treat execution and latency as first-class.
1.2 Roles and boundaries in a real HFT team
Even small HFT teams usually split responsibilities:
• Research: feature engineering from LOB, alpha modeling, offline validation.
• Trading: quoting logic, risk limits, real-time parameter tuning.
• Quant dev/engineering: feed handlers, OMS, risk checks, simulation framework.
• Infra: colocation, kernel bypass, time sync, observability.
A strong candidate can explain not only models but also production constraints.
2 LOB Microstructure Essentials (the world your strategy lives
in)
2.1 Basic primitives
A limit order book (LOB) is defined by:
• best bid bt , best ask at , spread st = at − bt ,
• depths at price levels Dk,t
bid , D ask ,
k,t
• trades (prints) with size and aggressor side.
at +bt
The midprice is mt = 2 , but mid is not tradable.
3
Practical HFT Fundamentals Microstructure, Execution, Production
2.2 Tick size and price-time priority
Tick size shapes everything:
• large tick assets: spread often one tick; queue competition dominates.
• small tick assets: spread varies; pricing and adverse selection dominate.
Most venues use price-time priority: better price first, then earlier time.
2.3 Microprice and imbalance
A common microstructure feature is the microprice:
bid + b · D ask
at · D1,t t 1,t
micropricet = bid + D ask
.
D1,t 1,t
If bid depth is larger, microprice tilts upward (and vice versa). It is often more informative than
mid at sub-second horizons.
Depth imbalance:
bid − D ask
D1,t 1,t
It = bid ask + ϵ
.
D1,t + D1,t
2.4 Adverse selection and toxicity
When you provide liquidity (resting limit orders), you earn rebates/spread only if you are not
consistently picked off by informed flow. A practical diagnostic is the markout after a fill:
Markout(∆) = sign(fill) · (mt+∆ − pfill ),
where sign is +1 for buys, −1 for sells. Negative markouts indicate toxicity.
3 Data and Time: Timestamps, Feeds, and Latency Budget
3.1 Market data types
• Top-of-book (L1): best bid/ask + sizes.
• Depth (L2/L3): multiple levels or full order-by-order.
• Trades (prints): executions, sometimes with aggressor side.
• Reference data: symbol mappings, tick size tables, trading status.
3.2 Timestamp taxonomy (critical in HFT)
You must distinguish:
• exchange timestamp (when event occurred at venue),
• receive timestamp (when your NIC/kernel saw it),
• process timestamp (when your strategy consumed it),
4
Practical HFT Fundamentals Microstructure, Execution, Production
• send timestamp and ack timestamp for orders.
Many backtest mistakes come from mixing these.
3.3 Latency budget
Your decision loop has a latency budget:
Total Latency = Lfeed + Lparse + Lcompute + Lrisk + Lsend + Lexchange .
Edge can be destroyed if latency variance (jitter) is high even when mean latency is low. In
interviews, talk about P50/P90/P99 latencies, not just averages.
Example 1 (Realistic latency constraint). Suppose your alpha half-life is 150ms. If your
round-trip (market data receive to order ack) is 300ms, you are trading stale information. Then,
you must either: (i) shift to longer-horizon features (seconds), or (ii) reduce complexity / colocate
/ optimize pipeline, or (iii) trade passively and let the market come to you (but then fill risk
increases).
4 PnL in HFT: Accounting that Matches Reality
4.1 Per-fill PnL decomposition
For a fill of size q at price p, define reference mid m at decision time. Short-horizon expected
PnL is:
PnL ≈ q · (mt+∆ − p) + rebate/fee − inventory cost − impact from your actions.
If you are a liquidity provider, rebates can be material. If you are taker-heavy, spread dominates.
4.2 Implementation shortfall (HFT version)
Decision price p⋆ could be mid or microprice at decision time. Implementation shortfall:
X
IS = ∆qi (pi − p⋆ ) + fees.
i
In HFT, you often compute IS at multiple horizons (instant, 50ms, 200ms, 1s) to separate spread
from adverse selection.
4.3 Inventory risk is not optional
Even “market-neutral” HFT strategies accumulate inventory intraday. Inventory cost can be
modeled as:
λ
InvCost ∼ qt2 ,
2
where λ scales with volatility and risk limits. This feeds directly into quoting skew.
5 HFT Alpha: What Can Be Predicted at Millisecond Horizons
HFT alpha is usually conditional and state-dependent. Common categories:
5
Practical HFT Fundamentals Microstructure, Execution, Production
• Order flow / imbalance: predicts short-term price pressure.
• Queue/priority effects: predicting fill probability and adverse selection.
• Cross-venue lead-lag: one venue moves first; others follow.
• Microstructure mean reversion: temporary dislocations revert.
5.1 Order flow imbalance as a short-horizon predictor
A simple but realistic feature:
bid ask
OFIt = ∆D1,t − ∆D1,t ,
where depth changes incorporate adds/cancels/trades at best quotes. A regression model might
be:
∆mt→t+h ≈ β0 + β1 OFIt + β2 It + β3 ∆st + ε.
But in production, you must worry about:
• feed gaps and out-of-order messages,
• exchange-specific definition of cancel/modify,
• hidden liquidity and midpoint facilities.
5.2 Microprice reversion and “fade” signals
In large-tick markets, microprice deviates from mid in a predictive way. A trade might be: if
microprice is above mid and imbalance positive, expect next mid uptick. But your realized edge
depends on whether you can be in the queue at the right time.
Example 2 (Non-toy alpha: microprice + trade sign + volatility regime). State variables:
• z1 = (microprice − mt )/tick,
• z2 = EWMA(trade sign) over last 50ms,
• z3 = short-term volatility regime indicator (e.g., high/low).
Decision: quote tighter only when (z1 , z2 ) agree and z3 is low; otherwise widen or cancel. This is
realistic because toxicity spikes in high-vol regimes, so you dynamically control exposure.
6 Fill Probability and Queue Position (the hidden core of HFT)
In HFT, you do not choose only price; you choose queue position.
6.1 Queue position model (practical approximation)
Let your queue position at best bid be Qt (shares ahead of you). Your limit buy fills when
cumulative sell market orders (and cancels ahead) exceed Qt . A simple hazard model:
Z t+∆
P(fill in [t, t + ∆]) = 1 − exp{− λ(u) du},
t
where intensity λ(u) depends on:
6
Practical HFT Fundamentals Microstructure, Execution, Production
• recent trade rate (aggressor sells),
• imbalance (predicts next aggressor side),
• spread and price level stability,
• queue churn (cancels ahead of you).
6.2 Queue priority vs. adverse selection
Improving queue position by joining early helps fills, but increases toxic exposure if you get
filled right before price moves against you. Therefore HFT often uses quote life management:
• join early in low-toxicity regimes,
• pull quickly when toxicity indicators spike,
• reprice when the book shifts to avoid being “last look” victim.
Example 3 (Realistic decision criterion: expected utility of resting vs crossing). Let:
• saved spread by resting: ≈ s/2,
• expected adverse selection conditional on fill: A (bps),
• fill probability within ∆: p,
• missed alpha if no fill: M (bps).
Expected net benefit of resting (vs not trading) approximately:
s
Net ≈ p · − A − (1 − p) · M.
2
If Net < 0, you either cross (market/IOC) or skip. This kind of reasoning impresses interviewers
because it ties microstructure to action.
7 Market Making: Quoting, Inventory, and Adverse Selection
7.1 Core market making problem
You quote bid/ask around a reference price and manage inventory. A common quoting form:
s s
pbid = mt − − δ bid , pask = mt + + δ ask ,
2 2
where δ adjusts for risk/toxicity/inventory.
7.2 Inventory skew
If you are long inventory, you want to sell: lower ask, raise bid (or widen bid). A simple skew:
δ ask = kqt , δ bid = −kqt ,
so quotes tilt to mean-revert inventory.
7
Practical HFT Fundamentals Microstructure, Execution, Production
7.3 Spread setting under toxicity
Spread is not only compensation for volatility; it is also defense against informed flow. In
high-toxicity regimes, optimal action can be:
• widen spread,
• reduce size,
• cancel entirely,
• switch to taker hedging in correlated instruments.
Example 4 (Non-toy market making: correlated hedge + inventory cap + time-of-day).
Instrument A: liquid futures contract; Instrument B: ETF or another venue. You make markets
in A but hedge inventory using B when:
• A’s spread is tight but toxicity rises (adverse selection marker spikes),
• your inventory hits soft limit,
• cross-venue lead-lag suggests B has moved first.
Risk controls: hard inventory limit; max order rate; kill-switch on abnormal fill/markout. This
is realistic and commonly implemented on professional desks.
8 Execution and Order Types: IOC/FOK, Pegs, Requotes, and
Throttles
8.1 Order types that matter in HFT
• Limit: join queue; uncertain fill.
• Market: immediate; pay spread + potential impact.
• IOC: take available liquidity up to size; cancel remainder.
• FOK: fill-or-kill; useful for strict size needs.
• Pegged: midpoint or primary peg; depends on venue rules.
8.2 Cancel/replace and message throttles
HFT strategies generate huge message traffic; venues enforce:
• max order rate limits,
• cancel-to-trade ratio constraints,
• penalties for excessive messaging or “quote stuffing” behavior.
Your strategy must incorporate throttling and batching.
8
Practical HFT Fundamentals Microstructure, Execution, Production
8.3 Quote life-cycle and state machine
In production, quoting is a state machine (per symbol):
• desired quote → send new/replace → pending ack,
• handle rejects, partial fills, cancels, trading halts,
• maintain local book vs. exchange book synchronization.
Listing 1: Illustrative quote state machine (conceptual)
if venue_status != " OPEN " :
cancel_all ()
return
desired = c o m p u t e _ d e s i r e d _ q u o t e s ( state )
if pending_ack :
if now - last_send_ts > ack_timeout_ms :
enter_safe_mode ( " stuck ack " )
return
if r i s k _ l i m i t s _ b r e a c h e d ( state ) :
w i d e n _ q u o t e s _ o r _ c a n c e l ()
return
if t o o _ m a n y _ m s g s _ r e c e n t l y () :
rate_limit ()
return
s e n d _ r e p l a c e _ i f _ n e e d e d ( desired , state )
9 Backtesting HFT: Event-Driven Simulation and Fill Modeling
9.1 Why bar-based backtests fail for HFT
Minute bars cannot reproduce:
• queue priority,
• partial fills,
• cancels and modifies,
• microprice dynamics and short-lived dislocations,
• latency and asynchronous multi-venue feeds.
HFT backtests must be event-driven.
9
Practical HFT Fundamentals Microstructure, Execution, Production
9.2 Event-driven backtest loop
At minimum, you simulate:
• market data events (book updates, trades),
• strategy decisions (quote updates),
• order acknowledgements,
• fills (full/partial) based on queue and aggressor flow,
• latency delays and time alignment.
9.3 Fill simulation: conservative and realistic
A pragmatic approach:
• maintain synthetic queue position based on visible depth,
• decrement queue with trades at your price level,
• incorporate churn with cancels ahead using empirically estimated rate,
• treat hidden liquidity as uncertainty (scenario-based).
Example 5 (Realistic fill model sanity checks). Your simulated strategy should reproduce:
• fill rate by spread state (1 tick vs 2 ticks),
• fill probability vs. queue position decile,
• average markout of fills (toxicity signature),
• distribution of holding times and inventory.
If your simulation produces “too good” fills at mid or ignores queue, it is invalid.
9.4 Latency in backtests (must-have)
In HFT, latency is part of the strategy. A useful model:
• add random jitter to feed-to-decision and decision-to-venue,
• use empirical distributions (P50/P95/P99),
• model occasional spikes (tail latency) because they cause worst losses.
10 Overfitting and Robustness: HFT-Specific Failure Modes
HFT models are easy to overfit because features are abundant and horizons are short.
10
Practical HFT Fundamentals Microstructure, Execution, Production
10.1 Common traps
• Leakage via event ordering: using a future exchange timestamp accidentally.
• Survivorship in symbol/venue mapping: ignoring halts or symbol changes.
• Modeling fills at favorable prices: “filled at touch” without queue.
• Ignoring throttles: a backtest may assume you can cancel infinitely fast.
• Non-stationary microstructure: tick size changes, fee schedule changes, participant
behavior evolves.
10.2 Robust validation practices
• Walk-forward across regimes: low vol vs high vol days.
• Venue perturbation: test sensitivity to fee/rebate and latency changes.
• Feature stability: does imbalance predictive power persist after costs?
• Stress execution: spreads widen 2–3x, depth drops 50%.
Remark 1. A strong interview answer includes: “what would break this strategy?” In HFT, the
biggest breaks are microstructure shifts and execution assumptions.
11 Risk Management in HFT: Real-time Limits and Kill Switches
11.1 Risk dimensions
• Inventory risk: position and delta exposure.
• PnL risk: intraday drawdown, loss rate.
• Operational risk: stale data, desync, stuck orders.
• Compliance/venue risk: order rate limits, self-trade prevention.
11.2 Real-time controls (practical)
• Hard position limits: per symbol and global.
• Order size / price bands: fat-finger protection.
• Order rate throttles: per second, per minute.
• Kill switch: on drawdown, stale data, abnormal rejects/fills.
Listing 2: Example kill-switch policy (conceptual)
if data_stale_ms > 200:
cancel_all () ; di sa bl e_ ne w_ or de rs ( " stale feed " )
if rejects_rate > 0.05:
cancel_all () ; di sa bl e_ ne w_ or de rs ( " venue rejects spike " )
11
Practical HFT Fundamentals Microstructure, Execution, Production
if pnl_drawdown > dd_limit or loss_rate > loss_rate_limit :
reduce_risk ( target_inventory =0)
di sa bl e_ ne w_ or de rs ( " risk breach " )
if s e l f _ t r a d e _ r i s k _ d e t e c t e d () :
enable_stp () ; widen_quotes ()
11.3 Markout-based toxicity circuit breaker
A modern market making desk often monitors rolling markout:
M ∆ = EWMA(Markout(∆)).
If M ∆ turns sharply negative, you reduce quoting or widen.
12 Production Systems: Feed Handlers, OMS, Observability
12.1 Architecture (high level)
Typical components:
• feed handler: parse multicast/unicast, sequence checks, book build,
• strategy engine: feature computation, decision logic, quoting,
• risk engine: pre-trade checks, limits,
• OMS: order state, acks, cancels, reconciliation,
• monitoring: metrics, logs, alerts, replay.
12.2 State consistency and reconciliation
Key production problem: local order state can diverge from exchange state due to:
• packet loss or out-of-order delivery,
• missed acks, disconnect/reconnect,
• exchange resend gaps, sequence resets.
Therefore you need periodic reconciliation (drop-copy, order query, or recovery protocol).
12.3 Observability metrics that matter
• latencies: feed & order RTT (P50/P95/P99),
• fill rate by state (spread, imbalance decile),
• cancel-to-trade ratio,
• markout curves by venue and time-of-day,
12
Practical HFT Fundamentals Microstructure, Execution, Production
• inventory distribution and limit hits,
• reject codes frequency (to catch rule violations).
13 End-to-End Case Study: Large-Tick Market Making with
Queue Modeling
13.1 Setup
Instrument: a large-tick futures contract where spread is usually 1 tick. Venue: maker-taker
fees; strong competition at touch.
13.2 Signals
• imbalance It (best-level depth imbalance),
• microprice deviation µt = (micropricet − mt ),
• short-term trade sign pressure,
• volatility regime flag (EWMA of mid returns).
13.3 Decision logic (realistic)
1. Compute desired quote size as a function of risk budget and volatility.
2. Place bid/ask at touch when toxicity is low; otherwise widen or cancel.
3. Use inventory skew: shift quotes to mean-revert inventory.
4. Use queue-aware logic: if queue ahead is too large, either (i) reduce size, or (ii) step back
(worse price), or (iii) switch to taker hedge elsewhere.
13.4 Backtest requirements
• event-driven LOB replay,
• queue position tracking and churn estimation,
• latency model with jitter and tail spikes,
• fees/rebates and throttles,
• markout curves and risk metrics.
13.5 Failure modes and mitigations
• Regime shift: sudden increase in informed flow → toxicity breaker.
• Latency spike: stale quotes → widen or pull on P99 latency.
• Feed gap: book desync → safe mode + resubscribe + cancel all.
• Crowding: everyone uses similar imbalance signals → edge decays; need novelty or better
execution.
13
Practical HFT Fundamentals Microstructure, Execution, Production
14 Interview Playbook: Questions, Drill-downs, and Strong
Answer Templates
14.1 How interviewers probe HFT candidates
They often start simple (define microprice) then drill into implementation:
• “How do you model fills?” → queue, latency, cancels, hidden liquidity.
• “How do you measure toxicity?” → markout, adverse selection, trade sign.
• “What breaks your backtest?” → leakage, ordering, throttles, optimistic fills.
• “How do you handle tail latency?” → P99 monitoring, safe modes.
14.2 Answer framework that works
Use: Objective → Constraints → Model → Validation → Failure modes. Provide a
small equation and a production detail. That combination signals competence.
14.3 Representative interview questions (with desk-oriented hints)
1. Explain microprice. When is it useful?
Mention large-tick markets, imbalance, short-horizon prediction, and fill constraints.
2. How would you build a fill probability model?
Discuss queue position, hazard/intensity, features, calibration, and validation by deciles.
3. Your market making PnL drops suddenly. Diagnose.
Check markouts, spreads, reject codes, latency spikes, fee changes, regime shifts.
4. How do you incorporate latency into backtests?
Use jitter/tail spikes, decision/ack times, multi-venue asynchrony.
5. What is adverse selection? How do you measure it?
Markout curves; condition on spread/imbalance/time-of-day.
15 Appendix: Practical Checklists
15.1 HFT research checklist
• Is the signal defined with correct timestamps (no leakage)?
• Does it survive realistic fees, rebates, and spread?
• Does it remain after conservative fill simulation?
• Does it survive stress: wider spreads, lower depth, higher latency?
• Are results stable across days/regimes/venues?
14
Practical HFT Fundamentals Microstructure, Execution, Production
15.2 Production readiness checklist
• Data health monitoring (staleness, sequence gaps).
• Kill switches and safe modes tested.
• Replay tooling for incident debugging.
• Limits: position, order rate, price bands, self-trade prevention.
• Observability: P99 latency, rejects, markouts, fill rates.
15