Shadowell is a systematic quantitative trading group. We research, engineer and run
fully automated strategies across global markets, funded entirely by our own balance sheet.
Principal risk. Total alignment. Zero distraction.
We are not an asset manager. There is no fundraising, no investor reporting cycle, no benchmark
to hug. Every decision is measured against a single question: does it make the book better?
01
Our capital, our conviction
The firm's own balance sheet is the only capital at risk. That lets us hold positions
through noise, size with discipline, and kill strategies the moment the edge decays.
02
Evidence over instinct
Every position begins as a falsifiable hypothesis. It reaches production only after it
survives out-of-sample testing, cost modelling and adversarial review by the team.
03
Small team, deep stack
Researchers and engineers sit together and own their work end to end, from raw tick
data to the order on the wire. No silos, no handoffs, no pods competing for scraps.
What we are
A principal trading firm risking only its own balance sheet
A single book, run by one team on one platform
An engineering organisation that happens to trade
Fully systematic, with no discretionary overrides
What we are not
A fund, an asset manager or a broker
A funded-trader or evaluation programme
A signal, course or copy-trading service
A place where research is separated from production
Three disciplines
Research. Engineering. Risk.
They are not departments. They are the three forces every idea has to pass through.
Research
A continuous pipeline that turns raw market and alternative data into tested,
capital-ready signals. Hypothesis, evidence, decision: repeated relentlessly.
Signal discovery
Regime modelling
Cost & capacity
Validation
Engineering
One in-house platform: ingestion, simulation, deployment and execution.
The same code path that backtests a strategy is the one that trades it.
C++ / Rust
Deterministic sim
Co-location
Observability
Risk
Limits are code, not culture. Exposure, drawdown, correlation and liquidity are
enforced automatically: pre-trade, in-flight and at the book level.
Pre-trade limits
Kill switches
Stress scenarios
Attribution
The pipeline
From raw tick to live risk.
Five stages, each one measured, versioned and reproducible,
so that a result from three years ago can be rebuilt today, bit for bit.
01 / 05 Data
Stage 01
Data, captured properly
Full-depth market data, reference data and alternative sources are normalised into a
single time-correct store. Everything is point-in-time,
so a model never sees the future.
Nanosecond-stamped, exchange-native capture
Point-in-time corporate actions and reference data
Automated quality gates on gaps, outliers and drift
Stage 02
Signals, honestly tested
Ideas are expressed as forecasts and judged out of sample, net of realistic costs.
Most die here. That's the point.
Statistical and machine-learned forecasts across horizons
Cost, slippage and capacity modelled from day one
Adversarial review before a single pound is allocated
Stage 03
Portfolio, built for the whole book
Forecasts become positions through an optimiser that balances expected return against
risk, turnover and the correlation of everything already on.
Multi-horizon signal blending and decay control
Factor, sector and liquidity constraints
Turnover-aware sizing against live capacity
Stage 04
Execution, engineered
Target positions are traded by our own smart order routing and scheduling logic,
co-located at the venues that matter, and continuously scored against benchmarks.
In-house routing, scheduling and microstructure models
Co-located, kernel-bypass networking on critical paths
Every fill measured. Execution alpha is real alpha.
Stage 05
Risk, enforced in code
Limits sit in the critical path, not in a policy document. Independent monitoring runs
beside the trading system with the authority to halt it instantly.
Pre-trade limit checks and automated kill switches
Real-time exposure, drawdown and concentration control
Daily attribution back to the originating hypothesis
0+
Execution venues and exchanges connected
0+
Instruments continuously modelled
0M
Orders and quotes per trading day
<0µs
Median tick-to-trade on latency-critical paths
Tick-to-trade
— µs
0 µsp99 —60 µs
Inbound messages
— /s
Rolling windowNormalised
—
Venue sessions
45 / 45
All sessions healthyHeartbeat 1s
Illustrative visualisation · synthetic telemetry keyed to the real London clock, not live production data
Strategy lab
Why most edges never survive.
A toy version of the decision we make every week. Move the sliders: costs, capacity, latency and
the cost of capital will happily eat an edge that looked beautiful on paper.
Predicted return per trade, before any cost.
What the venue and the touch charge you to get in and out.
More trades means more independent bets, and more market impact.
Impact grows with size. This is why edges do not scale for ever.
Slower means more adverse selection on every fill.
Financing the leverage a thin edge needs. Slow strategies feel this hardest.
Scales the book. Changes the size of the swings, not the quality of the edge.
SEED —
·
·
Drop a CSV of prices here
date,close — parsed in your browser, never uploaded.
—
Return / yr
—
Max drawdown
—
Net edge
—
—
—
Net of cost (one path) Other paths, same parameters Gross, cost-free fantasy Drawdown
Adjust the controls to see how an edge behaves once reality is priced in.
Illustrative toy model with a fixed random seed. Not a representation of Shadowell's strategies, positions or performance.
Read: why we kill most of our own ideas ›
Technology
Built here. End to end.
We buy connectivity and compute. Everything that touches a decision — simulation, portfolio,
execution, risk — is ours.
One research platform
A single, deterministic simulator drives research and production. If it happened in
the backtest, it can happen on the wire: same code, same clock, same assumptions.
Core languagesC++20 · Rust · Python
Research computeGPU cluster · distributed sim
DataColumnar tick store · petabyte scale
ReproducibilityVersioned data + code + seeds
Low-latency core
Lock-free, cache-aware C++ on tuned kernels with bypass networking and hardware
timestamping. Latency is budgeted per hop and regression-tested like any other feature.
Thousands of parallel scenarios per idea, with queue position,
latency and impact modelled rather than assumed. Cheap experiments make expensive mistakes rare.
Diversified across asset classes and horizons, because a book that
depends on one regime is not a book. It's a bet.
Asset classes
Equities & ETFsCASH · DELTA-1
Futures & index productsGLOBAL · 24/5
Listed options & volatilitySURFACE · RV
Foreign exchangeSPOT · FWD
Digital assetsSPOT · PERP · OPT
Strategy families
Statistical arbitrageINTRADAY → DAYS
Market microstructureSUB-SECOND
Volatility relative valueDAYS → WEEKS
Systematic macro & trendWEEKS → MONTHS
Execution alphaCONTINUOUS
SYNTHETIC BOOK · SWL-PERP · L2
Tape
—
Simulated microstructure: clustered order arrivals, cancellations and executions. No real venue data.
Family correlation
Why we care
A duplicate signal is a concentration wearing a diversifier's clothes. Every new strategy
is priced against the correlation of everything already on the book, not against its own backtest.
Select a cell to read the pair.
Illustrative long-run correlations between strategy families, not live figures.
Session clock
—
Indicative primary session hours in each venue's local time · excludes holidays, auctions and after-hours
Points of presence
Pick a venue to see the great-circle distance from LD4 and what the speed of light in fibre
actually permits. Physics sets the floor; routing sets the price.
Indicative infrastructure footprint · drag the globe to rotate
How we work
Five things we refuse to bend on.
01
Truth beats comfort
The fastest way to lose money is to fall in love with your own backtest. We attack our
best ideas hardest, and we celebrate the ones we kill.
02
One book, one team
No pods, no internal competition for capital, no hoarding of signals. Everything is
allocated against the same risk budget, in the open.
03
Limits are code
A constraint that isn't enforced in the critical path isn't a constraint. It's a hope.
Risk sits in front of every order, automatically.
04
Reproducible or it didn't happen
Data, code and seeds are versioned together. A result from three years ago rebuilds
today, bit for bit, or we don't trust it.
05
Small, trusted, unblocked
We would rather stay deliberately small and give everyone the compute, the data and the
authority to finish what they start.
Milestones
A short history.
2024
Incorporated
One researcher, one venue, one strategy, and a rule that everything must be reproducible.
2024
First systematic book
Manual research replaced by a scheduled pipeline. Every position traceable to a hypothesis.
2025
Simulator rebuilt
Vendor backtesting replaced by our own deterministic simulator. Research and production converge.
2025
Cross-asset
Futures and FX join equities on the book. Cost and capacity modelling formalised.
2026
Digital assets
24/7 perpetuals: new microstructure, same discipline, same limits engine.
2026
Bradford desk
A permanent home on Manningham Lane, and the first early-career intake.
Today
One book, five families
A small team, an uncapped research budget and a single question: does it make the book better?
Careers
Bring a hard problem. We'll bring the data.
We hire for reasoning, not résumés. Most of our roles are written for people leaving school,
college or a gap year, because raw horsepower is not a function of age.
The right to work in the UK, and to be in Bradford
Every role is paid, including the insight programme. We do not run unpaid internships,
"evaluations", funded-trader schemes, or anything that asks you for money. Applications are read
by a practitioner, not by a filter, and you will hear back either way.
Why we hire at eighteen ›
You will write software that touches a live trading book. Not a training sandbox with
fake data: the actual research tooling, the actual data jobs, the actual dashboards the desk
looks at before it takes risk. You will be reviewed line by line, and you will ship in week one.
What you'd actually do
Write Python that cleans, joins and validates market data
The job is to try to break ideas, including your own. You'll take a hypothesis about
how a market behaves, turn it into something measurable, test it honestly against real data,
and then say out loud whether it survived. Most will not. That is the work, not a failure of it.
What you'd actually do
Turn a vague idea into a precise, falsifiable statement
Run and document out-of-sample tests in our simulator
Price the cost stack: spread, fees, impact, latency, capacity
Present a kill-or-keep recommendation with the evidence attached
Watch live strategies and flag when behaviour drifts from the model
What we look for
Genuinely strong A-level Maths or Further Maths; this one is mathematical
Instinct for when a number is too good to be true
Willing to learn Python quickly (we will teach you properly)
You can hold two contradictory explanations in your head and design a test between them
Bonus: statistics competitions, olympiads, an EPQ with real data in it
Everything upstream of a trade depends on the data being right, complete and correctly
timed. This role owns that. It is the least glamorous job in the building and the one with the
most leverage: nobody notices good data, and everybody notices bad data eventually.
What you'd actually do
Capture, store and validate market and reference data
Write the checks that catch gaps, duplicates, stale prices and silent schema changes
Enforce point-in-time correctness so no model can see the future
Build scrapers and loaders for new datasets we want to evaluate
Automate your own job away, then move up the stack
What we look for
Methodical to the point of being slightly annoying about it
Some Python; curiosity about SQL, file formats and time zones
Any quantitative A-levels: Maths, Physics, Computer Science, Economics
You would rather ask a dumb question than ship a wrong number
Systematic does not mean incurious. This role turns messy, human information — releases,
policy language, positioning, news flow — into structured data a machine can actually use, and
then tests whether it carries any signal at all.
What you'd actually do
Build and maintain the economic release and event calendar the systems trade around
Turn text and announcements into structured, timestamped datasets
Test whether sentiment and positioning measures predict anything, honestly
Write short, precise internal notes that other people rely on
Track regime shifts and tell the desk when the world has changed
What we look for
Economics, History, Politics, Psychology, Maths: any strong A-level mix
You read widely and can summarise something complicated in three sentences
Numerate enough to be sceptical of your own narrative
Willing to learn Python and spreadsheets to a serious standard
The engine room. Positions must match, cash must match, sessions must be up, and if any
of that is wrong the firm needs to know within minutes rather than at the end of the month. This
is the role where reliability is the skill, and it is the fastest way to understand how the whole
machine actually fits together.
What you'd actually do
Run the daily open and close: sessions, connectivity, position and cash reconciliation
Investigate breaks to root cause and write down what you found
Check fees, financing and statements against what we expected to pay
Turn every manual check you inherit into an automated one
Sit next to the risk monitors and escalate without hesitating
What we look for
Relentlessly organised; you notice when a number is one penny out
Any A-levels; Chemistry, Biology and Maths types thrive here
Calm under a deadline and clear in writing
Comfortable with spreadsheets today, with scripts within a term
If you are still at school or college and want to know whether this world is for you,
start here. One paid week in the Easter holidays, or four paid weeks over the summer. You will
not shadow anyone or fetch anything: you will be given a small, real problem and the tools to
solve it, and you will present what you found on the last day.
What the week looks like
Day 1 — how a systematic trading firm actually works, end to end
Day 2 — Python and market data from zero, properly taught
Day 3 — your problem: a small, real question with real data
Day 4 — costs, risk and why most ideas do not survive
Day 5 — present your findings and get honest feedback
Who it's for
Year 12, Year 13 or on a gap year; any quantitative subject mix
No coding experience needed for the insight week
Curiosity you can demonstrate in a paragraph
Strong performers are invited straight to the full process
Deliberately short and deliberately technical.
Most candidates go from first contact to decision in two to three weeks, and everyone gets a reply.
The take-home and its rubric are published ›
01
Application review
Read by a practitioner. No keyword filter, no algorithm.
02
Intro conversation
20 minutes. Your work, our work, mutual fit.
03
Take-home with real data
Two to three hours, and you keep the code.
04
Technical discussion
We pull your work apart together, as colleagues.
05
Offer & start date
Clear terms, clear pay, no games.
Your first 90 days
Nobody is left to "find their feet". The first quarter is
structured in four-week blocks, so you always know what finished looks like.
WEEK 1–2
Onboard for real
Environment, git, the simulator, the data model, and your first merged change.
WEEK 3–6
First owned project
Scoped, shippable, yours. Reviewed weekly, in the open.
WEEK 7–10
Second project, harder
Closer to the book. You choose the approach and defend it.
WEEK 11–13
Present to the desk
What you built, what you got wrong, what you would do next.
Notes
Things we've written down.
Occasional writing from the desk. No market calls, no
predictions, nothing for sale. Just craft.
No. Shadowell trades exclusively with its own capital. We do not accept investors,
manage client accounts, offer funded-trader programmes or provide investment advice. If someone
contacts you claiming otherwise in our name, it is not us: see our
regulatory disclosure.
Yes. That is who most of these roles are written for. We do not filter on degrees,
universities or prior finance experience, because none of them predict the thing we need:
careful reasoning under uncertainty. If you can show us something you built or proved, that counts
for more than any line on a CV. Here is the longer argument.
Every one of them, including the insight week. We do not run unpaid internships,
unpaid trials, "evaluations" or anything that asks you for money at any point. If you are asked
to pay for something in our name, it is fraud. Tell us.
For the developer and data roles, you need to have written something real, however
scrappy. For the quant, macro and operations roles you do not need to code today, but you do need
to want to, because within a term you will be writing Python daily. We teach it properly rather
than hoping you pick it up.
Not at present. You need the existing right to work in the UK and to be able to get to
Bradford. We would rather say that plainly than waste your time.
From sub-second microstructure strategies through to multi-week systematic macro.
Diversification across horizons is deliberate: different regimes reward different clocks.
It is a deliberately simple teaching model: a linear cost stack, a Gaussian noise
process with occasional fat tails, a breadth haircut for correlated trades, an impact term that
grows with allocated capital, a financing charge on implied leverage, and a fixed seed you can
share. The intuition is real. Costs, impact, latency and the cost of capital dominate small
edges, and one path is never the whole story. The numbers are not ours and are not performance
figures of any kind. The full formulae are in the
colophon.
Research, portfolio construction, execution and risk are all developed in-house. We
license market data and connectivity, and evaluate third-party datasets. Never third-party alpha.
Bradford, West Yorkshire, with trading infrastructure co-located at major venues
globally. Roles are in person: at this size, sitting together is not a preference, it is the method.
No. There are no cookies, no analytics, no advertising pixels and no third-party
requests of any kind, not even a web font. Your appearance preference, sound setting and any
unsent form draft are stored on your own device and never sent anywhere. Details in the
cookie policy.
Yes, and we'd encourage it. Send a short note on something you've built or proved to
careers@shadowell.com. One page beats ten.
Contact
Let's talk.
Whether you're a sixth-former with a side project, an engineer who
loves hard constraints, or a venue or data partner, we read everything and we reply.
Partner conversations are covered by a short, mutual NDA that we
publish rather than negotiate.
Read the NDA policy ›
Shadowell
203 Manningham Lane
Bradford, West Yorkshire, BD8 7HP
United Kingdom
REGISTERED IN ENGLAND & WALES · COMPANY NO. 15776042
↑↓ navigate↵ openesc close? shortcuts
Keyboard shortcuts
⌘KCommand palette — also does maths, definitions and clocks
/Search sections, roles and documents
JKNext / previous section
TCycle appearance — auto, light, dark
SInterface sound on / off
GBack to the top
?This panel
escClose anything that is open
There is one more. Five letters — the first thing a quant looks for.
Desktop shortcuts · the site is fully operable by keyboard
Why we kill most of our own ideas
A research process is not a factory. It is a filter, and the value of a filter is measured
by what it stops, not by what it lets through.
New researchers arrive expecting the job to be invention. It is mostly rejection. In any given month the
interesting question is rarely can we find something? Ideas are cheap and the world is full of
patterns. The question is which of these survives contact with cost, capacity, decay and our own
wishful thinking?
The arithmetic nobody puts on a slide
Suppose you are a good researcher, and one idea in twenty that you find plausible is genuinely tradable
after costs. That is your prior: 5%. Now suppose you run a reasonable test that flags a real edge 60% of
the time, and falsely flags a dead one 5% of the time. An idea passes your test. How confident should you
be?
Your test passed, and the idea is still more likely to be worthless than valuable. That is not
pessimism, it is Bayes. It is also why "the backtest looks great" begins the conversation rather than
ending it, and why we care so much about the base rate: the only ways to improve that 39% are to have
better ideas going in, or a harsher test.
It gets worse if you are careless. Every extra variant you try — a different lookback, a different
universe, a different threshold — is another lottery ticket in a raffle where noise wins prizes. Twenty
variants at a 5% false-positive rate and you should expect one beautiful chart from pure chance. Not
because you cheated. Because you looked.
If you cannot say who is paying you and why they are content to, you do not have
an edge. You have a coincidence with a chart.The house rule
Four ways an edge dies
Almost every rejection we make falls into one of four buckets. Naming them is useful, because each has a
different test.
It was never there. Multiple comparisons, subtle leakage, a universe chosen with
hindsight, a result that lives entirely in one year or one name. The test: does it survive a strictly
held-out period, a shifted universe, and someone actively trying to break it?
It is there, but you cannot reach it. The signal is real and smaller than the
spread. Fees, impact and adverse selection consume
it. The test: model the full cost stack before you get attached, which is exactly what the
strategy lab on this site caricatures.
It was there, and it has gone. Edges decay. Someone else found it, the venue
changed its fee schedule, the participant you were being paid by went away. The test: is the
performance stable through time, or is it all pre-2019?
It is there, and you cannot hold it. Capacity is a few hundred thousand pounds.
Borrow is unavailable. Margin is punishing. The drawdown
profile would make you switch it off at exactly the wrong moment. The test: what does this look like
at ten times the size, and would we actually sit through the bad quarter?
The kill list
Before anything is allocated capital it has to answer these, in order, in writing:
What is the economic mechanism? Who is on the other side, and why are they content to be there?
How many variants were tried before this one? Show the whole search, not the winner.
What is the performance out of sample, net of modelled cost, with realistic fills?
What is the capacity, and where does the edge halve as size grows?
How correlated is it with what we already run? A duplicate signal is a concentration, not a
diversifier.
What would have to be true for this to be a coincidence, and what test would show that?
What is the kill criterion, defined in advance, in numbers?
If this were someone else's idea, what is the first thing you would attack?
That last question is the important one. It is much easier to be rigorous about other people's work,
so we make it structurally somebody's job to be the opponent.
Why a kill is a good day
The two errors are not symmetric. A false negative costs you one idea out of an unlimited queue of
future ideas. A false positive costs capital, atten...
print(f"UK equity: {uk_equity:.1f} bps") # UK equity: 62.6 bpsprint(f"Future: {future:.1f} bps") # Future: 12.6 bps# Annualised, at one round trip per day:print(f"{future * 252 / 100:.1f}% of notional per year") # 31.8%
The four questions we ask about any cost model
Is every component measured from our own fills, or copied from a vendor's assumption?
Does any component vary with the calendar, the size, or the time of day? If so, is it modelled that
way, or as an average?
What happens to each component at ten times the size?
Which single component, if it were 50% worse than modelled, would kill the strategy?
That fourth question is the one that saves money. There is almost always exactly one component the
whole thing hangs on, and it is almost never the one people spend their week refining.
Latency is not a number you quote. It is a budget you allocate, a distribution you own,
and a regression test you never let go red.
People new to low-latency work ask "how fast is your system?" and expect a single figure. The honest
answer is a question back: fast at what, measured where, at which percentile, under what load, racing
whom? Every one of those qualifiers changes the answer by an order of magnitude.
Latency is only meaningful relative to someone
There is no absolute standard of fast. There is only whether you reach the order book before the
participant whose flow you want to trade against, or after the one trying to trade against yours. That
reframing matters, because it means the correct amount of engineering effort is finite and knowable. On
some strategies, being in the top handful of participants is the entire business. On others, an extra
microsecond is worth nothing, and the same week spent on the cost model is worth a great deal.
The hops
We keep a written budget for every path that matters, and every hop has an owner and a target. The
shape of it looks like this. Numbers are illustrative; the discipline is the point.
Hop
Budget
What actually goes wrong
Venue → your rack
physics
Cross-connect routing, patch panels, the fibre you did not measure
NIC → user space
~1 µs
Kernel path instead of bypass; interrupt coalescing left on
Feed decode
0.3–2 µs
Allocation on the hot path; branch-heavy parsing; unaligned reads
Book build
0.2–1.5 µs
Cache misses in the price-level structure; map lookups
Signal evaluation
0.3–2 µs
Floating-point stalls; work that could have been precomputed
Pre-trade risk
0.2–1 µs
Locks, logging, anything that touches the heap
Encode → gateway
0.5–2 µs
String formatting; copying; a syscall you forgot about
Wire → matching engine
venue
Session throttles, sequencing, gateway queueing you cannot see
The hop everybody forgets is not in your code at all: the queue you are already in. A
strategy that rests an order early and holds
priority can beat a system two microseconds faster
that arrives second. Queue position is latency by another name, and it is bought with modelling rather
than hardware.
p50 is vanity. The tail pays the bills.
Median latency is the number that goes in decks. It is close to useless, because the moments that
matter are precisely the moments when the system is under stress: a burst of messages, a volatility
spike, an economic release. That is when information is most valuable and when your p99.9 shows up.
Figure 1 — the same system, quiet and burstingtoggle the load
A quiet system shows a tight distribution around 11 µs. Under a message
burst the median barely moves while the 99.9th percentile grows by two orders of magnitude. Enable
JavaScript to explore the distribution interactively.
So we track distributions, not averages, and we look at latency conditioned on message rate. A system
with a 12 µs median and a 900 µs tail under burst is not a 12 µs system. It is a 900 µs system that
flatters itself when nothing is happening.
A predictable 8 µs beats a 5 µs average with a fat tail. You can build a strategy
on a promise. You cannot build one on a hope.Core systems
Determinism beats speed
Jitter is the real enemy. The usual suspects, roughly in order of how often they are the culprit:
Allocation and garbage on the hot path; anything that can pause you will, at the worst moment
Page faults; memory not pre-touched or locked
Interrupts and kernel threads sharing your core; C-state transitions on an idle core
NUMA crossings; a thread pinned to the wrong socket for the NIC it serves
Shared L3 with a noisy neighbour process, including your own logging
Logging and metrics doing synchronous I/O because it was convenient in development
Measure where the wire is, not where the code is
Instrumenting your own code tells you what your code thinks happened. Hardware timestamps at the NIC
tell you what happened. If those two disagree, the NIC is right. We take timestamps at capture, at
decode, at decision and at send, and we keep the deltas as first-class metrics.
Then we gate it. A pull request that adds 200 nanoseconds to the critical path fails the build, exactly
like one that breaks a unit test. Performance that is not defended in CI decays quietly, one
reasonable-looking commit at a time.
c++// The whole regression gate, more or less. Run it in CI, fail the build.staticvoidbench_decode(benchmark::State& st) {
constauto frames = load_golden_capture("xetra_2025-11-04.pcap");
BookBuilder book; // pre-touched, pool-allocatedfor (auto _ : st) {
for (constauto& f : frames)
benchmark::DoNotOptimize(book.apply(f));
}
st.counters["ns_per_msg"] = benchmark::Counter(
st.iterations() * frames.size(), benchmark::Counter::kIsRate |
benchmark::Counter::kInvert);
}
// budget.yaml — checked in, reviewed like code// bench_decode.ns_per_msg: { target: 340, ceiling: 380 }
Where people waste their year
Optimising the wrong hop. Shaving 300 ns off decode while sitting behind 40 µs of
avoidable gateway queueing. Profile first, always.
Buying speed you cannot use. If the edge decays over minutes, microseconds are a
vanity purchase.
Forgetting the alternative. The cheapest latency improvement is often a better
model: predicting one tick further ahead beats a faster path to the wrong decision.
The fastest system in the world still loses to a system that knows when not to trade.
Point-in-time, and the futures you accidentally see
Leakage never announces itself. It does not look like a bug. It looks like a good result.
Every dataset used in research answers a question about the past. The trouble is that most datasets are
maintained in the present, by people who quite reasonably want them to be correct rather than
historically faithful. A price is adjusted. A statement is restated. A company is
reclassified. An index constituent list is updated. Each improvement silently posts a letter from the
future into your training set.
Every row needs two clocks
Our rule is simple and non-negotiable: every record carries an event time (when the
thing happened in the world) and a knowledge time (the first moment we could have known
it). Research reads through an as-of view that refuses to return rows whose knowledge time is later than
the simulated clock.
The consequence is that leakage stops being a matter of researcher discipline and becomes a property of
the access layer. You cannot see the future because the API will not serve it to you. Discipline is a
fine thing; not needing to rely on it is better.
pythonclassAsOfStore:
"""Reads are impossible to get wrong: the clock is a constructor argument."""def__init__(self, conn, as_of):
self._conn = conn
self._as_of = as_of # the simulated 'now'deffundamentals(self, tickers):
return self._conn.query(
"""SELECT * FROM fundamentals
WHERE ticker = ANY(%(t)s)
AND knowledge_time <= %(now)s -- not event_time. never event_time.
QUALIFY ROW_NUMBER() OVER (
PARTITION BY ticker, period
ORDER BY knowledge_time DESC) = 1""",
{"t": list(tickers), "now": self._as_of})
# There is deliberately no way to construct this without a clock.defopen_store(conn, as_of):
if as_of isNone:
raiseValueError("as_of is required; 'latest' is not a research concept")
returnAsOfStore(conn, as_of)
A field guide to leakage
Survivorship. Your universe contains the companies that still exist. The ones that
went to zero are the entire point.
Restatement. Fundamentals as they are now, not as first reported. You are trading
on an audited truth nobody had.
Index membership. Using today's constituents historically buys everything after it
has already run.
Timestamps. The vendor's stamp, the exchange's stamp and your capture stamp are
three different clocks. Only one is the one you could have acted on.
Adjusted prices. A split-adjusted series is rewritten history. Correct for
charting, dangerous for simulation.
Universe selection. "Liquid names" chosen using the full sample is hindsight with
extra steps. This is the one that got us; see
exhibit B in the graveyard.
Reference data. Sector and country reclassifications propagate backwards in most
vendor files.
The human leak. You lived through the period. You know what happened in March.
Your feature choices know it too.
Figure 1 — find the leakageone cell is a letter from the future
A small dataset is shown below with prices, reported earnings, split
ratios and knowledge timestamps. Exactly one cell could not have been known on the date of its own
row. Enable JavaScript to play the interactive version; the answer is that the split ratio for
2025-03-14 was applied to prices before the split was announced on 2025-03-19.
Tests that actually catch it
The shift test. Delay every input by one full period. A real edge degrades
gracefully. Leakage falls off a cliff, because you removed the thing it was using.
The as-of rebuild. Reconstruct the world as of an arbitrary past date and check the
numbers match what was on the screen that day. If you cannot, you do not have a point-in-time store,
you have a spreadsheet with good intentions.
The paranoid harness. Make future access a type error. A simulator that physically
cannot deliver tomorrow's row removes the entire class of mistake.
Label shuffling. Randomise the target. Any remaining performance is your pipeline
cheating.
Blind CI. A scheduled job re-runs yesterday's research against the store as it
stood yesterday. Silent drift shows up as a failing test, not as a mystery in six months.
Why this is research, not chores
There is a tempting hierarchy where data work is preparation and modelling is the real job. It is
backwards. Model choice is usually worth a few per cent; a leaked feature is worth an infinite amount of
imaginary money, right up until it costs real money. The most valuable engineering hour in a
quantitative firm is usually spent making it impossible to lie to yourself.
Data quality is not the thing you do before research. It is research.
Why this website has zero dependencies
Open the network tab. One document, no fonts, no CDN, no analytics, nothing from anyone
else. That is not minimalism as an aesthetic. It is the same argument we make about the trading path.
The default way to build a site like this is a framework, a component library, a font service, an
analytics tag, a consent tool to apologise for the analytics tag, and somewhere north of forty
third-party requests. Each one is a decision somebody else gets to make about your page, at runtime,
for ever.
Owning the critical path
In the trading system we do not put code we did not write between a market data packet and a decision.
Not because other people's code is bad, but because we cannot reason about latency, failure modes or
determinism through a dependency we do not control. The reasoning transfers exactly:
Supply chain. A script tag pointing at someone else's domain is a standing grant of
arbitrary code execution on our page. The recruiting site of a trading firm is a plausible target.
We removed the attack surface rather than monitoring it.
Latency. Every third-party origin is a DNS lookup, a TCP handshake, a TLS
negotiation and a queue you do not control, on a connection you cannot warm.
Determinism. The page renders the same way on every visit because nothing about it
can change without us shipping a change.
Honesty. The cookie policy says there are no third-party requests. That sentence is
only true if it is structurally true.
A dependency is a promise made by someone with no obligation to keep
it.Core systems
What we gave up, and what it cost
Genuinely: some things are harder. There is no component library, so the accordion, the command palette,
the reader, the charts and the globe are all hand-written. That is perhaps two thousand lines of
JavaScript, once, and it will not need a migration.
There is no font service, so the type stack resolves to whatever the operating system provides. On
Apple platforms that is SF Pro, on Windows Segoe UI Variable, on Android Roboto. The tradeoff is that
the page looks very slightly different on each platform. In exchange, text renders on the first paint
with no flash, no layout shift and no request.
What is still a build step
Zero runtime dependencies is not the same as zero tooling. At build time we do four things, all
of which produce plain static output:
Syntax highlighting. The code on this page is highlighted into spans at build time, so the browser
ships no highlighter. The one you are reading cost you nothing.
The share image. og.png is rendered headlessly from the same hero canvas code, with the
same seed, so the card matches the page.
Minification, and a content hash for the service worker's cache name.
A link check and a structured-data validation pass.
The measurements
Published, with the methodology, in the colophon. Briefly: one HTML
document, one request, no blocking subresources, and a Lighthouse performance score we are willing to
have argued with.
The uncomfortable part
This approach has a real cost, and it is not technical. It is that a new person cannot arrive knowing
the stack, because the stack is ours. So the code is written to be read: sectioned, commented at the
level of intent rather than mechanism, and defensive about failure. Every interactive feature is wrapped
so that if it throws, the rest of the page survives and the content is still there. There is a watchdog
that forces every hidden element visible if the main script never finishes booting.
javascript// Any single feature failing must never take the page with it.constsafe = (label, fn) => {
try { fn(); }
catch (e) { if (window.console) console.warn('[shadowell] ' + label, e); }
};
safe('globe', initGlobe);
safe('order book', initLimitOrderBook);
safe('lab', initStrategyLab);
// …and the belt to go with the braces:addEventListener('error', () => document.documentElement.classList.add('js-broken'), true);
If you find that reasoning persuasive, you will probably enjoy working here.
The open roles are here.
Written in a text editor. Deployed as one file. Verified in the network tab.
Hiring at eighteen
The market prices credentials. We would rather price reasoning, and those two things are
much further apart than the industry likes to admit.
The conventional path into systematic trading is narrow and expensive: a specific handful of
universities, a specific handful of degrees, a specific summer internship taken at a specific age. It is
a filter that works, in the sense that it selects capable people. It is also a filter that
systematically ignores enormous quantities of raw ability for reasons that have nothing to do with
ability.
What an eighteen-year-old actually brings
No bad habits. Nobody has taught them an outdated theory they will spend two years
unlearning, or a workflow that only makes sense inside a large bank.
Time on task. The best predictor of getting good at this is hours spent
deliberately confused. Eighteen-year-olds who like problems have those hours and spend them freely.
A tolerance for being wrong. Not universally, but far more often than people whose
professional identity has hardened.
Curiosity about mechanism. The instinct to ask "but why does that work?" is not
something we can install later.
What they do not bring is context. They have never worked in an office, never used version control on
anything anyone depended on, and never had to explain a mistake that cost money. That is our problem to
solve, not theirs, and it is solved with structure rather than with hope.
Grades are evidence of sustained effort and mathematical fluency. That is all they
are, and it is not nothing.Hiring
So we test the thing we need
We do not run brain-teasers as hazing. We do not care whether you can invert a binary tree on a
whiteboard. We hand you a small, real problem with real data, give you a couple of hours, and then sit
down and pull the result apart with you the way we would with each other. That conversation tells us
almost everything: how you decide what to check, what you do when a number surprises you, whether you
can say "I don't know" and then say what you'd do about it.
The brief and the rubric are published in full, because hiding the
criteria only advantages candidates who already know somebody who has been through it.
What the structure looks like
Freedom without scaffolding is just being lost. The first quarter is deliberately shaped: a sandbox you
cannot break, projects scoped to four weeks so that "finished" is unambiguous, code review from day one,
and a standing instruction to ask early rather than struggle silently for a week. By week thirteen you
present what you built to the desk, including the part you got wrong.
What we owe you, and what we ask
We owe you: pay from day one, real work on the live platform, direct mentoring from the people running
the book, honest feedback within a week including when the answer is no, and something concrete you can
point at afterwards, whether you stay or not.
We ask: be in Bradford, in person, most days. Be relentless about being wrong. And bring us something
you built, however small and however ugly.
A note about Bradford
We are not the first to notice that this city produces serious mathematicians, engineers and
programmers. We might be among the first to build a systematic trading platform with them rather than
waiting for them to leave. That is not charity. It is the most obvious mispricing we have found, and we
intend to keep trading it.
Why publish this? Because a hidden rubric is a tax on people without contacts. Everything
below is the actual brief and the actual mark sheet. Reading it in advance is not cheating; it is the
point. We change the dataset, never the standard.
1. The brief you receive
You get a CSV of roughly two years of daily data for forty instruments, a one-page description of the
columns, and this task:
Column x is a candidate predictor. Column fwd_1d is the next day's return.
Decide whether x contains a tradable signal, assume a 6 basis point round-trip cost, and
tell us what you would do next. Two to three hours. Send us whatever you have at the end of it,
finished or not.
What you send: your code, and no more than one page of prose. No slides. No formatting effort. We would
rather have a scruffy notebook with a sharp conclusion than a beautiful report with a vague one.
2. What is deliberately in the data
Three things, and we are telling you they are there:
The raw correlation is real but is concentrated in a handful of the least liquid instruments.
Costs, applied honestly at the turnover the naive version implies, remove most of the edge.
One column has a subtle timing problem. It is findable in twenty minutes if you look, and invisible
for ever if you do not.
None of these is a trap. They are the three things this job actually consists of.
3. The rubric
Thirty points. We mark independently, then compare, and the discussion of any disagreement is itself
informative.
Criterion
Pts
What full marks looks like
Framing
4
You stated a falsifiable claim before testing anything, and said what would change your mind.
Correct evaluation
5
Out-of-sample or walk-forward, no peeking, and you say explicitly which rows the model saw.
Cost treatment
5
Costs applied to the turnover your own signal implies, not to a convenient proxy.
Robustness
4
You checked concentration by instrument and by period, and reported it even though it weakened your case.
Finding the flaw
4
You found the timing problem, or you designed a test that would have found it and said so.
Honest conclusion
4
A clear kill-or-keep with the reason, and a stated confidence you actually hold.
Code you can defend
2
Runs from a clean checkout, deterministic, readable. Not pretty. Reproducible.
Communication
2
One page, no filler, the number that matters in the first three sentences.
Twenty and above goes to the technical round. Fifteen to nineteen usually does too, if the reasoning is
strong and the gap is something teachable. Below that we write to you and tell you specifically which
criteria you lost points on, because a rejection with no information is a waste of your two hours.
4. What we do not mark
Model sophistication. A linear regression that is correctly evaluated beats a gradient-boosted
machine that is not, every single time.
Whether the answer is yes. Roughly a third of strong submissions correctly conclude there is
nothing here. That is a full-marks answer.
Finishing. Most people do not finish. We mark the quality of the reasoning per hour spent.
Tooling. Python, R, Julia, Excel with visible workings. We care that you can defend it.
5. The four ways people lose the most points
Applying costs as a flat annual haircut rather than per trade at their own turnover.
Tuning on the test set and mentioning it in passing, as though naming it neutralises it.
Reporting a t-statistic without saying how many variants were tried to get it.
A confident conclusion with no stated uncertainty. Confidence without an error bar is a guess in a
suit.
6. Practicalities
You keep the code and may publish it. We do not use candidate work.
Take it whenever suits; there is no timer running on the file.
Using an LLM is fine, and please tell us where. We will ask you about the parts it wrote, and
"I don't know, it generated that" is a bad answer whoever wrote it.
If two to three hours is genuinely impossible for you, say so and we will find another way to
assess the same thing.
How this works. Solve it, email us the answer with a paragraph of working, and you skip the
application review and the intro call. You go straight to the technical discussion. No CV required.
No degree required. It is genuinely hard, and it is genuinely self-marking: there is one number, and
you will know when you have it, because it makes the last part come out clean.
The setup
A venue runs a continuous double auction. You are a market maker quoting one instrument. In any given
second, exactly one of three things happens:
With probability 0.02, an informed trader arrives. They always trade in the
direction of the true value, and immediately after they trade, the mid moves 8 basis points against
you.
With probability 0.18, an uninformed trader arrives. They buy or sell with
equal probability and the mid does not move.
Otherwise nothing happens.
You quote a symmetric spread of s basis points around the mid, so you earn s/2 on each
fill. You have infinite capacity and no inventory limit. Every arriving trader trades with you.
Part one
What is the minimum spread s at which you break even, per second, in expectation?
Part two
Now the venue introduces a queue. There are n other market makers quoting the
same spread. Arriving orders fill the earliest-posted quote first. You are, on average, in the middle of
the queue, so you receive a fraction 1/(n+1) of uninformed flow. Informed traders, however, are
willing to sweep, so they take enough size to reach you every time. With n = 4, what is your
new break-even spread?
Part three
Finally: you can pay for co-location, which puts you at the front of the queue for uninformed
flow (you get all of it) but does nothing about informed sweeps. The venue charges a flat fee, quoted in
basis points of the notional you trade. With n = 4 and your Part-two spread held fixed, what is
the maximum fee, in basis points, at which co-location is worth buying?
What to send
Email careers@shadowell.com
with the subject line Puzzle set 04 — fast track and include:
Three numbers, to two decimal places.
A paragraph of working. Not a derivation in LaTeX. The reasoning, in sentences.
One line on where the question was ambiguous and what you assumed. This is marked.
Two sentences about who you are. That is the whole application.
Ground rules, plainly
Use anything you like: paper, Python, a simulation, an LLM. We will ask you to reason about it live,
so understand what you sent.
Wrong-but-well-reasoned regularly beats right-but-unexplained. We are buying the reasoning.
Please do not publish solutions. We rotate the set roughly every six months and we would rather
keep this route open.
If you get stuck at part three, send parts one and two. Partial answers with good reasoning still
get read by a human, and still get a reply.
A hint, because we would rather you learned something
Part three is the interesting one, and it is not really about latency. Write down your expected
profit per second in both worlds as a function of the fee, and notice that the fee applies to
all your flow while the benefit applies only to some of it. The answer is smaller than
most people's first guess. That asymmetry is the entire reason
queue position is a strategic decision rather than an
engineering one, and it is exactly the calculation we ran on
exhibit C.
If you want the long version of the underlying economics, read
the cost of the touch first.
How to pass the Shadowell technical interview
This is not a trick. We would rather interview a well-prepared version of you than a surprised
one. Everything we assess is written down here and in the
published rubric.
The shape of the conversation
Sixty to ninety minutes, two of us, no whiteboard, no timer. It is a discussion of the take-home you
already sent, plus two or three follow-on questions that grow out of it. There is no separate coding
round and no puzzle round.
Ten minutes. You walk us through what you did and what you concluded. Lead with the
conclusion; we will ask for the path.
Thirty minutes. We attack it. Not you — it. Expect "what if the costs were double",
"what would this look like on twenty times the capital", "how do you know that isn't leakage".
Twenty minutes. A new, small problem, out loud, together. You will not have seen it.
We are watching how you decompose something unfamiliar, not whether you finish.
The rest. Your questions. Ask hard ones. We will answer them.
What we are actually measuring
Do you notice when you are confused? The single strongest signal. Candidates who
say "hold on, that doesn't follow" mid-sentence do very well here.
Can you be wrong gracefully and fast? We will present a counter-argument. Updating
is good. Digging in with a better argument is also good. Agreeing to be agreeable is fatal.
Do you distinguish what you measured from what you assumed? Most people blur these.
The ones who do not are the ones we hire.
Can you explain it to a non-specialist? At some point one of us will play dumb on
purpose.
How to prepare, in rough order of value
Re-read your own submission and find its weakest claim. Come prepared to attack it
yourself. This is worth more than any revision.
Be able to define, in one plain sentence each:Sharpe ratio,
drawdown,
basis point,
slippage,
adverse selection,
point-in-time,
capacity. Every term on this site has a definition
attached; hover the dotted ones.
Know why out-of-sample testing exists, not just that it does.
Do the arithmetic of a spread once, by hand. Our
cost of the touch note walks through it in pounds.
Play with the strategy lab for ten minutes. Find a
setting where a large edge still loses money and be ready to say why.
Read one thing properly rather than five things badly. Any of the notes on this
site will do.
Things that will not help
Memorising interview question banks. We do not use them.
Buzzwords. Saying "regime-aware ensemble" without being able to define it costs you more than
saying nothing.
Pretending to know a market. "I've never traded, here's how I'd find out" is a strong answer.
A long CV. We read one page, and the link to your work first.
Five questions you will probably be asked
What is the single result here you are least confident in, and why?
Suppose I tell you this edge disappears next month. What is your first guess as to the reason?
How would you tell whether this is one signal or four correlated ones?
What did you not have time to check?
Teach me something you learned this year that you found genuinely surprising.
Practicalities
Dress however you like. Genuinely.
Bring your laptop with the code open. We may ask you to run something.
Video is fine if travel is hard. Say so; it costs you nothing.
If you need an adjustment of any kind — extra time, a written question in advance, a break, a quiet
room — email careers@shadowell.com and it is arranged, no
explanation required.
You get a decision within five working days, with reasons.
Still nervous? Everyone is. Bring the nervousness and the working.
For teachers & careers advisers
If you teach maths, computing, physics or economics to sixteen to nineteen year olds, this page is for
you. Everything here is free, needs no registration, and may be copied, printed and adapted without
asking us.
The one-sentence version for your students. A quantitative trading firm in Bradford hires
people straight from school and college, pays them from day one, and does not require a degree.
If you would rather deliver it yourself than have us in
If any link fails, email careers@shadowell.com
with "teachers pack" in the subject and we will send the files directly, same day.
2. The 40-minute activity: "Why most good ideas lose money"
No computers required, though the interactive version on this site works well on a projector. Runs with
Year 12 or 13; needs GCSE-level arithmetic and nothing else.
0–5 min. Set-up. A share is quoted 412.4p to buy from you, 412.8p to sell to you.
Ask: who pays for that gap, and what does it cost on £100,000? (Answer: about £48.50, one way.)
5–15 min. Students are given a "signal" that predicts a 20 basis point move and a
cost sheet. In pairs, they work out whether trading it is profitable. Most say yes. Then you hand out
the second cost sheet, with stamp duty on it. Most flip to no. This is the entire lesson.
15–30 min. Open the strategy lab on the board. Ask a
student to find a setting where a very large edge still loses money. Ask another to explain why
turning the turnover slider up makes things worse rather than better.
30–40 min. Discussion: "one good result out of twenty tries". Roll dice to
demonstrate. Land the point that a striking result is evidence of nothing until you know how many
times somebody looked.
Curriculum hooks: percentages and compound growth; probability and conditional
probability; hypothesis testing and significance; the difference between correlation and causation;
A-level Economics on market efficiency; Computer Science on simulation and pseudorandomness.
3. What we will do for you, free
Come and speak. 30–60 minutes, in person anywhere in West Yorkshire, or by video
anywhere in the UK. A practitioner, not a recruiter. We do not sell anything and we do not run a
careers pitch at the end unless you ask for one.
Host a visit. Small groups can come to Manningham Lane and see the desk. Best for
groups of eight to fifteen.
Look at a student's project. Send us an EPQ, an NEA or a personal project with a
data or code component and one of us will write real feedback on it. No strings; we simply enjoy it.
Answer the awkward question. If a student asks you something about this industry
you would rather not guess at, forward it.
No degree is required for any of our early-career roles. Neither is finance experience.
Every role and programme is paid, including the one-week insight programme.
The best thing they can bring is something they built or proved. A scruffy Python script counts.
A well-argued EPQ counts.
Strong A-level maths matters for the quant roles specifically. Predicted grades are fine.
The assessment and its rubric are published, so they can see
exactly what is expected before applying.
Applications go to careers@shadowell.com and are read by
a practitioner. Everyone gets a reply.
5. Safeguarding and data
We are used to working with under-18s. In practice: we do not collect more information than a role
requires, we will not correspond with a student about employment without a parent, guardian or member of
staff copied in where that is appropriate, our visitors are accompanied, and we hold DBS checks for
staff attending schools. Our full position is in the privacy policy,
section 9. Any safeguarding question:
hello@shadowell.com.
If it would help to have this as one printed sheet, use the print button above; it lays
out as a single page with a QR code back to this document.
NDA policy for partner conversations
The short version. We will sign a short, mutual, two-year NDA before any substantive
conversation with a venue, broker or data provider. We publish our standard form rather than
negotiating one, and we will not sign anything that restricts what we are allowed to build.
1. When we use one
Before any discussion where either side would need to describe something non-public: pricing,
connectivity architecture, dataset contents, product roadmaps, or our own systems. We do not need one
to talk about publicly available facts, and we will say so rather than paper a conversation that does
not need it.
2. Our standard terms
Mutual. Identical obligations both ways. We will not sign a one-way agreement that
binds only us.
Two years from disclosure, then it lapses. Perpetual confidentiality obligations
are unmanageable for a firm of our size and we decline them.
Defined scope. The purpose is stated in one sentence on the front page. Anything
outside that purpose is not covered, in either direction.
Standard carve-outs. Information already public, independently developed, lawfully
received from a third party, or required to be disclosed by law or a regulator.
Residual knowledge is retained. Our people may not use your documents, but they
cannot and will not unlearn general skills and know-how.
English law, English courts.
3. What we will not agree to
Non-compete or exclusivity clauses. An NDA is for protecting information, not for
restricting what we are permitted to trade, build or research.
Residual-knowledge prohibitions, or clauses that would taint a whole team from
working in an area after a single conversation.
Assignment of intellectual property in anything we develop, whether or not it
touches a topic discussed.
Non-solicitation of our staff as a condition of a commercial discussion.
Perpetual terms, or indefinite survival of obligations after termination.
Unlimited or punitive liability, or liquidated damages disproportionate to the
information at stake.
Confidentiality over the existence of the NDA itself, which makes governance
impossible.
4. How to get one signed
Email partners@shadowell.com with the counterparty name,
the one-sentence purpose, and whose form you would like to use. We will send our standard form the same
working day. If you would rather use yours, send it and we will mark it up against this page rather
than against a preference; expect a response within two working days.
5. Candidates
We do not ask candidates to sign an NDA to interview, to receive the take-home, or to
discuss their own work. You keep and may publish everything you produce during our process. Employees
sign confidentiality terms as part of their contract of employment; those are provided in full with the
offer, never after acceptance.
6. Our obligations, in practice
Confidential material we receive is stored in access-controlled systems, shared only with the people
who need it for the stated purpose, and deleted or returned on request at the end of the engagement. We
keep a register of every NDA we are party to. If we ever cannot comply with a term, we will tell you
before we sign it rather than after.
Questions: hello@shadowell.com. This page
describes our policy and is not itself an agreement or legal advice.
Colophon, performance and how the toys work
What this page is made of, what it costs to load, and the exact formulae behind every simulation on it.
Proof beats promises, so here are the numbers and the methodology to argue with them.
1. Construction
One HTML document. No framework, no runtime dependency, no third-party request of any kind at
runtime, not even a font. Reasoning in
why this website has zero dependencies.
Type: the operating system's UI face — SF Pro, Segoe UI Variable, or Roboto — plus the system
monospace. Nothing is downloaded.
Layout: CSS grid with container queries, so components respond to their own width rather than the
viewport. Colour in OKLCH where supported, with sRGB fallbacks, and one Display-P3 accent on
high-dynamic-range screens.
Motion: native scroll-driven animations where the engine supports them, an IntersectionObserver
fallback where it does not, and nothing at all under prefers-reduced-motion.
Graphics: six 2D canvases (hero, technology, telemetry, order book, globe, lab) plus inline SVG.
All are paused when off-screen or when the tab is hidden, and all are disabled entirely under
Save-Data or on very low-memory devices.
Offline: a service worker caches the document so the page opens with no connection. The footer says
so when it happens.
2. Build steps
Zero runtime dependencies is not zero tooling. Four build steps, all producing static output:
Syntax highlighting. Code blocks are tokenised into spans at build time, so no
highlighter ships to your browser.
Share image.og.png (1200×630) is rendered headlessly from the same
hero-canvas code with the same seed, so the card matches the page.
Minification and hashing. The content hash becomes the service-worker cache name,
so a deploy invalidates cleanly.
Validation. Link check, structured-data check, and an axe accessibility pass. All
three must be clean to deploy.
3. Measured performance
Method: Chrome, Lighthouse mobile preset, 4× CPU throttle, Slow 4G, cold cache, median of five runs,
hitting the production origin. Re-measured every deploy; a regression blocks the deploy.
Metric
Result
Budget
Requests, cold load
1
1
Third-party requests
0
0
Transferred, compressed
~112 kB
< 150 kB
Largest Contentful Paint
1.1 s
< 1.8 s
Interaction to Next Paint
32 ms
< 100 ms
Cumulative Layout Shift
0.00
< 0.05
Total Blocking Time
40 ms
< 200 ms
JavaScript, uncompressed
~86 kB
< 120 kB
Cookies set
0
0
Also tested by hand on a £120 Android handset over a congested mobile network, because
synthetic throttling flatters everybody. Replace these figures with your own measurements before
launch — publishing numbers you have not taken would be exactly the sin the rest of this site
complains about.
4. The strategy lab, exactly
A teaching model, not a research tool. Every term is deliberately simple and visible. In basis points
per trade:
Regime multipliers: normal (1.0, 1.0, tail 2%), crisis (impact ×1.45,
adverse ×1.6, tail 9%, vol ×1.7), low vol (impact ×0.85, adverse ×0.9, tail 0.5%, vol
×0.6, and the spread term is not reduced, which is the whole lesson).
Sortino replaces the denominator with downside deviation only, computed on the realised path. Every run
uses 120 Monte-Carlo paths (48 on small or low-core devices) over 756 trading days, all from one seed,
so a shared link reproduces exactly what the sender saw. Dropped CSVs are parsed with the File API and
never leave the browser.
5. The other simulations
Order book. Clustered arrivals: base intensity plus a decaying excitation term
after each trade, so activity begets activity. Cancellations are exponential in distance from the
touch. Executions print to the tape. No venue data is used, licensed or implied.
Telemetry. Synthetic, but keyed to the real Europe/London clock: message rates and
tick-to-trade widen around 08:00 (equity open) and 13:30 (US data), and fall away overnight and at
weekends. It is a stylised model of load, not a recording.
Globe and latency. Great-circle distance from your selected point of presence,
converted at the speed of light in fibre (c ÷ 1.4675 ≈ 204,000 km/s) for the theoretical floor, then
multiplied by 1.25 for a plausible routed path. Real venue-to-venue figures depend on the carrier and
the route and are not ours to publish.
Correlation matrix. Illustrative long-run figures between strategy families, chosen
to make a teaching point about duplication. Not live measurements.
Session clock. Real. Local venue time and daylight saving come from
Intl.DateTimeFormat; session hours are hard-coded indicative primary hours and exclude
holidays and auctions.
6. Accessibility and privacy
Targets WCAG 2.2 AA; the detail and the known limitations are in the
accessibility statement. Three keys in
localStorage, no cookies, no analytics; see the
cookie policy.
7. Corrections
If any number on this page is wrong, or you can reproduce a worse result than the ones published, tell
us and we will fix the number or the site:
hello@shadowell.com.
Privacy policy
The short version. We collect only what you deliberately send us: a form, an email, an
application. This website sets no cookies, runs no analytics, and makes no third-party requests. We do
not sell, share or trade personal data, and we do not build profiles.
1. Who we are
Shadowell ("we", "us") is the data controller for the personal data described in this policy. We are
registered in England and Wales, company number 15776042, registered office
203 Manningham Lane, Bradford, West Yorkshire, BD8 7HP, United Kingdom. For any privacy matter, email
hello@shadowell.com marked "FAO Data Protection".
2. What we collect, and why
Enquiry form: your name, email address, chosen topic and message. Used solely to
read and reply. Lawful basis: legitimate interests (responding to a request you made of us).
Email correspondence: anything you send us, including attachments. Lawful basis:
legitimate interests.
Job applications and puzzle submissions: your CV or note, education, work history,
right-to-work information and interview notes. Lawful basis: legitimate interests in assessing
candidates, and legal obligation where checks are required by law.
Server logs: our hosting provider records IP address, user agent, requested path
and timestamp, for security and abuse prevention. Lawful basis: legitimate interests.
We do not use personal data for advertising, and we do not make automated decisions about you. Every
application is read by a human being.
3. What we deliberately do not do
No cookies, analytics, advertising pixels, session recording or fingerprinting
No third-party requests of any kind: no CDN, no hosted fonts, no embeds
No data sales, no data sharing for marketing, no enrichment from data brokers
No newsletter list, so nothing to unsubscribe from
4. Storage on your own device
The site uses your browser's local storage for four things, all of which stay on your device and are
never transmitted: sw-pref (your light / dark / auto preference),
sw-draft (an unsent contact-form draft, so you can come back to it),
sw-sound (whether interface sound is on) and sw-hint (whether you have
already seen a one-off hint). Clearing site data removes all four. See the
cookie policy.
5. Who else sees your data
Only the service providers we need to operate: our website host and our email provider, each acting as
a processor under contract. We may disclose data where we are legally required to. We do not transfer
personal data outside the UK except where a provider does so under an approved safeguard such as the UK
International Data Transfer Agreement or adequacy regulations.
6. How long we keep it
General enquiries: up to 12 months after our last exchange
Unsuccessful applications: up to 12 months, then deleted, unless you ask us to keep you in mind for
longer, which we are glad to do
Successful applications: retained as part of employment records for the periods required by law
Server logs: short retention set by our host, typically weeks
7. Your rights
You have the right to be informed, to access your data, to have it corrected or erased, to restrict or
object to processing, and to data portability where it applies. To exercise any of these, email
hello@shadowell.com. We will respond within one month and we
will not charge you. If you are unhappy with how we have handled your data, you can complain to the
Information Commissioner's Office at ico.org.uk,
though we would appreciate the chance to put it right first.
8. Security
Access to enquiry and application data is limited to the people who need it. Mail and files are held in
access-controlled accounts with multi-factor authentication. The site is served over HTTPS only. No
system is perfect; if we ever suffer a breach that presents a risk to you, we will tell you and the ICO
as the law requires.
9. Applicants aged 16 to 18
Our early-career roles are open to people finishing school or college, and we recognise some applicants
will be under 18. We collect no more than we need, and where the law or good practice requires it we
will ask for parental or guardian consent before proceeding. Where we correspond with a student about
employment we will normally ask that a parent, guardian or member of staff is copied in. We do not
knowingly collect personal data from anyone under 16.
10. Changes
If we change this policy we will update the version and date at the top. Material changes affecting
data we already hold will be communicated directly where we have your contact details.
Website terms of use
These terms govern your use of shadowell.com. By using the site you accept them. If you do not accept
them, please stop using the site.
1. Who we are
Shadowell, registered in England and Wales, company number 15776042, registered office
203 Manningham Lane, Bradford, West Yorkshire, BD8 7HP, United Kingdom. Contact:
hello@shadowell.com.
2. What this site is, and is not
This site exists to describe who we are and to recruit. It is information only. Nothing on it is:
an offer to sell, or a solicitation of an offer to buy, any security or financial instrument
an offer of investment management, brokerage, custody or advisory services
investment, legal, tax or accounting advice, or a recommendation of any strategy or instrument
an invitation to invest in Shadowell, or in anything else
Shadowell trades only its own capital and does not manage money for third parties. See our
regulatory disclosure.
3. Illustrative content
The interactive "strategy lab", the synthetic order book, the telemetry panels, the correlation matrix,
the session clock and the globe are illustrative. The lab is a simplified teaching model driven by a
pseudorandom seed; the telemetry and order book use synthetic inputs; latency figures are computed from
great-circle distance and stated assumptions; the session clock shows indicative primary hours and
excludes holidays and auctions. None of it describes Shadowell's actual strategies, positions, systems
or results, and none of it is performance information. Operational figures shown elsewhere are
approximate and are not performance figures. The full methodology is in the
colophon.
4. No reliance
We try to keep the site accurate and current, but we give no warranty that it is complete, accurate or
up to date, and you should not rely on it for any decision. We may change or withdraw any part of it at
any time without notice.
5. Intellectual property
All content on this site — text, code, design, the Shadowell name and mark — is owned by us or licensed
to us. You may read it, print it and quote short extracts with attribution and a link. You may not
republish it wholesale, use it commercially, or use our name or mark in a way that suggests endorsement
or association, without written permission. The teachers' pack is an explicit exception: it may be
copied, printed and adapted freely for educational use.
6. Acceptable use
Please do not attempt to interfere with the site, gain unauthorised access to anything, introduce
malicious code, or place unreasonable automated load on it. Sensible, well-behaved indexing is welcome;
our position on automated and AI crawlers is stated in /robots.txt and
/llms.txt.
7. Recruitment communications
Genuine communications from us come only from @shadowell.com addresses. We never ask
candidates for payment, for banking details, or for money in connection with recruitment, equipment,
training or "evaluation" of any kind. If you receive such a request in our name, it is fraudulent;
please forward it to hello@shadowell.com.
8. Links out
Where we link to third-party sites we do so for convenience. We do not control them and are not
responsible for their content or their privacy practices.
9. Liability
To the fullest extent permitted by law, we exclude liability for any loss or damage arising from your
use of, or inability to use, this site, including any loss of profit, business, goodwill or data.
Nothing in these terms limits liability for death or personal injury caused by negligence, for fraud, or
for anything else that cannot lawfully be limited.
10. Governing law
These terms and any dispute arising from them are governed by the laws of England and Wales, and the
courts of England and Wales have exclusive jurisdiction.
11. Changes
We may update these terms. The version and date at the top will change when we do.
Cookie policy
This website sets no cookies. Not necessary ones, not functional ones, not analytics, not
advertising. That is also why you have not been shown a consent banner: there is nothing to consent
to.
1. Why there is no banner
Consent banners exist because most sites store or read information on your device for purposes that are
not strictly necessary, usually analytics and advertising. We do neither. No cookie is set by this site,
and no third party is given the opportunity to set one, because the page loads nothing from anyone
else.
2. No third-party requests at all
This page is a single document served from our own domain. There is no content delivery network, no
hosted font, no analytics script, no tag manager, no social embed, no video player, no chat widget and
no advertising code. Nothing about your visit is disclosed to a third party by the act of loading the
page. You can verify this in your browser's network inspector, and we would encourage you to.
3. What we do store, on your device only
The site uses your browser's localStorage, which differs from a cookie in one important
way: it is never attached to a network request, so it cannot be used to track you across sites and it is
never sent to us.
Key
Purpose
Lifetime
sw-pref
Remembers whether you chose light, dark or automatic appearance
Until cleared
sw-draft
Keeps an unsent contact-form draft so you do not lose it
Until sent or cleared
sw-sound
Remembers whether you switched interface sound on
Until cleared
sw-hint
Notes that a one-off interface hint has already been shown
Until cleared
All four are strictly functional and are set only in response to something you did. You can remove them
at any time by clearing site data for shadowell.com in your browser settings; the site works perfectly
well without them.
4. The service worker cache
So that the page opens without a connection, a service worker stores a copy of this document in your
browser's cache. It contains only the public page you already downloaded, holds nothing about you, and
is never transmitted anywhere. Clearing site data removes it, and unregistering the service worker in
your browser's developer tools disables it entirely.
5. Server logs
Our hosting provider keeps standard access logs — IP address, time, path, user agent — for security and
abuse prevention. That is a property of running a web server, not a cookie, and we do not use those logs
for analytics or profiling.
6. If this ever changes
If we ever introduce anything that stores non-essential information on your device, we will ask for
your consent first, in a way you can decline without losing access to anything, and we will update this
page. Questions: hello@shadowell.com.
Modern slavery & human trafficking statement
Shadowell publishes this statement voluntarily. Our turnover is below the £36 million threshold at
which section 54 of the Modern Slavery Act 2015 requires a statement, and we are therefore not obliged
to produce one. We publish it because the standard should not depend on the threshold.
1. Our organisation
Shadowell is a proprietary trading group registered in England and Wales (company number
15776042), operating from Bradford, West Yorkshire. We trade financial instruments with
our own capital. We have no manufacturing, no physical products and no complex multi-tier supply chain.
2. Our supply chain
Our suppliers are almost entirely providers of professional and technical services: market data
vendors, exchanges and trading venues, brokers and clearing providers, co-location and cloud
infrastructure, software licences, professional advisers, and office services. The principal areas where
modern slavery risk could realistically arise are facilities, cleaning, security, and hardware
manufacture further down the chain of our infrastructure providers.
3. Our commitments
We do not tolerate slavery, servitude, forced or compulsory labour, or human trafficking in our
business or in any part of our supply chain.
All work at Shadowell is voluntary. Nobody is required to lodge identity documents or deposits, and
anyone may leave on the notice set out in their contract.
We pay for all work, including internships and insight programmes. We do not use unpaid labour,
unpaid trials or unpaid "evaluations".
We never charge any fee to any candidate, at any stage, for any reason.
We verify the right to work of everyone we engage, and we do so in a way that respects their
privacy.
4. Due diligence
Given our size, our due diligence is proportionate: we prefer established, reputable suppliers who
publish their own policies; we review the modern slavery position of any supplier providing on-site
labour before engaging them; and we will terminate a relationship with any supplier unwilling to confirm
compliance or to cooperate with a concern we raise.
5. Raising a concern
Anyone — employee, applicant, supplier or member of the public — may raise a concern about modern
slavery in connection with our business by emailing
hello@shadowell.com. Concerns can be raised in confidence.
Nobody will suffer detriment for raising one in good faith, and we will support the use of external
reporting routes including the Modern Slavery Helpline and the police where appropriate.
6. Review
This statement is reviewed annually and approved by the board of Shadowell. It will be updated as the
firm grows, and it will become a statutory statement if and when we cross the threshold.
Regulatory disclosure
This page sets out plainly what Shadowell is, what it is not, and what that means for anyone who
contacts us. If anything here conflicts with something you have been told in our name, this page is
correct and you should tell us.
1. Proprietary capital only
Shadowell trades financial instruments exclusively with its own capital, for its own account. We do not
manage, pool, hold or have access to money or assets belonging to any third party.
2. Regulatory status
Shadowell is not authorised or regulated by the Financial Conduct Authority and does
not appear on the Financial Services Register. We do not carry on regulated activities that require
authorisation. Specifically, we do not:
deal in investments as agent, or arrange deals for anyone else
manage investments for clients, or provide investment advice or personal recommendations
hold client money or client assets, or provide custody or safeguarding services
operate a collective investment scheme, fund, or any pooled vehicle
provide payment, lending, insurance or credit services of any kind
Because we have no clients, protections that exist for clients of regulated firms — including the
Financial Ombudsman Service and the Financial Services Compensation Scheme — are not relevant to any
dealings with us, and nobody should expect them to be.
3. Not an offer or a financial promotion
Nothing on this website, and nothing said by anyone at Shadowell, is an offer or invitation to engage
in investment activity, an offer of securities, or a solicitation of investment. We do not raise capital
from external investors and we are not seeking to.
4. Not a funded-trader or evaluation programme
We do not sell trading challenges, evaluations, "funded accounts", signals, courses, mentoring, copy
trading or any similar product. We will never ask you to deposit, transfer, stake or risk your own money
with us or through us. If you have been offered any of these under the Shadowell name, it is not us.
5. Illustrative material
The interactive tools and visualisations on our website are simulations built from synthetic or
pseudorandom inputs, documented in full in the colophon. They are
provided to explain concepts, not to describe our systems or results. Nothing on the site is, or should
be read as, a representation of trading performance or a forecast.
6. Market conduct
We maintain internal policies on market conduct and market abuse, on the handling of confidential and
inside information, on personal account dealing, and on the fair and orderly use of the venues we
connect to. Where a venue, broker or data provider imposes conduct obligations on us by contract, we
treat them as binding.
7. Recruitment fraud
Genuine Shadowell communications come only from @shadowell.com addresses. We never request
payment, banking details or identity documents by message during recruitment, and we never ask
candidates to buy equipment or training. Report anything suspicious to
hello@shadowell.com.
8. Company information
Shadowell, registered in England and Wales, company number 15776042. Registered
office: 203 Manningham Lane, Bradford, West Yorkshire, BD8 7HP, United Kingdom. Our public record is at
Companies House.
This disclosure is governed by the laws of England and Wales.
Accessibility statement
We want this site to be usable by everyone, including people using screen readers, keyboard-only
navigation, magnification, high-contrast modes or reduced-motion settings. We treat accessibility as an
engineering requirement rather than a compliance exercise.
1. What we aim for
We target conformance with the Web Content Accessibility Guidelines (WCAG) 2.2 at level AA. We test
with keyboard only, with the operating system's reduced-motion and appearance settings, at increased
text sizes, and with a screen reader. An automated pass runs on every deploy and must be clean.
2. What we have done
Keyboard operable throughout. Every control, menu, accordion, filter, slider,
overlay, glossary term and document reader can be reached and used from the keyboard, with a visible
focus indicator and no keyboard traps. Overlays trap focus deliberately while open and return focus
where you left it.
Semantic structure. Real landmarks, a logical heading order, native buttons and
links, labelled form fields, and a skip link to the main content.
Collapsed text is findable. Accordion panels use hidden="until-found",
so your browser's find-in-page will locate text inside a closed panel and open it for you.
Reduced motion. If your system asks for reduced motion, animation, parallax,
marquees, magnetic buttons, view transitions and canvas motion are switched off rather than merely
shortened.
Appearance. The site follows your operating system's light or dark setting by
default, and you can override it. Colour is never the only way information is conveyed, and we honour
increased-contrast and forced-colours settings.
Dynamic content is announced. Filter results, copy confirmations, form validation,
the latency and cost figures, and the strategy lab's outputs are exposed through live regions, so a
screen reader user hears the same information a sighted user sees. Sliders announce their real mapped
value and unit, not the raw position.
Decorative visuals are hidden. Canvas animations and diagrammatic SVGs are marked
decorative so they are not read out as noise; where a visual carries meaning it has a text
equivalent.
Sound is off by default and there is an explicit control for it in the footer.
Nothing plays audio without being asked.
No time limits. Nothing on the page expires, times out or moves on without you.
Works without JavaScript. All written content, including every note and document,
is present in the page source and readable with scripting disabled.
3. Known limitations
The rotating globe is decorative. It has a text description, and the same information is available
as a keyboard-operable list of locations next to it, including the latency figures.
The synthetic order book and the telemetry panels are illustrative animations. They summarise their
values as text, but they do not offer an underlying data table because there is no underlying real
data.
The strategy lab is inherently visual. Every input is a labelled slider with a text value, and the
results are announced as a sentence, but the chart itself is not sonified. The path data can be
exported as CSV.
The interactive figures inside notes each carry a static text fallback that states the finding, so
the point of the figure survives without interaction.
Long-form documents open in an overlay reader. It is fully keyboard accessible and each document has
its own shareable link, but it is a single scrolling view rather than paginated.
The correlation matrix conveys magnitude with colour and with a printed number in each cell; the
interpretation of any pair is announced as a sentence when selected.
4. Tell us if we have got something wrong
If any part of this site is difficult or impossible for you to use, email
hello@shadowell.com and describe what happened and what you
were using. We will reply, and we will fix it: accessibility problems are treated as defects, not
feature requests. If you need any content on this site in another format, ask and we will send it. If
you need an adjustment to our recruitment process, that is arranged without question and without
explanation.