Start Here

A thread quotes a 71% win rate for fading bitcoin funding. A bank note says buy AUDJPY for the yield. On $100,000, a 4% gap between Australian and yen rates is about $4,000 a year, and a normal day moves the pair by roughly fifty times that paycheck. Into 2008 the gap had been near 7% a year, and AUDJPY then lost nearly half its value in about three months. The paycheck was the premium for warehousing that crash. The person forwarding either pitch usually cannot name the payer, or the date the payer stops.

This page is the web version of The Quant Map: the map's own stations, numbered and named as on the plate, each with the articles that answer it. Pick the question you cannot answer about a trade you already have, and read that row. A blank link means the piece is written and queued. The printable PDF carries the plate in vector, five worked descents (example research from observation to live rule), and the full 835-entry strategy library. It waits at the end of the page.

Keep four words apart. A theory explains why someone pays. A feature is a number you could have computed that morning. A strategy turns the number into a repeated decision, with a size and a wrapper. A factor asks the same question across a universe. Between the feature and the strategy sit a forecast, a threshold, and a cost check. Any one of those can zero the trade.

The grand map

Every box on the plate is a section of this page. Open the plate full size to read it closely. The PDF at the end of the page carries the same plate as a vector A2 sheet, printable for the wall.

The spine, stage by stage

Seventeen stations, numbered and named as on the plate. Find the station where your trade broke, or walk them in order and stop at the first one you cannot pass.

I · Foundations

the toolkit you think with, pedagogy, not alpha

A strategy that wins 55% of trades still hits eight losers in a row about 0.17% of the time, once or twice per thousand trades. That run is the strategy working. Risk 2% of a $5,000 account and the streak costs about $800. Risk 15% because the setup looked clean and the same streak is the account. A second trap sits in the average: +60% then −40% advertises +10% a year, and $100 becomes $96. Compounding pays the variance. The average does not.

The full station, every article in reading order: I · Foundations.

II · Market architecture

where edges live: underlying markets × instruments, two different axes

The market is the thing whose price you care about. The wrapper is the contract you hold: spot, forward, future, perpetual, ETF, CFD. Four ways to be long bitcoin are one view and four P&Ls. Spot has custody. A quarterly future makes you pay or earn the basis every roll. A spot ETF sleeps while bitcoin does not. A perpetual never expires, and the crowded side pays funding, about 11% a year when the 8-hour print is 0.01%, and much more in a mania. In April 2020, spot oil recovered from the negative print. Holders of the rolled-futures wrapper did not get that recovery back. Being right about the market and wrong about the wrapper still loses the money.

The full station, every article in reading order: II · Market architecture.

III · Market theory

how markets can work in principle, efficiency is the null hypothesis

Events arrive in bursts. Volatility clusters. A liquidity shock can rewire correlations inside an afternoon. Clock time treats those as evenly spaced samples, and a lot of the structure lives in the waiting time, the path between prints, and the shape of the book of correlations. The articles here are frontier reading: they assume you already know what a null is, what a feature is, and what a cost line is. Read them cold and you collect vocabulary. Vocabulary does not size a position.

The full station, every article in reading order: III · Market theory.

IV · Edge mechanisms

why should an edge exist?, the root question of the map

Name the payer before you name the indicator. Reachable money is usually a risk someone better-resourced refuses to warehouse, a capacity limit that makes the trade a rounding error for them, or a flow that has to deal. A $5 million niche earning 40% a year is a living at your size and 0.01% of a $20 billion fund, which is why the fund leaves it. If the answer to "why does this pay" is "the backtest went up," you do not have a mechanism yet.

The full station, every article in reading order: IV · Edge mechanisms.

V · Observable world

what can we measure?, point-in-time, or it does not exist

An observable is whatever was knowable then: price, a rate, funding, open interest, a commitment-of-traders print, the book. A revised series, a delisting filled in later, a hedge ratio fit on the whole sample, none of those existed that morning. Point-in-time is the filter. If the series fails it, the feature you build on top is a look at the answer.

The full station, every article in reading order: V · Observable world.

VI · Feature engineering

transforms of observable state, one raw variable, many research questions

One raw series supports many questions, and they are different hypotheses. Carry level asks how big the paycheck is. Carry over volatility asks how big it is per unit of pain. Carry rank asks whether this pair is the best paid in the set. Carry change asks whether the gap is widening. Carry conditioned on trend asks whether price already agrees. Five features, two rates and a spot. RSI, MACD, and a stochastic on the same close are often one band agreeing with itself. Stacking them feels like confirmation. The frequency response says it is one feature in costume.

The filter shelf past this table is the same question asked of one operator: which band it keeps, and how late. Butterworth, sinc, Hilbert, roofing, and the named oscillators are later visits. Start with the primer and the frequency response. A filter output is a feature. The forecast is a later station.

The full station, every article in reading order: VI · Feature engineering.

VII · Hypothesis

“if mechanism X is real, feature F should predict Y over horizon H”

Write the death condition before the run. "Ranked by carry-to-volatility, the top third of G10 beats the bottom third over the next one to three months, after costs" can die: the buckets fail to rise in order, or the whole result lives in 2002 to 2007. "This chart looks like a head and shoulders" cannot die. Explanation is cheap after the print. A prediction is a dated, specific claim you cannot edit. The scientific bar is the pre-commit: hypothesis, null, statistic, threshold, held-out window, trial count, in a file whose timestamp is earlier than the result.

The full station, every article in reading order: VII · Hypothesis.

VIII · Validation

is it real, and does it survive costs?, most ideas die here, and that is the point

Most ideas die here, and the archive is the output. A single Sharpe is a sample statistic with a wide interval: a Sharpe near 1 on a few years of daily data is often consistent with a true Sharpe near 0. The best of forty-one thresholds is a maximum. At a 5% bar the family-wise false-positive rate on that scan is about 88% before you correct it. Add the spread, the fees, and an honest slippage before you like the curve. A printed machine-learning paper can fail on its own numbers: a do-nothing baseline, a base rate, a cost line that was never subtracted.

The full station, every article in reading order: VIII · Validation.

IX · Forecast / signal

a statement about the future, not a trade yet

AUDJPY at +1.3 and EURCHF at −0.8 are comparable units, and neither is a position. The station is the map from feature to expected value; in a cross-section, the stronger version is often an order, not a number.

The full station, every article in reading order: IX · Forecast / signal.

X · Decision policy

forecast → position: trade only if E[r] clears costs + threshold

The decision is the arithmetic after the score: expected edge next month, about a 4.1% gap divided by 12, near 34 basis points, against a couple of basis points to get in and out at institutional size. Clear the toll with room, then size by risk so one wild pair cannot own the book. Refuse a rule whose expected gross edge per trade sits under about three times the all-in round trip. That filter deletes most of what gets marketed as a five-minute system.

The full station, every article in reading order: X · Decision policy.

XI · Strategy

a decision rule repeated through time, and not everything that trades is alpha

A strategy is the decision frozen: monthly, rank by carry-to-vol, long the top three, short the bottom three, equal risk, rebalance. Morning mood is no longer an input. A clever way to buy is plumbing, a regime model is weather, and a risk limit is a seat belt. Useful, and none of them is a reason someone is paying you. High-noise markets want reversion. Low-noise markets want trend. One indicator on every chart is a refusal to read the observable.

The full station, every article in reading order: XI · Strategy.

XII · Factors

the same mechanism, systematically, across an entire universe

Factors come after one asset works end to end. A carry factor is the same question asked of every pair: high carry against low carry, and you no longer care whether AUDJPY rises.

The full station, every article in reading order: XII · Factors.

XIII · Multi-factor models

composites: carry + momentum + value, regime-conditioned

Composition reopens the search. A leftover alpha in a regression is often a factor you left out. A zoo of published factors is a counted search; the t-stat you remember was the maximum of many tries. Timing the factors, turning them on and off with a second model, mostly fails once you count that second search. Cross-sectional machine learning, in the study that measured it carefully, lifts about 20%.

The full station, every article in reading order: XIII · Multi-factor models.

XIV · Alpha ensembles

combining edges into one forecast engine

Two edges with a Sharpe near 0.5, roughly uncorrelated, combine to about 0.5 times the square root of 2, near 0.7. Five copies of the same idea, correlation 0.85, give you almost nothing and one decay date.

The full station, every article in reading order: XIV · Alpha ensembles.

XV · Portfolio construction

from forecasts to risk: budget it, do not just add it

Budget a portfolio volatility, cut every position when realized vol doubles, cap one name's share of risk, and write the drawdown rule before the drawdown. Signal quality and portfolio quality come apart: a real signal with the wrong weights is a bad portfolio. Overbetting is the failure mode the Kelly fraction exists to refuse.

The full station, every article in reading order: XV · Portfolio construction.

XVI · Implementation

where paper alpha meets the market

On Hyperliquid BTC in a studied window, a refused post-only buy returned about 0.31 basis points over ten seconds. The taker fee sat around 1.44 to 2.7 basis points. Stand aside as a taker; the edge does not clear the toll. Markouts are the truth serum: where the mid is a short time after you filled. Your own size walks the book at something near the square root of the fraction of volume you took. A simulation that fills you at mid, with no queue and no clock drift, is a different market from the one that sends the reject.

The full station, every article in reading order: XVI · Implementation.

XVII · Live research

strategies are mortal, monitor, adapt, retire

The live rule is an experiment that does not end. Watch the paycheck (the gap can close because policy changed), the crowding (positioning stretched means a narrow exit), and the weather. Two exits are both legitimate. Adapt: go back to the mechanism, the observable, and the feature with what the break taught you. Retire: archive the rule with a post-mortem so the next version of you does not pay the tuition twice. A crowded edge grinding lower as the trade gets popular does not heal by waiting. A barriered edge inside a historical-looking drawdown is often noise. The difference is the diagnosis, and it has to be written before you are in the drawdown.

The full station, every article in reading order: XVII · Live research.

The two rings (and where ML sits)

Two conditions wrap every station of the spine.

The regime / state ring

which market are you in today?

The same carry rule is an income stream in calm markets and a crash in a yen squeeze. Into August 2024 the Bank of Japan lifted rates to about a quarter of a percent, USDJPY fell from about 162 to about 142 in weeks, and the Nikkei dropped about 12% in a day. The mechanism was intact. The weather had changed. A backtest that lived inside one regime is a sample of that weather. Regime conditions the feature, the forecast, and the size at once.

The research governance ring

is the research telling the truth?

Governance polices every arrow between the stations. Before the run: the hypothesis, the null, the benchmark, the number of trials, the point-in-time rule, and the result that would kill it. A parameter that works only at 13 and fails from 11 to 17 memorized the sample. A loose region, Sharpe stable across a band, is the one you can defend. Simplicity is the degrees-of-freedom cap: effective sample size in markets is a few hundred independent bets, and each free parameter spends some of them.

Where machine learning sits

A fitted model weights and conditions signals that already have a payer. "I trained XGBoost" is a method. The mechanism is still the question in the first section. On this site the fit shows up late: after the feature is stationary, after the band is named, after the null exists. Trees will spend their splits on a raw price level and call it skill. Regularization is the complexity penalty you can write down. Below a few minutes the function class can matter; above an hour the feature usually matters more. A model that only works as an interaction still has to clear costs as a trade.

The markets

The map walks fourteen markets and the instrument layer on top of them. One family gets its own shelf here: prediction markets, where a fair book is arithmetic you can check.

Alternative / event markets

On a prediction market the cleanest money is a book that cannot all happen at once: two contracts that each look fair and together cost more or less than the certain payoff. That is a Dutch book, and the trade is a projection back onto the set of prices that add up. Forecasting the event is further down the list, and it is the edge that crowds fastest. The same spine still applies. Who pays, what you could have measured, what would kill it, whether the fill clears the fee. A copied arbitrage that crosses the spread is a different trade from the one the detector saw.

Playing a different game: the retail edge

Game selection beats skill. A small account has real advantages, and none of them is speed.

Which table can this account win?

One taker round trip a day at 10 basis points costs about 25% a year (252 times 10 basis points). Ten a day costs 252%, before the rule has earned a dollar. Latency games, market making on major venues, and reacting to a scheduled print put you across from a machine that saw it first and pays a fraction of your fee. The tables a small account can win are the ones where patience, a tiny capacity, and a public calendar matter, and speed does not: days to months, neglected names, funding you can collect, pre-announced rebalances. Boredom is the cheap toll. A prop-firm fee is a product. Read the daily-loss rule as a sizing constraint before you pay it.

Take the map with you

The spine, the two rings, fourteen markets, five worked descents, and a library of about 835 strategy entries, each with a named payer and a named failure mode, in one 65-page PDF. The grand-map plate inside is vector at A2. Print that page for the wall and read the rest as a booklet. The tables above are rows of that library written out in full; the PDF tells you which rows to want next.

Download The Quant Map (PDF, free, no email)