Outputs expected returns or ranks. It does not responsibly decide the final position size by itself.
FROM HYPOTHESIS TO POSITION · AND BACK AGAIN
A hedge fund is a decision factory.
Not every hedge fund is quantitative, and no two funds are identical. But a systematic fund can be understood as a loop: turn information into forecasts, forecasts into a risk-aware portfolio, and trading outcomes into evidence.
THE JOBmake better
decisions
decisions
RESEARCHWhat might predict returns?
MODELHow strong is each forecast?
PORTFOLIOWhat risks may we own?
TRADEWhat can actually fill?
LEARNDid it survive reality?
The public can see the architecture. The proprietary edge—data, signals, parameters, constraints, and execution details—usually stays private.
01
Eight stages.
THE FULL SYSTEM
Eight stages.
One skeptical loop.
02
A choir of weak signals
ONE RUNNING EXAMPLE
A choir of weak signals
becomes a portfolio.
Follow a fictional $1,000,000 U.S. equity fund. It evaluates AAPL, MSFT, NVDA, TSLA, GLD, and SPY using three deliberately simple signals. The tickers are real; every number below is an invented teaching input, frozen for this page—not a current forecast.
Mix the signals.
The sliders are model coefficients. Production models estimate and validate these weights instead of choosing them because a backtest looks good.
| Ticker | 20d reversal | 6m momentum | Earnings revisions | Combined score | Alpha forecast |
|---|
Outputs factor exposures, covariance, specific risk, and portfolio-risk forecasts—not an order ticket.
Balances forecast, risk, costs, leverage, liquidity, and mandate constraints to produce target weights.
EXPOSURE MATRIX B · 6 SECURITIES × 4 FACTORS
Many stocks. Few common drivers.
A production equity model may use dozens of style and industry factors. Four are enough to see the dimensionality reduction.
| Ticker | Market | Technology | Momentum | Gold |
|---|
FACTOR COVARIANCE F · 4 × 4
How the drivers move together.
Annualized illustrative covariance. Diagonal cells are factor variances; off-diagonals capture co-movement.
| Factor | Market | Technology | Momentum | Gold |
|---|
Σ = B F Bᵀ + DThe asset covariance matrix Σ is reconstructed from security-to-factor exposures B, the much smaller factor covariance F, and diagonal specific variance D. With 3,000 stocks and 60 factors, B is 3,000 × 60 while F is only 60 × 60—not one 60 × 60 matrix for the stocks themselves.
Constrain the enthusiasm.
This teaching solver approximates a mean–variance objective with projection. Real systems use stronger solvers, richer costs, liquidity limits, and mandate-specific constraints.
maximize μᵀw − λwᵀΣw − cost(w − w₀)
subject to position, gross, and beta limits
subject to position, gross, and beta limits
Forecast alpha—μᵀw · annualized example
Forecast vol—√wᵀΣw · annualized example
Portfolio beta—illustrative market exposure
Gross exposure—sum of absolute weights
One rebalance cost—linear estimate, not realized fill
| Ticker | Forecast | Volatility | Beta | Target weight | Approx. dollars |
|---|
03
Backtest the decision process.
THE TIME MACHINE
Backtest the decision process.
Not the final idea.
ONE MONTHLY REBALANCE · POINT-IN-TIME ORDER
Nothing may travel backward in time.
T−1 CLOSEFreeze the information set
Only records available by the cutoff enter the snapshot.
T 06:00Build features
Lag prices, fundamentals, estimates, and the universe.
T 06:05Forecast alpha
Apply the already-fitted model. Do not retrain on the future.
T 06:06Optimize
Combine forecasts, risk, costs, holdings, and limits.
T OPEN+Simulate execution
Trade only at a price that could actually be observed and filled.
T → T+20Mark and account
Apply fills, fees, borrow, cash, corporate actions, and P&L.
NEXT CYCLELog and repeat
Store attribution and diagnostics before moving the clock.
Step 1/7 · Snapshot contains only information knowable by the T−1 close.
Watch a paper edge shrink.
These controls make explicit arithmetic, not a universal performance model. Signal decay, costs, borrow, and turnover vary by strategy and market.
IN-SAMPLE GROSS—research result
AFTER OOS DECAY—before implementation
TRADING + BORROW DRAG—explicit arithmetic
NET OOS EXCESS—illustrative, not forecast
Does the edge remain positive after the stated decay and costs?
Does it remain positive when trading and borrow costs double?
More trials raise selection-bias risk; one ordinary holdout may not be enough.
A sealed period cannot be revised into another training set after inspection.
Paper or shadow orders test whether simulated prices resemble executable reality.
MODEL PROMOTION DECISION—
WALK-FORWARD VALIDATION
Train on the past. Move forward. Keep a final window sealed.
FOLD 1
FOLD 2
FOLD 3
FOLD 4
trainpurge / embargovalidationfinal untouched testnot used in this fold
04
The outer machinery is known.
WHAT THE PUBLIC RECORD SUPPORTS
The outer machinery is known.
The edge is not.
Data → models → portfolio → execution
Two Sigma publicly describes sourcing and preparing data, modeling persistent signals, combining forecasts with risk and trading costs, and executing efficiently. That directly supports the architecture—not any claim about its private features or parameters.
Open the firm’s overview ↗Many characteristics, many securities
AQR describes a repeatable process that evaluates hundreds of stocks using measurable characteristics such as profitability, momentum, quality, risk, and sentiment, then builds diversified portfolios.
Open AQR’s learning center ↗Exposure, risk, attribution
MSCI describes Barra models as tools for measuring factor exposures, managing risk, finding unintended bets, and attributing performance. Those are risk-system jobs; they are not a source of guaranteed alpha.
Open MSCI’s factor-model overview ↗A holdout can still lie
The authors show why repeated strategy selection can overfit historical simulations and propose a framework for estimating the probability of backtest overfitting. Validation must account for the research process, not only one final line of code.
Open the paper ↗Trading is an optimization problem too
Their classic execution model balances expected market-impact cost against price uncertainty. It anchors the practical point that a target weight is not the same as an executable fill.
Open the journal abstract ↗Factors can be inspected and reproduced
French’s public data library provides factor returns, portfolio sorts, definitions, and historical archives. It is a useful transparent reference for the difference between a factor series and a proprietary security-level forecast.
Open the data library ↗
05
A mental model.
METHOD + BOUNDARIES
A mental model.
Not a trading system.
One species in a varied ecosystem
Hedge funds are private pooled investment vehicles with flexible strategies that can include leverage, short selling, derivatives, discretionary judgment, and systematic models. This page focuses on a liquid, systematic quant workflow; macro, event-driven, credit, activist, relative-value, and discretionary funds can operate very differently.
Examples, not market facts
The tickers are familiar real instruments. Signal values, alpha forecasts, factor exposures, covariance, volatility, beta, costs, and resulting weights are invented and frozen teaching inputs. They were not estimated from current or historical market data and should not be used to trade.
Real equation, simplified solver
The objective is a standard teaching form: expected return minus risk and trading-cost penalties. The page uses a projected iterative approximation with position, gross, and beta limits. It omits nonlinear impact, borrow availability, taxes, margin, integer shares, lots, scenario constraints, concentration groups, and many production details.
Arithmetic, not a statistical guarantee
The lab multiplies in-sample alpha by an explicit retention assumption, then subtracts turnover × cost and short book × borrow rate. The number of trials triggers a warning but no invented “multiple-testing haircut.” A real evaluation needs timestamps, point-in-time universes and fundamentals, corporate actions, executable prices, uncertainty estimates, capacity analysis, independent review, and prospective fills.
SOURCES · ACCESSED AUGUST 11, 2026
Investor.gov · hedge-fund structure, flexibility, leverage, short selling, and investor risks ↗
Two Sigma · public investment-management pipeline and forecast/risk/cost combination ↗
AQR · systematic equity process and broad-universe characteristics ↗
AQR · public value, momentum, quality, beta, and time-series-momentum datasets ↗
MSCI · Barra equity factor models for exposure, risk, construction, and attribution ↗
Kenneth R. French Data Library · factor returns, portfolio sorts, definitions, and archives ↗
Bailey, Borwein, López de Prado & Zhu · The Probability of Backtest Overfitting ↗
Almgren & Chriss · Optimal Execution of Portfolio Transactions ↗
Asness, Moskowitz & Pedersen · Value and Momentum Everywhere ↗
Educational illustration only. It is not investment advice, a recommendation, a solicitation, or a claim that the example signals work. Historical simulations are not executable evidence; past performance does not predict future results. Hedge funds can use leverage, short sales, derivatives, illiquid assets, and limited redemption terms, increasing the risk of substantial or total loss.
PAIR IT WITHVolatility is not a crash →