Autonomy in an investment process is not a switch but a ladder, and each stage of the process, from research to execution to monitoring, sits on a different rung. We borrow the levels-of-autonomy vocabulary from driving, apply it stage by stage, ask what a system would have to demonstrate to climb, and set out the boundaries that should stay under human control no matter how capable the machine becomes.
- Offices:
- Chicago
- London
- Bahrain
Library
Showing all 30 pieces
202612 min read
Predictive relationships in markets are found, written up, sold, and copied, and at each step some of their power leaks away. We set out a simple way to measure how fast, show why the answer differs so much across signal families, and argue that the question investors should ask is not whether alpha is vanishing but how quickly it must be replaced.
Quantitative signals are built to be independent bets, but as the same data and the same methods spread, more capital ends up positioned on the same forecasts. We separate the two ways crowding erodes expected return, the decay of the prediction itself and the decay of the trade that acts on it, model the point at which a still-accurate forecast stops being worth trading, and show why crowding appears first not in returns but in correlations.
Being faster is widely assumed to be the same as being more profitable. It is not. Using a simple decay model, we show that speed pays only when an opportunity is disappearing on a timescale close to a system's reaction time, that the marginal value of speed is concentrated in a narrow band around that horizon, that most of a diversified inventory sits well outside it, and that latency is best treated as an engineering budget with a physical floor.
When a liquid futures contract jumps away from the price implied by related instruments, the opportunity begins to decay at once, but not at one speed. We build a stylized model of how much of such a dislocation a participant can expect to capture as a function of reaction delay, from a microsecond to a minute, and find three distinct shelves rather than a slope. The shelves, not the microseconds, determine where a firm should choose to compete.
The fundamental law of active management says that expected risk-adjusted return grows with skill and with the square root of the number of independent bets. That arithmetic is why systematic investors prefer enormous numbers of small, weakly correlated forecasts to a handful of convictions. It also carries fine print: bets are rarely independent, skill rarely holds up across thousands of them, and the law describes a ceiling rather than a promise.
The correlations that make a portfolio look diversified are not parameters but outcomes: they reflect which kind of news is moving prices, and that changes with the regime. Using stylized simulations, we show how the stock-bond correlation flips sign across growth, policy, liquidity, and crisis regimes, why the whole cross-asset matrix compresses when volatility rises, and what a portfolio process has to do differently once it accepts that correlation is a state, not a constant.
Portfolios built from many apparently unrelated positions routinely discover, in a crisis, that they held one position many times over. We set out a way of counting the bets a portfolio actually contains, show how that count collapses in a stylized stress episode, and trace the collapse to exposures, above all liquidity, that are invisible in calm data because they are only expressed under stress.
A trade is crowded when its holders would all want to leave at the same time, and the loss that follows is the cost of the exit rather than the failure of the thesis. We set out six observable symptoms of crowding, show how to combine them into a single score, and use a stylized model to illustrate why a crowded position's drawdown is deeper, wider in its range of outcomes, and only partly permanent.
Every trade that demands immediacy pays for it, in the spread and in the price impact of the order itself. We set out a square-root model of that cost, show how it moves as volatility rises and resting depth thins, and follow a hypothetical strategy as capital scales past the point where impact consumes its forecast. The conclusion: a predictive relationship is not alpha until it has paid for its own execution.
The best result of a large search looks like a discovery even when nothing was found. We work through the statistics of that fact: how the expected Sharpe ratio of the luckiest of N random strategies grows with N, how the evidentiary hurdle has to rise with the number of tests, why base rates decide how many declared discoveries are false, and which forms of selection bias a research process has to be designed against.
Historical simulations see the future through channels that are easy to miss and consistently flattering: universes that omit the securities that failed, economic figures stored in revised rather than first-printed form, timestamps that describe a period rather than a release, and corporate actions adjusted backward from today. We catalog the leaks, show in stylized form how much each can add to an apparent track record, and describe the point-in-time pipeline we consider the minimum for taking a backtest seriously.
Richer models fit financial history better, and a better fit is exactly what should worry a researcher. We walk through the arithmetic that separates in-sample from out-of-sample accuracy, show why markets punish complexity harder than most domains, compare model classes on a stylized panel, and argue that the complexity that matters most is often the kind that never appears in the model at all: the number of things that were tried.
Market commentary names regimes after they have ended. A systematic investor needs to recognize them while they are still in progress. Using a two-state hidden Markov model as the simplest precise definition, we separate what a real-time filter can know from what hindsight knows, measure how long recognition takes, and show why the choice of detection threshold is a policy on false alarms rather than a technical setting.
Volatility is the one quantity a market forecasts about itself, and it does so in several places at once: in the record of past returns, in option prices at each tenor and strike, and in the way those prices move together across markets. We set out what each of these measures knows, where they disagree, and why the disagreements, rather than the levels, are the part a systematic investor should attend to.
Market crashes share a direction and little else. A liquidity crash unfolds in minutes and largely reverses; a fundamental repricing unfolds over months and does not; a policy-driven episode falls until an official decision turns it. We compare the three archetypes on speed, order-book depth, correlation, volatility, recovery, and transmission, using stylized models, and argue that the shape of a drawdown, read from its microstructure, tells an investor more than its size.
The U.S. equity session begins with an auction and a burst of trading in which a large share of the day's price discovery is compressed into a few minutes. We describe what that burst looks like in volatility and spreads, how the opening cross arrives at the first price, why large and small stocks complete the process at different speeds, and what the variance ratio of early returns reveals about overshoot in different volatility regimes.
An equity's price is discovered for six and a half hours a day and carried, undiscovered, for the other seventeen and a half. Splitting returns at the open and the close produces two series with different characters: in a stylized model the drift lives overnight and the variance lives in the session. We set out the decomposition, show how the split shifts as information moves outside regular hours, and weigh the explanations on offer.
A growing share of U.S. equity volume no longer trades continuously but in one auction at the close, where index funds, ETFs, and benchmark-tracking institutions meet at a single price. We describe why the close attracted that liquidity, how the auction is run, what the published imbalance does to prices in the final minutes, and why a systematic investor has to treat the closing print as a distinct market with its own cost, information, and risk.
Most of the order flow in liquid electronic markets now originates in software, and the systems that generate it resemble one another far more than the people they replaced. We describe how the composition of participants has shifted, why messages have outrun trades by orders of magnitude, how one system's action becomes another's input, and which collective behaviors, from liquidity mirages to correlated withdrawal, that loop produces at machine speed.
Machine-learned models sometimes forecast returns reliably out of sample while offering no account of why. Refusing them wastes real predictive power; accepting them on the strength of a track record alone repeats the oldest mistake in quantitative finance. We set out what an explanation actually buys, why the data cannot substitute for one in time, and a way to convert a missing explanation into a smaller, more closely watched risk budget rather than a yes-or-no verdict.
A combination of several mediocre forecasting models often predicts better than the best of them alone, a result that seems paradoxical until one looks at how their errors relate. We set out the arithmetic of forecast combination, show that the value of a new model depends far more on what it shares with the existing ones than on its own strength, and describe the point at which adding another model begins to make the system worse.
A firm can screen thousands of alternative datasets, but the number of independent things they say is far smaller. Most of what any dataset measures is a common economic exposure other data already capture, and much of the rest is shared with its neighbors. We count the dimensions that survive, show why the price of an independent bit varies widely across families, and argue that this count, not the catalog, is the budget a research process works within.
A price changes at an exchange and, microseconds later, an order arrives back at the same matching engine. In between sits a chain of eight distinct pieces of machinery, each with its own job, its own time budget, and its own way of failing. We walk that chain hop by hop, show where a stylized round trip spends its time, explain why the tail of the latency distribution matters more than the median, and pair each hop's characteristic failure with the control that catches it.
Trading moved into data centers, but information still travels at the speed of light in glass, about 200 kilometers per millisecond. We map the hubs that matter to a firm in Chicago, London, and Bahrain, compute the theoretical minimum time between them, and show that what a microsecond is worth depends on the strategy's horizon and on where the participant stands.
Every microsecond removed from a trading system has a price, and the next one costs more than the last. We set out a simple model of what a microsecond is worth to strategies at different horizons, what it costs to remove through software, network, hardware, and colocation, and where the two curves cross. The optimum differs for every strategy, it moves when competitors invest, and for some races the right amount of speed to buy is none.
A systematic investor carries thousands of forecasts at once, and none of them knows about the others. Turning them into one set of positions takes a hierarchy: combination, a risk model, an optimizer, a constraint layer, and execution. We describe what each layer is for, what each constraint protects against, and what each costs in expected performance and turnover, and argue that constraints are the design decisions that make a portfolio a single decision rather than an accidental sum.
Every systematic strategy has a size beyond which it stops being the strategy that was tested. We use the square-root impact model to show the arithmetic: impact's share of alpha grows with the square root of capital, and the size that earns the most dollars is a fixed fraction of the size that earns none. Capacity is a property of speed, breadth, and execution, and a number to estimate before capital is committed rather than discover afterward.
Gross exposure, net exposure, notional, and leverage are the numbers a portfolio is usually described by, and none of them measures risk. We build three stylized books that are identical on every headline figure and very different in what they can lose, walk through what each measure captures and what it misses, and argue that the useful description of a position is what it is exposed to, how it behaves when correlations change, and how long it takes to leave.
Most portfolios rest on a handful of relationships that are treated as constants: bonds hedge equities, diversification lowers risk, liquidity is there when needed. We stress-test nine such assumptions in a stylized model against four archetypal crises, a credit crisis, a liquidity shock, an inflation shock, and a rate shock, and find that what survives depends less on the assumption than on the archetype. The relationships that persist are fewer, and plainer, than the ones portfolios are usually built on.
This site uses only strictly necessary cookies. By using it you agree to the Terms of Use and acknowledge the Privacy Notice.