Before costs, active money as a whole cannot beat the market it belongs to; after costs, it must trail it. This essay works through that arithmetic, separates a fair test from an unfair one, and takes failure apart: fees against gross skill, crowding, capacity, drift, self-selected backtests, and capital that chases good years. It closes with the standard the survivors share.
The alternative to hiring a manager used to be doing nothing, and doing nothing was easy to beat. Now it is a competent machine that builds a diversified, rebalanced, tax-aware portfolio almost for free. This essay argues where a professional gap can still honestly survive: in clean data, in validation that outlasts regimes, in crowding and capacity, and in investor behaviour.
News is the only dataset we refuse, and the reason is provenance, not noise. Headlines come from an industry paid for attention and arrive pre-interpreted by people with other horizons. This essay argues that ignoring the news is a data-governance decision, takes on the strongest case for trading events, and answers it with the one carrier of news we trust: price.
The danger of a generative model in an investment process is not that it is weak but that it is fluent. A system that can justify any position on demand destroys the value of justifications. This essay draws our line between outputs that can be mechanically verified and outputs that are judgments, and answers the objection that refusing AI at the core means being outcompeted.
Diversification is reported as a count, and a count is a photograph of a relationship that moves. A portfolio can add markets and become less diversified, because what matters is whether losses arrive together or apart. This essay argues that breadth is only the starting condition, and that staying diversified takes a maintained practice of sizing, rebalancing, and trimming winners.
Autonomy is usually argued as a question of capability, how far a system can run itself. We argue it is a question of failure: a system may run itself exactly as far as its failures announce themselves, and no further. The lines that should never move guard the decisions whose errors look like success for a long time.
Obituaries for alpha confuse two different deaths. Predictive power rarely vanishes outright; it decays and migrates toward shorter horizons, which quietly raises the cost of harvesting it until the signal is real and worthless at the same time. The useful measurement is not whether an effect exists but its half-life, and what remains after the faster clock is paid for.
A forecast has no value in the abstract. Its value depends on the capital already acting on it. The identical prediction is an edge in an empty room and an exit problem in a crowded one, and the trade does not change when the room fills; the risk does. This piece is about the moment a good prediction becomes a bad trade without being wrong.
Speed is an input, not an edge. It pays only where an opportunity dies faster than competitors can react, and it cannot repair a strategy that loses for ordinary reasons. Much of what is bought as speed is actually insurance against a fear nobody has priced. This piece separates the trades where milliseconds are the strategy from the many where they are a decoration on one.
A dislocation has a price curve in time, and it is not a gentle slope. Being late costs different amounts at a microsecond, a millisecond, and a second, and the costs are not proportional: most of the value is lost at one scale and most of the money is spent competing at another. Knowing which scale your strategy lives at is the difference between buying speed and buying theatre.
Breadth multiplies skill. That is the law, and the fine print is the essay: the multiplication works only when the bets are independent and honestly sized, and correlated bets count once no matter how many lines they occupy in the portfolio. A million small bets that are secretly one bet is one bet, with better marketing.
Correlation is quoted as a constant and behaves as a regime-conditional variable. A portfolio built on average correlations is built for the average day, and the average day is not the day the portfolio was built to survive. This piece asks what diversification is worth when the number it rests on is itself one of the things that changes.
Under stress, portfolios do not diversify less because the arithmetic changes. They diversify less because the binding constraint changes: in calm markets the constraint is information, and in a crisis it is funding, and funding is the same for everyone at the same time. Independence that depends on cheap liquidity was never independence. It was a fair-weather property, rented.
The risk of a position includes the exit behaviour of strangers. Part of crowding can be measured, in positioning, in borrow, in flows. The rest cannot, and must be assumed to be worse than it looks, because the holders you cannot see are the ones who will meet you at the door. This piece is about pricing a risk whose size depends on other people's fear.
Paper alpha is quoted gross of the toll. Immediacy costs more exactly when signals are strongest, because volatility is when everyone wants the same side of the same trade, so the strategies that look best before costs are often the ones most taxed by them. Liquidity gets paid first. What is left is the return. This piece insists on counting in that order.
Given enough trials, noise manufactures beautiful backtests on a predictable schedule. Which means a backtest's beauty is evidence about the search process before it is evidence about the market: the more results were examined and discarded, the more the survivor's polish tells you about the size of the pile it was picked from. This piece computes how the pile changes what a result is allowed to mean.
Most backtests could not have been run by anyone alive at the time they simulate. They use revised data presented as original, universes with the dead removed, and a quiet knowledge of which questions turned out to matter. Point-in-time honesty is not a refinement of a backtest; it is the difference between a measurement and a story about one.
A richer model fits better almost by definition, which is why fit is not the achievement it presents itself as. Prediction is a smaller and ruder claim, and every parameter added to earn it spends from a budget that variance collects. This piece treats complexity as a cost centre with an invoice, and asks what a model bought with the money beyond a better description of the past.
Market regimes are easy to identify in hindsight and expensive to identify in time, and the gap between those two facts is where regime-based strategies go to die. The usable question is not which regime we are in. It is what still works across regimes we cannot label until they are over.
Volatility is the market's forecast of itself, and it is unusually honest for a forecast, because it is priced rather than opined: somebody paid for it. Reading it well means knowing when it is information about the future and when it is a premium for fear wearing a number. This piece is about telling those two apart, and about what each is worth.
Two drawdowns of equal depth are different events. Depth is a statistic; the path is the risk, because speed decides who can respond, what financing does while it happens, and whether a strategy survives to see the recovery it was sized for. A slow loss is a problem. A fast one is a different animal wearing the same number.
The opening minutes of a trading day are not ordinary trading with the volume turned up. They are an information-processing event with its own mechanics, in which overnight knowledge is priced under time pressure, and the speed and quality of that pricing differ by stock and by regime. Treating the open as a normal slice of the day misprices everything measured from it.
A striking share of the equity return has historically arrived while the exchange was closed, which is an awkward fact for models built on intraday behaviour. Someone is being paid for bearing the hours when nothing can be adjusted, and the candidates for why are narrower and more interesting than the usual appeal to mystery. This piece asks who, and for what.
The equity day increasingly ends in a single auction, and so the closing price is no longer merely a price; it is an institution, one that benchmarks, index funds, and performance reports are all anchored to. When one print carries that much weight, its mechanics stop being plumbing and become part of what the market is. This piece is about how that happened and what it quietly changed.
When most participants in a market are systems, the market inherits the properties of the systems: similar logic reacting to similar inputs at similar speeds, liquidity that is plentiful until the moment it withdraws in unison, and feedback loops that run at timescales no human supervises. The participants changed species. This piece asks what behaviour a market has when almost nobody in it is deliberating.
A model can be right and still be unownable. Trust has to be rationed by verifiability, and performance evidence substitutes for understanding only up to a point: it cannot tell you when the model's world has ended, because the model cannot either. This piece sets out how much unexplained performance we would accept, where the substitution stops, and why the stopping point is a risk decision and not a philosophical one.
Averaging imperfect models beats the best of them, when their errors are diverse. The condition is the whole result: a new model that agrees with the ensemble's existing mistakes makes the ensemble worse while making its backtest look better, and it will be added enthusiastically for exactly that reason. Diversity of error, not count of models, is the asset.
Thousands of datasets say far fewer independent things than thousands. Data accumulates by collection; information accumulates only by independence, and most new datasets are old statements wearing new formats. The budget is the count of independent things actually said, and for most research programmes it is small, finite, and spent faster than anyone admits. Knowing its size changes what you bother to test.
Between a price changing and a trade existing there are eight hops, and each one is a place where time leaks, and occasionally the trade does too. This is an anatomy, not an argument: the full path, named hop by hop, so that arguments about speed can be held against a map of where the speed actually goes.
Electronic markets were supposed to abolish distance. They priced it instead, in microseconds per mile of fibre, and geography became a line item in strategies that never mention a place. The map did not disappear from trading. It was converted into time, and time into money, and this piece follows the conversion.
Speed has diminishing returns and rising costs, which means there is a point where the next microsecond costs more than the trades it wins. Rational buyers stop there. Many do not, and continue buying speed as insurance against being the slowest, which is a different purchase with a different logic, and should be admitted as one. This piece separates the two purchases and prices each.
Forecasts do not add. Between a thousand independent predictions and one set of positions stand layers of decisions about what to believe, how much, and what to offset against what, and errors behave differently at each layer. The portfolio is not the sum of its signals. It is the residue of a hierarchy, and this piece walks the hierarchy in the order the decisions are actually made.
A strategy's returns are partly a function of its own size, because at some scale it begins moving the prices it trades on and paying its own footprint as a cost. Capacity is therefore not a marketing figure. It is the size at which the strategy starts trading against itself, and honest managers measure it before their investors discover it.
The size of a portfolio is the most available number about it and one of the least informative. The same notional total can be conservative or fragile depending on how the parts correlate, how they are financed, and how losses would arrive in sequence. Risk lives in the relationships between positions, not in their sum, and the position that matters most is often the one no report lists: the assumed relationship itself.
Most stress tests examine scenarios their designers already imagined, which makes them examinations of imagination. The sharper test runs the other way: list the assumptions the portfolio rests on, ask which of them, if false, ends it, and whether the portfolio would announce the falseness in time to act. Scenarios are stories. Assumptions are load-bearing. This piece tests the load-bearing ones.
This website uses strictly necessary cookies only, for security and its core functions. We do not use third-party analytics cookies, and we will not set any non-essential cookie without your consent where consent is required. See our Privacy Notice and Terms of Use.