Insights

Occasional notes on risk, process, and discipline. Not market commentary, not a forecast, and not a description of any specific strategy.

Sep 2026

On drawdown, and why the plan comes before the loss

Most discussion of risk management centres on the rules themselves — position limits, stop levels, correlation caps. Less is said about timing. A rule decided in the middle of a loss is not a rule; it is a negotiation with the loss, and the loss usually wins.

The distinction matters because the two situations produce identical-looking numbers with entirely different meanings. A drawdown limit fixed before any capital is committed is a constraint the future self cannot argue with. The same number, arrived at after several difficult weeks, is simply where the discomfort became unbearable — a description of feeling, not a decision about risk.

This is why limits are set once, calibrated to volatility and correlation at the point a strategy is deployed, and are not revisited in response to how that strategy is performing. Discipline, in this sense, has very little to do with mathematics. The maths is the easy part. The discipline is refusing to hold a second conversation about a number that was only ever meant to be discussed once.

Sep 2026

Why backtested doesn't mean true

A backtest is not evidence. It is a hypothesis given the appearance of evidence, produced by fitting a rule to a history that has already happened and can never happen differently. The past does not confirm a strategy works; it confirms only that the strategy would have worked on that one specific version of the past.

The risk grows with the number of choices made along the way — which parameters, which universe, which period, which adjustment after the first result disappointed. Each choice is a chance to find something that fits the data by coincidence rather than by any genuine relationship, and a process with enough free parameters will usually find something. The result can look excellent and mean nothing.

The correct posture toward a promising backtest is closer to suspicion than celebration. A result that survives being tested for exactly this failure — on data it has never seen, under conditions it was not built to fit — is worth more than a result that impresses at first glance. Testing exists to eliminate ideas, not to justify the ones already liked.

Sep 2026

Correlation is not diversification

Diversification is often treated as an exercise in counting: enough names, enough sectors, enough geographies, and the risk is assumed to have been spread. But diversification is a claim about how positions behave relative to one another, not about how many of them exist.

Correlation is not a fixed property of two assets; it is a condition, and it changes with the environment. Positions that move independently in calm markets can move together sharply under stress — often at precisely the moment protection is needed most. A portfolio that looks diversified on a spreadsheet can behave, in practice, like a single position.

Treating correlation as something that shifts with regime, rather than a number measured once and trusted indefinitely, is a different kind of work than simply holding more things. The difference between the two is invisible until the exact moment it matters, which is also the worst possible moment to discover it.