Systematic vs Discretionary Trading

Every trade comes down to one question: who decides? In a discretionary book, a person looks at the situation and decides — this position, this size, today. In a systematic book, the decision was made months ago, written down as a rule, and today a machine applies it to whatever the data says. Both make money, in structurally different ways. Confusing them is how people end up with a "systematic" strategy that only works when someone overrides it.

The dividing line is the decision, not the computer

Discretionary traders use screens, models and Python. Systematic funds employ humans who research signals, argue about them and switch them off. The distinction is not tools. It is whether the rule that fires the order was fully specified before the data arrived, and applied without exception.

Two tests settle it in practice:

• The holiday test. The PM goes offline for three weeks. Does the book keep trading, rebalancing and hedging? If yes, it is systematic.
• The reproducibility test. Given the historical data, can you regenerate every position the strategy would have held, with no human input? If yes, the strategy has a backtest. If no, it has only a track record — a much weaker thing, because you cannot ask it counterfactual questions.

Breadth is the whole argument

The reason systematic trading exists is arithmetic, not ideology. Grinold's fundamental law of active management says that risk-adjusted performance scales as

$IR \approx IC \times \sqrt{BR}$

where $IR$ is the information ratio (return per unit of risk), $IC$ is the information coefficient — the correlation between your forecast and the return that actually happens — and $BR$ is the number of independent bets you take per year. Skill per call and number of calls trade off against each other, and the number of calls enters as a square root.

Work it through with two realistic books.

A discretionary macro PM. Twenty genuinely independent positions a year, each one researched for weeks. Suppose the forecasts are good: $IC = 0.10$, which is high for anything in markets. Then $IR = 0.10 \times \sqrt{20} = 0.45$.

A systematic equity book. Fifteen hundred names, rebalanced monthly, but the bets are far from independent — say 800 effective independent bets a year after accounting for shared factor exposure. The per-name skill is feeble: $IC = 0.04$. Then $IR = 0.04 \times \sqrt{800} = 1.13$.

The systematic manager is worse at every single call, by a factor of two and a half, and still ends up with more than double the risk-adjusted return. That is the entire case for systematic trading in one line.

It helps to see what an $IC$ of 0.04 actually looks like. Set the correlation slider near zero and look at the cloud: that is a real equity signal plotted against next month's return.

Now drag it up to 0.4. That is what people imagine an edge looks like. No cross-sectional equity signal has ever sustained it.

Two real books

Discretionary, September 1992. Soros's fund, run day-to-day by Stanley Druckenmiller, built a roughly $10bn short in sterling against the ERM band the Bank of England was defending. The insight was political, not statistical: Britain could not hold the parity and run its domestic economy at once. On 16 September the Bank took base rate from 10% to 12%, announced 15%, then left the ERM. The fund made about $1bn on a single position, with no backtest and no way to have built one.

Systematic, any month you like. Rank 2,000 US names on twelve-month momentum and book-to-price, neutralise the scores within sector, go long the top decile and short the bottom, rebalance monthly. Any individual name is close to a coin flip — call it 52/48. The book makes money because it holds 200 names a side and redraws the hand every month, so a 2% edge is expressed 400 times instead of once.

What erodes each edge

Systematic books die of overfitting — the backtest found a pattern in the noise — and of crowding. In August 2007 nearly every quantitative equity book held the same value and momentum names; one large fund delevered, and margin calls forced the rest to delever into each other. Nominally market-neutral books lost 20–30% in three days. The rules were fine; the problem was that everyone else had the same rules.

Discretionary books die of things that are harder to measure. Attribution is unclean, so you rarely know whether last year was skill or beta. Capacity is bounded by one person's attention, not by liquidity. Style drifts — the macro fund quietly becomes a long-equity fund. And the behavioural errors a rule would have prevented, averaging into losers and cutting winners early, are the ones a person under pressure commits.

In interviews

Lead with the fundamental law and make it concrete: an $IC$ of 0.10 on 20 bets loses to an $IC$ of 0.04 on 800, because breadth enters as a square root. Define the boundary by the holiday test, not by whether code is involved. Then name the failure modes on both sides — overfitting and crowding for systematic, attribution and capacity for discretionary.