QuantHard

3 min read · 533 words

What the seat actually does

A systematic strategy is a rule: given these inputs, hold this. It is researched, tested against history, implemented in code, and then followed — including on the days when following it feels wrong, which is the entire point.

The hard part is not finding a pattern. Anything can be found in enough data. The hard part is establishing that the pattern is a return for bearing something rather than an artefact of the search, and that it survives the cost of actually trading it.

• Signal research — a hypothesis first, then a test. The other order is how backtests lie.
• Portfolio construction — turning signals into positions under constraints on risk, turnover and concentration.
• Execution — the cost of trading, which decides whether a strategy that works on paper works.
• Monitoring — whether the live results resemble the test, and what to do when they stop.

A day, and where it goes

• The overnight run — data in, signals computed, target portfolio produced, differences to trade.
• Execution — spread over the day to reduce impact, because being the flow is a cost.
• Research — the majority of the time, on ideas most of which will be discarded.
• Attribution — did the return come from the signal, from implementation, or from a factor nobody intended to own.

What it is measured on

• Risk-adjusted return, and its stability across periods rather than its level in the best one. See risk measures.
• Capacity — how much money the strategy holds before its own trading destroys the edge. Many good strategies are small.
• Turnover against cost. A signal that decays in two days is a signal you must pay to harvest.
• Live versus backtest. The gap between them is the most informative number the seat produces, and the least discussed.

What it touches on this site

• How prices are made — market microstructure, which is where execution cost comes from.
• The statistics — risk measures and the arithmetic of drawdowns.
• Why people behave predictably — behavioural finance, which is what a great many signals are ultimately harvesting.
• When the models agree with each other — 2010 and 2018.

How it goes wrong

• Overfitting. A hundred tested ideas produce five that look significant by chance alone, and the five are the ones that get funded.
• Survivorship and look-ahead in the data. A universe containing only the companies that still exist, or a fundamental figure used before it was published, makes any strategy look wonderful.
• Costs assumed rather than measured. A strategy whose edge is smaller than its spread is a spreadsheet, not a business.
• Crowding. Everybody running similar models deleverages on the same day, which is a risk no backtest contains.

Concepts to master

• Hypothesis before test. A reason the return should exist, stated first, is the only real defence against overfitting.
• Out-of-sample means untouched, and a sample looked at twice is in-sample.
• Every strategy has a capacity, and it is a property of the market rather than of the manager.
• Implementation shortfall is the honest cost measure — the difference between the price when the decision was made and where it was actually done.