Rules Are a Risk Control, Not a Crystal Ball

Most people arrive at systematic investing through the promise of a better forecast. A rule, tested over twenty years of data, that would have turned one unit of capital into many. The implicit claim is predictive: the model sees something.

That claim is almost always the weakest part of the proposition, and it is also the part that decays fastest. Edges are competed away. Data becomes cheaper. What looked like a discovery in 2015 is an execution detail by 2025. If the only reason to hold a rule is that it predicts, you will abandon it the first time it stops predicting — which is to say, quite soon.

There is a much better reason to be systematic, and it survives the edge disappearing entirely.

Consider the decisions that do the most damage to real portfolios. Almost none of them are analytical failures. They are timing failures, and they cluster in two places: the moment after a sharp fall, and the moment after a long rise.

• —Suspending contributions during a drawdown — cutting purchases at the only prices that were genuinely lower.
• —Increasing position sizes after a strong run, when the risk in every holding is higher and the margin of safety is thinner.
• —Abandoning a diversifying holding because it has 'done nothing', which is the observable signature of a holding doing its job.
• —Selling the whole position because the *narrative* deteriorated, in a week when nothing in the accounts changed.

A written rule does not make any of these decisions better. It makes them unavailable. That is a different and more durable kind of value than a forecast, because it does not depend on the rule being right — only on it having been written down in advance.

Any historical dataset can be made to yield an attractive strategy, and the process for doing so requires no dishonesty at all — only persistence. Try enough parameter combinations and one of them will look excellent, purely by chance. The resulting chart is indistinguishable from a real discovery.

There is an honest limit to all of this, and it is worth stating plainly rather than glossing over. A rule cannot tell you when it has stopped working.

Every systematic strategy has periods of underperformance that are statistically indistinguishable from the strategy being broken. The decision about whether a drawdown is normal variance or structural decay is a judgement, and it cannot be delegated to the system, because the system is the thing under examination.

1. 1

   Define the retirement condition in advance

   At the same time you write the rule, write what would cause you to stop using it — stated as a magnitude and a duration, not as a feeling. 'Worse than the historical worst drawdown, sustained beyond the historical worst recovery period' is a defensible formulation. 'When it stops feeling right' is not.

2. 2

   Separate decay from variance with a mechanism, not a mood

   The most common failure is retiring a strategy at the bottom of a normal drawdown, which converts an ordinary bad period into a permanent loss and simultaneously forfeits the recovery.

3. 3

   Keep the system small enough to fail safely

   Any rule you are prepared to run should be sized so that its complete failure is survivable. A systematic allocation large enough that its decay would be catastrophic is not a systematic allocation — it is a concentrated bet on a model.

Strip systematic investing of the marketing and a durable core is left, and it is available to anyone, with or without a model.

1. 1.Decide in advance. The quality of a decision made in calm conditions is structurally higher than the same decision made under stress, and pre-commitment is the only way to transfer the first into the second.
2. 2.Write it down. An unwritten rule is a preference, and preferences are renegotiated the moment they become expensive.
3. 3.Apply it consistently, including when it is uncomfortable — which is the only circumstance in which it was ever going to add value.
4. 4.Review it on a schedule you set, not on a schedule the market sets by making you anxious.

None of those four require a signal, a dataset or a model. They require a document. That is a considerably lower barrier than the industry implies, and considerably more of the benefit than the industry admits.