When to change a rule
Lesson 17 · about 9 min
A plan that never changes is a plan that stopped learning. A plan that changes every week is not a plan. The gap between those is a set of thresholds: how many trades, how large an effect, and how much of the difference could be noise. This lesson sets them.
The two failure modes
Churn. A bad week produces a rule change. The next week is different, so it produces another. After three months the plan has been through eleven versions, none with more than fifteen trades, and the journal cannot say anything about any of them. This is the common failure; it feels like diligence and it is the opposite.
Calcification. The plan was written once, the trader is proud of it, and the numbers slowly go wrong while every review says "keep going, it's variance". Twelve months in, the edge has been gone for eight of them.
The fix for both is the same: rules change only at scheduled reviews, only when a minimum sample supports the change, and only when the effect is large enough that noise is an unlikely explanation.
Minimum samples
| Change proposed | Minimum trades in the relevant subset | Why |
|---|---|---|
| Execution rule (routine, checklist, breaker) | None; any weekly review | Does not affect what the setup is |
| Exit rule (target, trail, time stop) | 30 trades on the current rule | Exits are measurable against planned R on every trade |
| Filter threshold | 30 trades on each side of the threshold | You are comparing two groups; both need a sample |
| Setup criterion (add, remove, change a box) | 50 trades on the current setup | Changes what is being measured; resets the sample |
| Risk numbers (1R, heat, limits) | Per the risk plan's scaling rules | Governed by drawdown and months, not trades |
| Adding a second setup | 100 trades on the first | The first setup must be understood before a second dilutes attention |
"Trades in the relevant subset" matters. If you want to move the filter from 0.4% to 0.5%, the question is how the trades between 0.4% and 0.5% performed, and you need thirty of those, not thirty in total.
How large is large enough
With thirty trades in a group, the average R has a standard error of roughly the spread of your R outcomes divided by the square root of thirty, which for a typical setup with a spread of about 1.5R comes to about 0.27R. That means a difference between two groups of less than about half an R per trade is inside the noise and should not move a rule.
Practical thresholds:
- A subgroup with 30 or more trades and an average R more than 0.5R worse than the rest is a candidate for exclusion (tighten the filter).
- A subgroup with 30 or more trades and an average R more than 0.5R better than the rest is interesting, and is not on its own a reason to trade only that subgroup; it is a reason to keep logging and see whether it holds at 60.
- Anything under 0.3R difference is not a finding. Log it, wait.
These are not rigorous statistics; they are a rule of thumb sized so that a real effect gets through and a run of luck does not. If you want something closer to rigorous, the next lesson shows how to compare two rules directly with a prediction and a decision threshold written in advance.
Key idea: Thirty trades in the relevant subset, and a difference of at least half an R per trade, before any exit or filter rule changes. Fifty for a setup criterion. Anything smaller is noise, and noise is not a reason.
What can change at which review
| Review | May change |
|---|---|
| Weekly | Execution actions only. Nothing on the plan. |
| Monthly | Exit rules, filter thresholds, breakers, with the samples above. |
| Quarterly | Setup criteria, markets, sessions, adding a setup, risk numbers. |
This is the calendar side of the threshold. Even with the sample in hand, a setup criterion waits for the quarterly review. The wait costs a few weeks and buys the certainty that the change was considered by someone who was not reacting to a specific week.
Execution before plan
Before any plan rule changes, check the plan-follow scores for the period. A plan that was followed at score 3 on 60% of trades has not been tested; the results are a blend of the plan and of the deviations. The first change in that situation is always an execution action, never a plan rule. Get plan-follow to 85% or better, then evaluate the plan.
Most "the setup isn't working" conclusions in the first six months turn out, on inspection, to be "the setup was taken correctly 55% of the time, and those trades were fine."
The candidates list, revisited
Every idea goes on the list with a date, a reason and the current trade count in the relevant subset. At the monthly review, each candidate is either: still short of sample (wait), at sample but under threshold (drop or keep watching), or at sample and over threshold (test it, next lesson). Most candidates get dropped. That is the list working.
Try it: Take your candidates list, or write down three changes you have been tempted to make. For each, identify the relevant subset, count the trades in it, and compute the average R difference against the rest. Mark each one wait, drop or test. If all three are "wait", that is the normal answer for the first two months.
Recap
- Rules change only at scheduled reviews, with a minimum sample in the relevant subset, and only for a difference large enough to be unlikely noise.
- Thirty trades for exit and filter rules, fifty for setup criteria, a hundred before a second setup.
- Half an R per trade between subgroups is the practical threshold. Under 0.3R is not a finding.
- Weekly reviews change execution only; monthly reviews change exits, filters and breakers; quarterly reviews change setup, markets and risk.
- Check plan-follow first. A plan followed 60% of the time has not been tested, and the fix is execution, not the plan.
See it drawn
Original diagrams for the ideas on this page. Illustrative, not real market data.