Published 2026-10-27
Interpretability
Being able to say why a decision was made. Without it you cannot tell a strategy that learned a pattern from one that memorised noise.
What it is
Interpretability is the ability to state why a decision was made, in terms that can be checked.
Not “the model scored it highly”. Not “the signal fired”. A reason that another person could examine and disagree with.
Two layers
| Layer | Question it answers | Failure looks like |
|---|---|---|
| Model | Which inputs moved the output, and how much | “It is a black box” |
| Strategy | Why this rule should work at all | “It worked in the backtest” |
Both are required, and the second is the one most often skipped. A perfectly interpretable model can still sit inside a strategy nobody can justify — for example, one built by trying 500 rules and keeping the winner.
Why it matters
Interpretability is the only defence against overfitting that works before the money is lost.
If you can say why a rule should work, you can ask whether the reason survives outside the sample. If you cannot, the backtest result is the entire evidence, and a backtest result can always be produced by searching hard enough — see data snooping.
“It worked in the backtest” is not a reason. It is a measurement. The two are easy to confuse because the second one is a number and numbers feel like proof.
How to ask for it
Require a chain, not a conclusion:
- What is the claim? One sentence, falsifiable.
- Why would it be true? A mechanism, not a correlation.
- What would make it false? Conditions under which you would abandon it.
If a strategy cannot answer all three, the honest description is “unknown”, not “promising”. This is also why the standards require a stated failure condition rather than a track record.