Walk-forward efficiency (WFE) summarizes how much of your in-sample edge survives out-of-sample. Thresholds like 0.7 and 0.5 are useful communication shortcuts, but they are not laws of physics.

Treat them as guardrails that must be justified with window design and economic context.

What WFE is trying to measure (intuitively)

If you optimize on IS and then evaluate on OOS, you expect degradation. WFE is a normalized way to ask:

"Did a meaningful fraction of the IS edge remain under honest OOS conditions?"

When 0.7 can be a PASS

A higher threshold makes sense when:

you have enough independent windows (not two lucky splits)
costs are modeled conservatively
parameter search intensity was controlled

In that world, 0.7 indicates the strategy is not entirely a curve-fit mirage.

When 0.5 can still be ACCEPTABLE

A lower threshold can be acceptable when:

the strategy is intentionally low Sharpe but stable (carry-like profiles)
execution noise dominates short OOS segments
you pair WFE with strong stability evidence (PSI, narrow parameter plateaus)

The key is that "acceptable" means "eligible for next stage review," not "deploy full size."

What breaks threshold reasoning instantly

tiny trade counts
different cost assumptions between IS and OOS
optimizing on the entire series then slicing for display

Practical recommendation

Use thresholds as sorting metrics inside a checklist, not as a single boolean.

Report a distribution, not a headline

A single WFE number from one run is fragile.

Prefer reporting WFE across windows (median, interquartile range, worst window) so you can see whether one lucky slice is carrying the verdict.

Asset class and turnover change the bar

High-turnover crypto strategies often degrade faster under realistic costs than slow equity trend systems.

If your turnover is high, thresholds should move in the direction of demanding stronger cost stress and more windows, not looser headline numbers.

Pair thresholds with economic minimums

WFE can look fine while absolute OOS profit is too small to matter after fees.

Add minimum trade count and minimum net expectancy checks so you do not green-light statistically plausible but economically useless edges.