This article is written as a field report pattern, not as a claim about a specific proprietary dataset. The value is in the recurring patterns teams see when they move from one-off backtests to disciplined walk-forward batches.

Pattern 1: most strategies fail on costs, not on signals

When the same logic is re-run with stressed slippage and fees, a large share of "good" strategies collapses. The survivors usually share lower turnover or cleaner execution assumptions.

Lesson: validate costs early, not as a last step.

Pattern 2: hyperparameter-heavy strategies rarely survive first OOS

Complex search spaces routinely produce great IS scores and weak OOS stability.

Lesson: track search budget explicitly and penalize fragility.

Pattern 3: the best-looking equity curves often hide regime concentration

A strategy can be "robust" statistically while depending on one macro regime.

Lesson: add regime-tagged reporting for each OOS window.

Pattern 4: stable mediocre often beats brilliant fragile

Teams that deploy capital often pick the boring strategy with stable OOS degradation over the flashy strategy with cliff risk.

What we would measure next in a real batch

distribution of WFE across strategies
correlation of failure modes (liquidity vs trend vs costs)
sensitivity of rankings to slippage tier

Methodology note: apples-to-apples batching

If you run a real batch study, lock:

identical fee and slippage models per strategy class
identical universe rules
identical walk-forward window machinery

Otherwise your leaderboard compares strategies that were not tested under the same constraints.

What not to conclude from a leaderboard

A batch study can rank candidates, but it cannot replace deployment discipline. The output should be a smaller set of candidates for deeper review, not immediate capital allocation.

Publish failure modes, not only winners

If a batch study only circulates the top decile, the organization learns the wrong lesson.

Keep an internal archive of why strategies died: costs, fragility, regime, data.