Data snooping bias trading is what happens when you reuse the same dataset to generate hypotheses and to confirm them. P-hacking trading strategy is the special case where you keep testing until something is "significant."

How snooping shows up

Running many strategies and publishing the best Sharpe
Tweaking indicators after seeing the equity curve
Quietly changing the sample period when results disappoint

Habits that reduce snooping

Pre-register your hypothesis: instruments, timeframe, costs, success criteria
Hold out a true validation path (In-sample vs out-of-sample)
Report fragility and stability, not only peak performance (PSI)

Connection to multiple testing

Teams and incentives

Snooping is not only an individual mistake. Organizations reward "positive results." The fix is process: versioned notebooks, immutable datasets for a study, and a culture that publishes negative findings internally.

Pre-registration in plain language

You do not need an academic journal. You need a dated note: what you will test, what counts as success, what costs you assume, and what you will not change after you see the curve. That single habit removes a surprising amount of hidden multiple testing.

Experiment registry (minimum viable)

Create a table with one row per run:

run id, date, author
hypothesis
dataset snapshot id
parameter search budget
outcome metrics (including failures)

If you cannot reconstruct your research trail, you cannot defend your conclusions.

"Research mode" vs "confirmation mode"

Split tooling or folders so confirmation runs cannot accidentally reuse the same workspace state as exploratory runs.

The goal is to make cheating the process slightly harder than following it.

Benchmarks are hypotheses too

If you tune against a benchmark curve you peeked at repeatedly, you imported snooping through the side door.

Treat benchmark selection like instrument selection: choose it early, write it down, and hold yourself to it.