Out-of-sample testing: inspect a frozen NVDA strategy

TL;DR

The frozen NVDA optimizer winner gained 22.06% on April to June 2024 dates, but bought once and never sold, turning a daytrading idea into an overnight holding. It failed four of five predeclared requirements, including activity, drawdown and the benchmark comparison. Holding NVDA gained 36.00%. Keep the selected rules and failed behavior in the record; a profitable later return alone does not show that optimization preserved the strategy's purpose.

A profitable result can still fail your strategy’s purpose

Step 1 of 5

Freeze what the optimizer actually selected

The original seed bought above a 20-minute average during the morning and sold later that session. The optimizer changed the entry to a 789-minute equality test at market open and the exit to a position-value threshold far above the initial capital. A familiar portfolio name did not preserve the original behavior.

What you should see

Both complete strategies remain in the record. The selected exit needs a position worth more than $378,389.83 before it can sell.

Step 2 of 5

Keep selection dates separate from the later test

Training developed candidates; validation helped choose among the final six. Neither is the final test. We froze the selected configuration before the April-to-June check. The intervening dates are unused in this displayed plan, not an invented embargo setting.

What you should see

The later test follows both selection periods and uses unchanged rules. These public historical dates are now reproduction dates for readers.

Frozen genetic version 2024-06-28, 20:00:00 EDT

12,205.98

Original morning seed 2024-06-28, 20:00:00 EDT

8,980.94

NVDA exported comparison 2024-06-28, 20:00:00 EDT

13,600.01

Frozen genetic versionOriginal morning seedNVDA exported comparisonHistorical simulationDates shown in New York time

Step 3 of 5

Compare the frozen version with both alternatives

The selected version gained 22.06%; the original morning seed lost 10.19%. NVDA buy and hold gained 36.00%. The selected version’s profit alone did not establish an improvement over holding NVDA.

What you should see

All three curves cover the same later dates and start at $10,000. The seed and selected version are separate completed ordinary backtests.

Step 4 of 5

Check the trades, not just the green return

The complete Filled-order query returned one NVDA buy and no sells. Shares remained held across sessions. The version passed positive return but failed the benchmark, drawdown, trading-activity and no-overnight requirements, all set before inspecting the later result.

Filled orders

1 buy / 0 sells

Closed round trips

0; required at least 10

Maximum drawdown

12.73%; required at most 10%

Observed fee

$2.98

Holding behavior

Shares remained held to test end

Requirement decision

1 pass / 4 failures

What you should see

Four of five requirements fail. A profitable holding position missed the daytrading goal. This does not isolate overfitting as the sole cause.

Step 5 of 5

Inspect the two frozen strategies in Aurora

Open the request carrying both complete configurations and the predeclared checks. Review it, then send. It asks for costs before any reproduction. For your own unseen check, freeze the rules before looking at reserved dates. If you retune after a disappointing result, keep it in the record and choose another untouched period.

View saved optimization results →

What you should see

Aurora receives both exact strategies and the later-test plan. Opening the request does not create a portfolio, run research or send a message.

Inspect the frozen NVDA experiment in Aurora →

We chose a genetic NVDA version on January to March 2024, froze it, then tested April to June. It gained 22.06%, but bought once and never sold. The later test exposed a change from daytrading to holding shares.

Plan a later test of your own strategy

Use the guide below to freeze your own rules and reserve dates that have not affected their selection. The dates in this completed case are already public reproduction dates.

Keep the failed version and its full record

The selected candidate’s validation participation was zero. A top rank among six finalists did not establish profitable validation trades. The final leaderboard does not count every unique candidate evaluated during the search.

These later dates were excluded from this search, but the historical experiment cannot prove the prices were unknown to everyone. A chronological split alone cannot establish untouched information. Keep the failed objective rather than changing its requirements after seeing the gain.

Inspect a completed validation experiment

This case examines what a frozen NVDA strategy actually did on later dates, including the daytrading requirements it failed. Use the practical guide below when you want to run an unseen test on your own saved strategy.

Continue exploring