TL;DR
A high training return can select an inactive or unintended strategy. This NVDA candidate returned 38.14% in training, 0% with no participation in validation and 22.06% on frozen April to June 2024 dates. It bought once and never sold, holding overnight instead of completing the requested intraday trades. Its exit threshold exceeded $378,389 in a $10,000 account. Inspect source coverage, entry conditions and reachable exits before attributing the failure solely to statistical overfitting; return alone does not establish the intended behavior.
A top training return did not establish an active validation strategy
Show the table · 3 rows, 4 columns
Scroll sideways to compare all columns. Select a heading to sort.
| Training | 2024-01-02T00:00:00.000Z to 2024-03-06T12:00:00.000Z | 38.14% | Develop candidates |
| Validation | 2024-03-07T12:00:00.000Z to 2024-03-28T00:00:00.000Z | 0.00% | Select among final candidates |
| Frozen later test | 2024-04-01 to 2024-06-28 | 22.06% | Apply the predeclared acceptance criteria |
No matching rows. Clear the filter to see all records.
Training returned 38.14%; validation returned 0.00% with zero participation; the frozen later test returned 22.06%. The other five final candidates had negative validation returns, so the selected zero-return candidate won the validation ordering without demonstrating profitable validation trades.
These observations diagnose a selection problem; they do not by themselves prove that statistical overfitting is the only cause. Inspect inactive validation, source availability, changed entry conditions and unreachable exits before naming the failure.
Inspect the frozen later test
Portfolio value (USD)
- Genetic winner 2024-06-28, 20:00:00 EDT
- $12,205.98
- Original morning strategy 2024-06-28, 20:00:00 EDT
- $8,980.94
- NVDA buy and hold 2024-06-28, 20:00:00 EDT
- $13,600.01
- SPY buy and hold 2024-06-28, 20:00:00 EDT
- $10,371.61
Genetic winner: 12.40% largest decline in plotted observations. 2024-06-20, 04:01:00 EDT ($13,328.24) to 2024-06-25, 04:01:00 EDT ($11,675.87).
The selected candidate earned 22.06% in the later test. The stored order record contains one filled buy and no filled sells, so this positive result came with an overnight position and zero completed round trips. Compare the winner, unchanged morning seed and the stored NVDA/SPY baselines on their shared later dates.
Check whether optimization changed the strategy's meaning
Inspect the exact historical measurements
The selected exit required position value above $378,389.83 in a $10,000 account. The entry guard changed from an empty holding to position value ≥ 0. Read those changes alongside the one-buy/no-sell result: a candidate named as an intraday strategy actually held across sessions.
Scroll sideways to compare all columns. Select a heading to sort.
| Entry time | Minutes after open ≥ 30 and < 60 | Minutes after open ≤ 0 and ≤ 139 |
| Price rule | Price > 20-minute SMA | Price = 789-minute SMA |
| Position guard | NVDA position value = 0 | NVDA position value ≥ 0 |
| Buy size | 95% of buying power | $5,961.15 |
| Exit | At or after 200 minutes; position value > 0 | Exactly 213 minutes; position value > $378,389.83 |
| Sell size | 100% of the position | 86.23% of the position |
No matching rows. Clear the filter to see all records.
Separate three explanations before the next experiment
Show the table · 3 rows, 3 columns
Scroll sideways to compare all columns. Select a heading to sort.
| Validation participation 0 | Whether complete source history and entry operands existed | Count evaluated, eligible and active decisions; do not rank inactivity as successful trading |
| Only one buy and no sells | Whether the selected exit is reachable under the declared capital and price path | Require observed closed round trips and no overnight holdings for an intraday task |
| Training ranked higher than validation | How many variants, restarts and rule changes used the same validation dates | Freeze a new rule family and untouched final evaluation before selecting again |
No matching rows. Clear the filter to see all records.
No comments yet.