AI-assisted research continues to raise the bar.

Building on the framework already in place, AI proposes candidates, runs experiments, and iterates on the results. The development record shows how that process has changed the strategies over time.

Validation and holdout results are shown separately so development progress is not presented as final out-of-sample proof.

Development runs: 821
First 100 median: 4.554%
Last 100 median: 6.383%
Latest validation high: 7.097%

Progress over time

Across 821 development-eligible runs, the median validation average monthly return moved from 4.554% in the first 100 runs to 6.383% in the last 100. The record high reached 7.097%.

Highest observed validation avgM7.097%

Validation highs20

Selection statusDevelopment

Candidate validation resultRolling-median trendNew validation high5% and 8% reference lines

What it shows

Individual candidates remain uneven, but both the rolling median and the sequence of new highs move upward. That means later research produced stronger validation outcomes more often than the earlier runs. It does not show that every new candidate improved or that AI alone caused the change.

Does the improvement generalize?

Training results are useful only when they carry into unseen validation data. These views compare the two periods and then test whether higher returns also came with stronger risk-adjusted results.

Validation by training rankRuns sorted by training avgM ascending; validation avgM plotted per rank

Validation per runRolling-median trend5% and 8% reference lines

Training vs validation scatterTraining avgM on x-axis, validation avgM on y-axis; each point is one run. The line summarizes the in-sample relationship.

One development runLeast-squares trend (descriptive)5% and 8% validation reference lines

Average trendDevelopment runColor = MaxDD (dark = deep, light = shallow)

What it shows

Stronger training runs generally align with stronger validation runs, while the scatter shows meaningful variation around that relationship. Higher validation return also tends to align with higher Sharpe. Together, the data suggests that some improvement transferred beyond training and was not limited to return alone. The dispersion remains important evidence that generalization is incomplete.

The candidate frontier

20 candidates established a new validation high. The next views examine those records against training performance, drawdown, and other risk measures rather than treating the highest return as sufficient evidence.

Average monthly returnFee-adjusted NAV return

TrainingValidation

Annualized returnCompound annual growth rate, normalized for period length

TrainingValidation

Sharpe ratioAnnualized from monthly returns

TrainingValidation

Maximum drawdownNAV drawdown within period

TrainingValidation

What it shows

19 of 20 validation-record candidates finished above their corresponding training return. The records also show the trade-off between return and drawdown. Because this is a selected frontier of validation highs, it should be read as development history, not as an unbiased estimate of future performance.

Current candidate

The latest preserved candidate is shown across training, validation, and holdout. Keeping the periods separate makes deterioration visible instead of hiding it inside one headline result.

Training

6.054%

Apr 2003 to Dec 2018 · complete months only

Validation

6.674%

Jan 2019 to Dec 2023 · complete months only

Holdout

6.548%

Jan 2024 to Jun 2026 · complete months only

Run #1028

Average monthly NAV return

Mean monthly return after fees. The primary development metric; consistency across periods signals generalization.

6.05%

Training

6.67%

Validation

6.55%

Holdout

Annualized return (CAGR)

Compound annual growth rate, normalized for period length. Directly comparable across training (188mo), validation (60mo), and holdout (30mo).

91.80%

Training

101.77%

Validation

102.75%

Holdout

Sharpe ratio

Annualized return per unit of volatility. Higher means smoother returns, not just higher returns.

2.00

Training

1.88

Validation

2.20

Holdout

Maximum drawdown

Worst peak-to-trough NAV decline. Deeper bars mean larger losses endured; shallower is better.

-30.65%

Training

-26.04%

Validation

-19.71%

Holdout

On this candidate, holdout average monthly return was 6.55%, 0.13% below the validation result. Holdout Sharpe was 2.20 versus 1.88 in validation. This is one held-out check, not evidence of repeatability across future candidates.

This is a backtested research result, not a live track record or investment recommendation.