ORION v2 (founder's own strategy)
- 98% deflated confidence that the edge is real, after 180 variant(s).
- The record (15.3 years) is longer than the 10.1 years needed.
Parameter tuning
The edge comes from the strategy family, not the chosen settings. The best in-sample variant is a coin flip out of sample (54% overfitting probability), but it stayed profitable in 100% of out-of-sample tests. Expect results like the typical variant (Sharpe 0.67), not the best one.
Performance as presented
Growth of $1 (log scale, blue) and drawdown from the previous peak (red).
Is the edge real?
| Test | Result | What it means |
|---|---|---|
| Probabilistic Sharpe ratio | 99.8% | Chance the true Sharpe is above zero, ignoring how many variants were tried. |
| Deflated Sharpe ratio | 97.8% | The same, after correcting for 180 variant(s) tried and for fat tails. |
| Luck threshold (Sharpe) | 0.22 | The best Sharpe that 180 skill-less variants would show by chance. |
| Minimum track record | 10.1 yrs | Years of data needed to beat the luck threshold with 95% confidence. This record: 15.3 yrs. |
| Skew / kurtosis | -0.44 / 6.3 | Negative skew and kurtosis above 3 mean crash risk the Sharpe ratio hides. |
| Probability of backtest overfitting | 53.9% | Share of 12,870 in/out-of-sample splits where the in-sample winner fell below the median. |
| Out-of-sample loss rate | 0.0% | How often the in-sample winner lost money out of sample. |
If more variants were tried
Confidence the edge is real, depending on how many versions were tested before this one.
| Variants tried | Confidence |
|---|---|
| 1 | 99.8% |
| 10 | 99.1% |
| 100 | 98.1% |
| 1,000 | 96.9% |
Next 5 years, resampled
5,000 paths built from blocks of this strategy's own history.
| Annual return, bad case (5th pct) | -8.9% |
| Annual return, typical (median) | 30.9% |
| Annual return, good case (95th pct) | 88.2% |
| Chance of losing money | 11% |
| Typical worst drawdown | -58% |
| Bad-case worst drawdown | -82% |
Method
Deflated and probabilistic Sharpe ratios and minimum track record follow Bailey & López de Prado (2012, 2014). Overfitting probability uses combinatorially symmetric cross-validation (Bailey, Borwein, López de Prado & Zhu, 2017). Sharpe-based tests use returns in excess of the stated risk-free rate. Results describe statistics of the returns supplied; they cannot detect look-ahead bias, survivorship bias or costs missing from those returns. This report is not investment advice.