Does this thing actually work?
Realized forward returns of past candidates, bucketed by composite-score decile (10 = strongest signal). Computed from real cached bars — never modelled.
| Composite decile | Candidates | avg +5d % | avg +20d % |
|---|---|---|---|
| 10 | 672 | +2.29 | +11.97 |
| 9 | 664 | +1.43 | +7.15 |
| 8 | 646 | +1.69 | +5.49 |
| 7 | 648 | +1.15 | +3.74 |
| 6 | 611 | +1.13 | +2.44 |
| 5 | 608 | +0.83 | +3.55 |
| 4 | 618 | +0.60 | +2.65 |
| 3 | 665 | +1.08 | +4.09 |
| 2 | 691 | +0.35 | +1.31 |
| 1 | 715 | -0.69 | +1.83 |
If the top deciles don't outperform the bottom, the signal isn't working — that is the point of this page.