Study · Performance Attribution · July 2026
Return Concentration: Where the Book's Gains Actually Came From
The Concentration Premium asked whether this book's position sizing is really as diversified as its HHI implies. This study asks the same question about the returns themselves: of the roughly 32 positions in this book, how many are actually responsible for the gains, and are those positions independent bets or the same correlated trade paying off six different ways? The answer uses nothing but AEA's own published Holdings weights, returns, and correlation matrix — no new data, just a different cut of it.
1. A simple attribution, honestly bounded
The Return contribution section of the Holdings page ranks each position by today's allocation weight × its total return since cost — a simplified attribution, not a precise time-weighted decomposition, and the page says so directly: weights shift as positions compound, so this ranking doesn't reconcile exactly to the account's own +72.14% all-time dollar-weighted return. With that caveat stated once and carried through everything below, here are the top ten contributors by that measure:
| Ticker | Alloc % | Return % | Contribution (pp) |
|---|---|---|---|
| SNDK | 7.94% | +151.41% | +12.02 |
| META | 8.42% | +137.59% | +11.59 |
| AMD | 7.78% | +138.86% | +10.80 |
| MU | 5.38% | +152.59% | +8.21 |
| INTC | 4.13% | +140.98% | +5.82 |
| PANW | 5.42% | +106.98% | +5.80 |
| DDOG | 2.63% | +124.68% | +3.28 |
| CAKE | 2.65% | +123.39% | +3.27 |
| NBIL | 2.06% | +128.17% | +2.64 |
| FTNT | 1.59% | +92.11% | +1.46 |
These ten positions sum to roughly 64.9 of the ~76.3 total percentage points of positive simplified contribution across the book — about 85% of the gross gain, from 31% of the position count. That alone is a normal, expected shape for a concentrated book; the more interesting question is whether the biggest of these contributors are actually different bets.
2. The top six aren't six different ideas
Pull the six largest contributors — SNDK, META, AMD, MU, INTC, and PANW — and check every pair against AEA's own published correlation matrix. Below is the complete set of 15 pairs among those six names.
| Pair | Correlation |
|---|---|
| SNDK ↔ MU | 0.72 |
| AMD ↔ MU | 0.53 |
| MU ↔ INTC | 0.44 |
| SNDK ↔ INTC | 0.39 |
| AMD ↔ INTC | 0.36 |
| SNDK ↔ AMD | 0.25 |
| META ↔ MU | 0.22 |
| META ↔ AMD | 0.18 |
| INTC ↔ PANW | 0.14 |
| META ↔ INTC | 0.10 |
| SNDK ↔ META | 0.03 |
| SNDK ↔ PANW | −0.03 |
| AMD ↔ PANW | −0.02 |
| MU ↔ PANW | −0.12 |
| META ↔ PANW | −0.22 |
The average across all 15 pairs is 0.20. That headline number hides the real shape, though: four of these six — SNDK, AMD, MU, and INTC — are a genuine semiconductor/memory cluster, correlated 0.25–0.72 with each other. NBIL, the seventh-ranked contributor, belongs in that same cluster too: it correlates 0.34–0.58 against all four semiconductor names. The other two of the top six, META and PANW, both sit near-zero or negative against that cluster and against each other (−0.22) — they aren't part of it, and they aren't a pair either.
3. Two real exceptions, and a second cluster hiding lower in the list
META is the second-largest contributor at +11.59pp, and it is genuinely independent — its correlation to the four-name semiconductor cluster ranges from 0.03 (SNDK) to 0.22 (MU), low enough that the gain reads as its own outcome, not the same trade wearing another ticker. CAKE (+3.27pp, ranked 8th) is mostly independent too, with one exception: it correlates 0.37 with AMD, a tie worth naming rather than ignoring.
PANW looks independent sitting inside the top six alone, but it isn't independent of the book as a whole: PANW, DDOG (+3.28pp), and FTNT (+1.46pp, ranked 10th) form a second, tightly correlated cluster — PANW–DDOG at 0.67, PANW–FTNT at 0.72, and DDOG–FTNT at 0.82, all enterprise-software names benefiting from the same platformization and AI-tooling demand rather than three separate stories.
Netting it out: of the ~64.9pp of simplified contribution in the top 10, roughly 39.5pp (SNDK, AMD, MU, INTC, NBIL) comes from one tightly correlated semiconductor/memory cluster, 10.5pp (PANW, DDOG, FTNT) from a second, smaller correlated cluster, and only META's 11.6pp and CAKE's 3.3pp read as genuinely independent of both. Restated plainly: most of this book's gain this cycle came from two correlated clusters paying off together, not from a wide spread of independent calls being right.
This is the same conclusion as The Concentration Premium, from the other direction
That study showed the book's position sizing behaves like roughly 5 effective independent bets instead of 32, once correlation is priced in. This study shows the book's actual gains tell the same story from the return side: the positions that made the most money are largely the same correlated cluster, not a diversified list of separately-right calls. Two different calculations, same book, same answer — which is itself a form of confirmation, not just repetition.
4. What this does and doesn't mean
It does not mean the thesis was wrong — the semiconductor/AI-infrastructure buildout call has, so far, been the correct call, and the IPS explicitly endorses concentrated, high-conviction positioning over shallow diversification. What it does mean is that this book's track record so far is much closer to "one macro call executed well across five vehicles" than "five independent research successes." That distinction matters for judging the record going forward: the next test isn't whether SNDK, AMD, MU, INTC, and NBIL all keep working — if the underlying AI-capex/memory-pricing thesis breaks, they are likely to break together, not one at a time. The AI Capex Reality Check and Data Center Alley studies already examine that shared thesis directly; this study is the reminder that the book's actual results, not just its position sizing, are riding on it.
Methodology & limitations
Return contribution. Contribution (pp) = current allocation % × total return % since cost, taken directly from AEA's Holdings page. This is a simplified attribution using today's weights, not a time-weighted performance bridge — it will not sum exactly to the account's own +72.14% all-time dollar-weighted return, and the Holdings page states this limitation directly.
Correlation data. All 15 pairwise correlations are taken directly from the full pairwise Pearson correlation matrix (daily returns, year-to-date) already published on the Holdings page, restricted to the 6 largest contributors identified in Section 1. This is a targeted subset, not a full-book average — the 0.173 book-wide comparison figure is itself a 3-row, 90-pair sample from The Concentration Premium, not a complete 496-pair average.
NBIL's leverage. NBIL is a 2× daily-reset leveraged ETF tracking NBIS; Section 2 groups it with the semiconductor cluster because it correlates 0.34–0.58 against those four names, not because it tracks the same underlying company — the leveraged-sleeve mechanics themselves are covered separately by the Risk X-Ray tool.
What this study does not claim. It does not claim the underlying thesis is wrong, and it does not recommend any change in position sizing. It restates, from the returns side, the same structural finding The Concentration Premium reached from the position-sizing side.
Not investment advice. Nothing here is a recommendation to buy, sell, or avoid any security.