Four Bets, One Ticker: Why Two People Who "Own NVDA" Don't Actually Own the Same Thing

A note on hierarchical risk decomposition for fund managers, allocators, and concentrated PMs.

By Conrad Gann — Founder, RiskModels.app

Two portfolio managers tell you, on the same day, "I'm long NVDA."

You'd assume they're running the same bet. They aren't. They might be running entirely different investment theses — with different exit conditions, different correct hedges, and different right-and-wrong answers to the question "is this position working?"

The reason is structural. Every equity position is a layered combination of four exposures stacked on top of each other, and most casual conversations about positions collapse all four into one. Once you separate them, "I'm long NVDA" stops being a claim about one bet and becomes a claim about which of four bets you actually intended to take.

This piece walks through the four bets, what they mean, and why disentangling them is more important than most retail-tier risk frameworks let on.

The decomposition

Take the daily returns of NVDA over a year. We can write them as a sum of four contributions:

NVDA return  =  market exposure              (layer 1)
              + sector tilt vs market         (layer 2)
              + subsector tilt vs sector      (layer 3)
              + residual                      (layer 4)

Each layer is what's left after the layers above it have been stripped out. Layer 1 is just being long equities — what you'd get from holding SPY in proportion to NVDA's market beta. Layer 2 is the tilt you get from preferring technology to the broad market — what's left after the market component is removed, replicated as long XLK / short SPY in matched amounts. Layer 3 is the further tilt from preferring semiconductors specifically — long SMH / short XLK, again with the previous layers stripped. Layer 4 — the residual — is whatever's left after all three structural exposures are removed.

Mathematically this is just nested regressions with orthogonal residuals, and any quant with a Bloomberg terminal can run it. The interesting thing is what each layer means once you have it.

What each layer is, financially

Layer 1 — market. The bet that equities go up. NVDA delivers this regardless of anything specific to the company; it's the cost of being in stocks at all. A retired teacher with an S&P 500 index fund holds more total layer-1 exposure than most active PMs.

Layer 2 — sector. The bet that technology beats the market. This is a top-down view: macro themes, rate regimes, growth-vs-value rotations. A PM who's overweight tech is taking layer-2 risk whether or not they ever look at a single stock chart.

Layer 3 — subsector. The bet that semiconductors beat technology. This is where most "thematic" investing lives — AI infrastructure, cloud capex cycles, semi capacity build-out. A long-NVDA / long-SMH / long-AVGO sleeve is mostly a layer-3 bet dressed up as stock picking.

Layer 4 — residual. What's specific to NVDA, after the structural exposures are removed. Earnings surprises, executive moves, supply-chain wins, customer-concentration shocks. This is where selection skill lives. A manager who consistently produces positive residual returns is genuinely picking — versus a manager whose returns evaporate once you control for sector and subsector tilts.

The four layers are mathematically independent. They can rise or fall in any combination. NVDA in 2023 was unusual specifically because layers 1 through 3 were positive and layer 4 was massive — the company-specific bet paid off on top of the structural tailwinds. Most equities don't have that pattern.

"I'm long NVDA" as four different theses

Now go back to the two PMs.

PM A says: "I love AI; I'm long NVDA, and I'm not hedging." Translated: A's conviction is in layer 4. They believe NVDA-specific catalysts (data-center demand, software moat, training-cluster orders) will drive returns in excess of what semis as a category deliver. Their right-and-wrong is: did NVDA outperform AVGO and AMD on a sector- and market-neutral basis? If yes, the thesis was right. If no — even if NVDA was up 50% — the thesis was wrong; they got lucky riding semis.

PM B says: "I'm long NVDA as expression of my AI-infrastructure thesis." Translated: B is taking a layer-3 bet. They want semiconductor exposure with a slight skew toward the cleanest AI play. Their right-and-wrong is: did semis as a group beat broader tech? If yes, the thesis was right whether or not NVDA outperformed its peers. NVDA was the vehicle, not the bet.

PM C says: "I'm long NVDA because I think this market is going up and growth leadership belongs to tech." That's mostly layers 1 and 2 — a leveraged equity-plus-tech bet, and NVDA happens to be a high-beta way to express it. Hedging out NVDA-specific risk doesn't compromise the thesis; it sharpens it.

Three managers, same realized return on the same security, three completely different skill claims. The only way to evaluate any of them honestly is to run the decomposition and ask: did the layer you said you cared about pay off?

The post-mortem trap

Most position reviews in practice never do this. A manager up 28% on NVDA writes a year-end letter about their AI conviction, and investors read it as a layer-4 success story. The four-bet decomposition often tells a different story: most of the move was a layer-3 (semi) tailwind that any sector-tilted ETF captured automatically. The manager's actual layer-4 contribution might have been small, neutral, or even negative — meaning that on a risk-adjusted basis, the manager added no skill above what a low-fee semi ETF could have delivered.

This is the asymmetry: structural returns get retroactively claimed as alpha unless someone runs the math on the four bets explicitly. Allocators who don't do this work end up paying 2-and-20 for layer-2 and layer-3 exposures they could have replicated cheaply.

The mirror trap is worse. A manager who legitimately picked NVDA over its peers — generating real layer-4 alpha — but in a year when semis underperformed broader tech can show a flat or negative result and look unskilled. Their actual stock picking was excellent; the thematic backdrop was wrong; the headline number tells neither story.

What disentangling buys you

Once the four bets are explicit, several questions become answerable that aren't otherwise:

• Which of my positions are pure thematic exposure dressed up as stock picking? — find positions where layer-3 tilt explains most of the return.
• Which of my managers consistently generate residual? — layer-4 returns persist across regimes; layer-1/2/3 returns are environment-dependent.
• What's the correct hedge if I want to keep my actual conviction? — hedge the layers you didn't intend to take, leave the layer you did.
• How does this position compare to its peer cohort? — on a layer-4 basis. That's the only basis on which "comparable" is meaningful, because it strips out the structural exposures everyone in the cohort shares.

The work is straightforward in principle. It's also straightforward to get wrong: layer definitions have to be consistent across positions, the decomposition has to be orthogonal at each level, peer cohorts have to be cap-weighted properly, and historical comparisons have to respect when each input was actually publicly knowable on the simulated date.

That last point is its own essay.

Why this matters more now

Concentrated portfolios — including the family-office and concentrated-PM segment that's increasingly important post-2022 — make this decomposition more urgent, not less. When 30% of a portfolio is in a single name, "I'm long NVDA" can mean half a thesis. The four bets behind that one ticker probably represent four different decisions, four different time horizons, and four different reasons to exit. Conflating them is how concentrated PMs end up taking risks they didn't intend, hedging exposures they wanted to keep, and selling positions for reasons that have nothing to do with their original conviction.

This is the work we do at RiskModels.app for fund managers, allocators, and concentrated PMs. We publish position-level analyses that make the four bets explicit on every name, with peer-cohort comparison, residual quality scoring, and hedge-ratio decomposition out of the box. If you've ever stared at a "+18% YTD" in a tech fund and wondered which of the four bets really paid off, you already understand why this is a question worth answering precisely.

Same ticker, different bet. The label tells you almost nothing.

If this resonates and you'd like to see worked examples on real positions — including the cases where the four bets disagree dramatically — the RiskModels.app snapshot library is the place to start.

Next in the series: Risk Structure in 13F Filings — the same decomposition across the 13F books of Buffett, Ackman, Lone Pine, Tiger Global, and Baupost: market, thematic, and stock-specific risk, and what survives the 45-day filing lag.