What will be the largest company in the world by December 2026?

1. Abstract

The world’s most valuable company is among the most legible forecasting questions in existence: a small candidate set, a public daily resolution, and a liquid prediction market that already prices it. We use that legibility to demonstrate a pre-registered, market-benchmarked forecasting method, the same freeze-before-read discipline we apply to capital, on one question: who will be the world’s largest company by market capitalization on 31 December 2026. The method is constant. State a dated, objectively resolvable claim; take the null from the market mid rather than a naive even-odds prior; freeze the claim, the null, the probability, and a kill-criterion before the outcome is known; and score the result later with the Brier score against the frozen null. At the 22 June 2026 freeze, Nvidia leads the field at roughly a five-trillion-dollar valuation, about 0.6 trillion ahead of Alphabet, and the devigged Polymarket market prices it at 0.72 to still be first in December. We forecast 0.85, modestly overweight the incumbent, justified by a declared reference class in which a leader holding a double-digit margin retains the top spot about 88 percent of the time across semiannual checkpoints since 2011. The forecast is a single observation: under the method’s own rule it is exploratory and sizes to zero. Its one honest claim is that it was recorded before the answer was known.

2. Introduction

Ask a layperson which company is the biggest in the world and they will name one without hesitation, and a year from now anyone can check whether they were right. That property, a question everyone understands whose answer is public and dated, is rare and valuable to a forecaster. Most macro questions worth asking resolve over years, in disputed units, against no clean benchmark. The identity of the world’s largest public company resolves every trading day, in dollars, against a liquid prediction market that already publishes a price. It is the ideal vehicle for showing what a forecasting discipline actually consists of, because nothing about the question itself is hard to state.

This paper runs that one question through a disciplined forecasting method we also apply to capital, operating here in forecasting-only mode with no money at risk. The method’s master rule is constant across everything it touches: a forecast is worthless as evidence unless you say what it must beat before you make it (ATOL Research, 2026). For a binary or categorical claim the thing to beat is the market mid, and the metric is the Brier score measured against that mid. We never grade a forecast against a naive even-odds prior, because a confident call that merely tracks the crowd is not skill.

The contribution is a worked, public, pre-registered application of that discipline on a legible question, with the prediction recorded here before it resolves. We do not claim an edge. A single forecast cannot establish one, and we say so in the verdict rather than burying it. What we claim is narrower and, we think, more useful: a demonstration of the full apparatus, from the dated claim through the market-implied null, the freeze, and the scoring rule, on a question the reader can check for themselves on 31 December 2026.

Section 3 places the method in the forecasting and prediction-market literature. Section 4 describes the method, the discipline, the null, the freeze, and the scoring rule end to end. Section 5 reports the standings at the freeze, the market-implied null, our forecast and its tilt against the market, and the score the forecast will earn under each resolution. Section 6 interprets the tilt and the question of whether a deviation from the market is skill or hubris. Sections 7 and 8 state the limitations and what resolution will settle.

3. Background and Related Work

The evaluation machinery is standard and old. Forecast quality for probabilistic claims is measured with proper scoring rules, which are minimized in expectation only by honest probabilities. The Brier score (Brier, 1950) is the mean squared error between the forecast probability and the realized binary outcome; the logarithmic score is its sterner sibling, punishing confident mistakes harder. Murphy’s decomposition (Murphy, 1973) splits a Brier score into reliability, resolution, and uncertainty, which is how a forecaster learns whether their 70 percents come true 70 percent of the time. Gneiting and Raftery (2007) give the general theory of proper scoring rules that underwrites the choice. None of this is novel here; the method simply enforces it in code.

The choice of null is where forecasting discipline most often fails, and where the prediction-market literature is decisive. Market prices aggregate dispersed information and are hard to beat (Wolfers and Zitzewitz, 2004); an unfunded crowd median is a weaker benchmark than a liquid book, because money disciplines the quote. The method reads its null from the venues that price the question, preferring an order-book mid to a community median, and refuses to invent an even-odds prior when no market exists. The pre-registration habit, freezing the claim and the benchmark before the outcome is known, is the operational core of the superforecasting program (Tetlock and Gardner, 2015): scoring is honest only against a number you committed to before you could see the answer.

On the specific object of the forecast, the single largest company by market capitalization, there is no tidy academic neighbor to predicting next December’s leader, so we lean on the historical record itself as the reference class (the public market-capitalization record; companiesmarketcap.com, 2026; Wikipedia, 2026). That record shows long incumbencies broken by rare regime changes, which is the empirical regularity our base rate encodes.

Within the ATOL corpus, this paper continues a line of market-benchmarked forecasts. MKT-004 (ATOL Research, 2026) built a tournament engine and graded it against the devigged betting line; MKT-006 (ATOL Research, 2026) placed an ATOL forecaster beside a Goldman Sachs model on a single shared question; MKT-005 (ATOL Research, 2026) documented the autonomous engineering practice behind these systems. This paper is the first to apply the general method, rather than a bespoke single-purpose engine, on a general macro question, and the first whose headline forecast is published unresolved.

4. Methods

4.1 The question and its resolution

The claim is a single dated, objectively resolvable sentence: Nvidia is the world’s largest publicly traded company by market capitalization on 31 December 2026. It resolves yes if Nvidia has the highest closing market capitalization of any company on the last trading session of 2026, and no otherwise, read from the public market-capitalization record and corroborated by the resolution of the corresponding Polymarket market. The candidate set is small: at the freeze, only Nvidia, Alphabet, and Apple are within plausible reach of the top spot over the horizon (Tab. 1).

4.2 The method and the discipline

The forecast was produced with our forecasting method, run in forecasting-only mode: no capital, no live trading, every step enforced in code rather than asserted. Its pipeline is scout, freeze, score, report, and its master rule is constant (Fig. 4): beat a pre-declared null, net of cost, with overfitting correction. The ruler changes with the instrument. For a binary or categorical claim such as this one, the null is the market mid and the metric is the Brier score and its paired difference against the mid; for a continuous instrument the null would be a declared baseline and the metric a deflated Sharpe ratio. This paper is entirely on the binary branch.

4.3 Where the null comes from

The method takes its null from a cross-venue prior provider that reads the market-implied probability from the venues that price the question, preferring a true order-book mid (Kalshi, Polymarket) to an unfunded community median (Metaculus, which it labels as a crowd prior rather than a money market). For this question a liquid Polymarket market exists, “Largest Company end of December 2026,” carrying about 3.6 million dollars of volume at the freeze, so the null is its devigged mid (Tab. 1). The raw yes prices across candidates sum to more than one because of the bookmaker’s overround; we remove it by normalizing each price by the sum,

where $q_i$ is the raw yes price for candidate $i$. When no liquid market prices a claim, the method does not fall back to even odds: it requires an explicitly declared reference-class base rate, and refuses the forecast if neither a market nor a named base rate is available. Here we declare such a base rate as a second, independent anchor even though the market is the binding null: across semiannual checkpoints since 2011, a leader holding a double-digit percentage margin over the second-place company retained the top spot about 88 percent of the time (Fig. 3).

4.4 The freeze

Every forecast is pre-registered. At the moment of forecasting, the method freezes a record that fixes the hypothesis and its objective resolution criterion, the null and its value at forecast time, the forecaster’s probability p_true, the metric, the declared sample count, and a kill-criterion that would falsify the claim early (Fig. 4). The freeze is content-addressed, so any drift in the null, horizon, or criterion is a new freeze rather than a quiet edit. This is what closes the worst leak in long-horizon forecasting: at scoring time the scorer reads the frozen null, never a live quote that has drifted toward the truth in the interim. Re-reading the null at scoring time would flatter every forecaster.

The freeze for this paper is dated 22 June 2026, carries a single operator id, sets the null at 0.72 and p_true at 0.85, and declares the kill-criterion that Nvidia’s lead over the second-place company falling below roughly 200 billion dollars before resolution would put the call in immediate danger (Fig. 4). The method also records how many times a claim was attempted; a bespoke forecast such as this one is a single observation, N = 1, and that count is on the record from the start because no overfitting correction is computable without it.

4.5 Scoring

When the claim resolves, the forward scorer computes, against the frozen null, the Brier score for the forecaster and for the null,

where $y_i \in {0, 1}$ is the realized outcome, and the paired difference

which is positive when the forecaster beats the null. The headline is $\Delta B$ with a confidence interval, and the interval uses an effective sample size rather than the raw count, because forecasts made in one macro regime are correlated and forty calls in a trending world are not forty independent bets,

with $\rho_1$ the lag-one autocorrelation of the paired-difference series. The verdict rule is concrete and is the same one that, in a trading context, keeps an under-evidenced edge from ever touching capital (Tab. 4): with effective $N$ below 30 the verdict is exploratory and the forecast sizes to zero regardless of how good $\Delta B$ looks, and only with effective $N$ at or above 30 and a confidence interval whose lower bound clears zero does a forecast become confirmed. A lone forecast is therefore always exploratory. Skill is a property of a track record, not of any one brave call.

4.6 Software environment and reproducibility

The analysis notebook (notebook.ipynb) consumes only the committed CSVs under data/ and regenerates every figure and the scoring scenarios with Python 3.12 and pandas, running top to bottom with no network access. The market mid, the standings, and the historical leader record were retrieved on the dates stated in data/manifest.yaml and frozen into those CSVs; the notebook does not re-fetch them. The method’s scorer is pinned to committed golden fixtures by a fail-closed cross-check that fails rather than skip-greens when its corpus is missing, and is validated by retrodiction on known-answer resolved sets; that scorer is part of our internal forecasting toolchain and is referenced here, not reproduced.

5. Results

5.1 The standings at the freeze

At the 22 June 2026 freeze, Nvidia is the world’s most valuable public company at about 5.23 trillion dollars, ahead of Alphabet at 4.63 trillion and Apple at 4.53 trillion, with Microsoft at 3.11 trillion and Amazon at 2.87 trillion completing the top five (Fig. 1, Tab. 1). Nvidia’s lead over second-place Alphabet is about 0.6 trillion dollars, roughly 13 percent; earlier in 2026 that lead briefly exceeded a trillion dollars as Nvidia touched a 5.5-trillion-dollar valuation intraday (CNBC, 2026). The shape that matters for the forecast is the gap at the top: one company clear of a tightly bunched pair, with a third tier well behind.

The longer record sets the reference class (Fig. 3). The top spot is sticky. ExxonMobil held it into the early 2010s; Apple held it for most of the years from 2012 to 2024, with only the 2018 year-end going to Microsoft and the end of 2019 going to the freshly listed Saudi Aramco; Nvidia took it during 2025 on the artificial-intelligence build-out and has held it since (companiesmarketcap.com, 2026; Wikipedia, 2026). Across the roughly thirty semiannual transitions since 2011 the identity of the leader changed only a handful of times, and almost every change occurred when the leader’s margin was thin. A leader holding a double-digit margin, as Nvidia does now, retained the top spot at the next semiannual checkpoint about 88 percent of the time, the declared base rate of Section 4.3.

Table 1. Candidates for the world’s largest company on 31 December 2026: market capitalization at the freeze, the raw and devigged Polymarket mid, the ATOL forecast, and the deviation. Market caps are in trillions of US dollars; the “Field” row aggregates Amazon, Saudi Aramco, Microsoft, and Tesla, each priced at about 1 percent. Probabilities, not percentages.

5.2 The market-implied null

The Polymarket market “Largest Company end of December 2026” prices Nvidia at a raw 0.74, Alphabet at 0.11, and Apple at 0.093, with a curious 0.046 on SpaceX and about 1 percent each on Amazon, Saudi Aramco, Microsoft, and Tesla (Tab. 1). Devigged by Equation (1), Nvidia’s mid is 0.72, which is the binding null for our claim. The SpaceX price is a market artifact rather than a live possibility: SpaceX is private, and for it to be the largest company by December it would need a valuation near Nvidia’s, which is not reachable on the horizon; we read its 0.046 as a lottery bid and fade it accordingly.

The near-term market is far more confident than the December market, exactly as it should be. An earlier reading of the same family of markets on 17 June 2026 found the corresponding “largest company by end of June” market pricing Nvidia at 0.96 to hold the crown thirteen days out (Polymarket, 2026). That the same incumbent is priced at 0.96 over two weeks and 0.72 over six months is the market correctly assigning more challenger optionality to the longer horizon, and it frames our task: the only interesting question is whether 0.72 is the right number for December, not whether Nvidia is ahead today.

5.3 The forecast and its tilt

We forecast Nvidia at 0.85 to be the world’s largest company on 31 December 2026, 0.13 above the devigged market mid, and we are underweight every other candidate: Alphabet at 0.08, Apple at 0.04, SpaceX at 0.01, and the rest of the field at 0.02 (Fig. 2, Tab. 1). The tilt is a single coherent bet: overweight the incumbent, underweight the challengers and the lottery bid. Its justification is the margin-conditioned reference class. A leader with a double-digit lead retains the top spot about 88 percent of the time across the historical checkpoints (Fig. 3), and Nvidia carries both that margin and a structural moat in the artificial-intelligence build-out that is driving the entire top of the market. We shade our 0.85 below the raw 88 percent base rate, not above it, out of respect for two things the base rate does not see: Nvidia is the highest-beta name at the top of the market, so a single disappointing print or a wobble in data-center capital spending would compress its valuation faster than Alphabet’s or Apple’s, and six months is enough time for the roughly 13 percent gap to close. The market’s 0.72 prices that challenger optionality more heavily than we do; that disagreement is the entire content of the forecast.

5.4 What the score will be

Because the forecast is more confident than the market, it will be scored more sharply in both directions (Fig. 5, Tab. 3). If Nvidia retains the crown, the forecaster’s Brier score by Equation (2) is 0.0225 against the market’s 0.0784, for a paired advantage of 0.0559 by Equation (3): we win, and by more than the market because we were more right. If Nvidia loses it, the forecaster’s Brier is 0.7225 against the market’s 0.5184, a paired deficit of 0.2041: we lose, and by more than the market because we were more wrong. Confidence above the market magnifies the score in both directions.

Whether that confidence is wise depends on whose beliefs you use, and the symmetry is exact. Under our own probability of 0.85, the expected paired advantage of the deviation is plus 0.0169 Brier points; under the market’s 0.72, the same deviation has an expected value of minus 0.0169 (Tab. 3). By our lights the deviation is worth making; by the market’s lights it is a mistake. Nothing in a single observation can adjudicate between those two views, which is the point of the verdict.

Table 3. The Brier score (lower is better) for the forecaster and the market null under each resolution, the paired difference, and the expected paired difference under each party’s beliefs. Computed from the frozen p_true of 0.85 and null of 0.72 in the notebook scoring cell.

5.5 The verdict

The forecast is a single observation. Its effective sample size is one, far below the threshold of thirty that the method requires before any confidence interval on a paired Brier advantage is even reported, so the verdict is exploratory and the forecast sizes to zero (Tab. 4). This is not a hedge or a failure; it is the correct ruling. A lone forecast, however well reasoned, carries no statistical license to claim skill, and the method is built so that nothing about being confident, or being right once, can change that ruling. The forecast’s one honest claim is procedural: it was frozen, in the open, against a number we declared before the answer was known, and it will be scored against that number in December.

Table 4. The method’s verdict rule, reproduced from its framework. The effective sample size, not the raw count, decides whether a forecast can ever be called confirmed. This forecast sits in the first row.

6. Discussion

The forecast is a bet on the incumbent against the market’s pricing of challenger optionality. The market’s 0.72 is a coherent read: six months is long enough that Alphabet’s continued rerating on artificial-intelligence integration, or an Apple resurgence, or a derating of Nvidia’s multiple, could each close a 13 percent gap, and the market spreads its remaining mass across those paths. Our 0.85 is the claim that the market is slightly underweighting how rarely a leader with a clear margin actually loses the top spot over a single semiannual window (Fig. 3). Both numbers can be sensible at once; the gap between them is small in probability and large only in the score swing it creates (Fig. 5).

The SpaceX line is worth dwelling on because it shows what the discipline is for. A naive reading of the market would carry SpaceX at about 4.5 percent into the forecast. The discipline asks what real-world path resolves that bet yes, finds that none does on this horizon, and fades it to near zero. Markets price lottery bids; a forecaster’s job is to notice which prices are information and which are noise, and to say which is which before the outcome, not after.

For the practitioner the takeaway is not the number but the legibility. This question resolves in months, not years, which means a method pointed at a standing set of such questions can accrue a scored track record quickly. The single forecast in this paper proves nothing about skill, by construction. A ledger of fifty of them, each frozen against a market mid and scored on resolution, would, and that is the only way the exploratory verdict of Table 4 ever becomes a confirmed one.

7. Limitations

1. The forecast is a single observation. Effective sample size is one, so under the method’s own rule (Tab. 4) it is exploratory and licenses no claim of skill; the paired Brier advantage of Section 5.4 is illustrative of the scoring rule, not evidence of an edge.

2. The null rests on one venue. We take the market mid from a single Polymarket market and devig it by a simple overround normalization (Eq. 1); a different venue or a different devigging method would move the null by a point or two and with it the headline deviation.

3. The reference-class base rate is a hand-built count, not a fitted model. The roughly 88 percent retention figure (Fig. 3) rests on a small number of regime changes since 2011 and on judgment calls about which year-end transitions count; it is a declared anchor, not an estimated probability with an interval.

4. Market-capitalization rankings move intraday, so the freeze is a snapshot. The standings of Figure 1 and the lead of Section 5.1 are as of 22 June 2026 and will have drifted by the time this is read; the claim and its resolution date, however, are fixed by the freeze.

5. The resolution admits edge cases. A private candidate such as SpaceX has no public market capitalization, and intraday-versus-close timing on 31 December 2026 could matter in a near-tie; the resolution criterion of Section 4.1 ties to the public close and to the corresponding Polymarket resolution to keep it objective.

6. This is the binary ruler only, and a single operator’s view. The forecast is the method’s binary branch with a market-mid null; the continuous-instrument machinery and the heavier overfitting battery are out of scope. The probability is a single operator’s and is not blended with any other view, because calibration is personal and the ledger refuses to pool operators.

7. The ATOL pre-registration flag is false by design. The forecast was frozen under the method’s freeze-before-read protocol (Section 4.4), but this paper does not commit a separate ATOL hypothesis.md with a git SHA, so the repository-level preregistered field is false; elevating the freeze to a formally pre-registered ATOL hypothesis is a future step.

8. Conclusion

We applied a disciplined forecasting method to one legible question and recorded its answer in the open before it could resolve. At the 22 June 2026 freeze the world’s largest company is Nvidia at about 5.23 trillion dollars (Fig. 1), the December prediction market prices it at 0.72 to still be first (Tab. 1), and we forecast 0.85, modestly overweight the incumbent on a margin-conditioned reading of a sticky historical record (Fig. 2, Fig. 3). The deviation wins 0.0559 Brier points if Nvidia holds and loses 0.2041 if it does not (Fig. 5, Tab. 3).

What remains open is the only thing that matters: the answer. The claim resolves on 31 December 2026, and the paired Brier advantage will be scored then against the null frozen here, not against a quote that has drifted toward the truth. The verdict today is exploratory and the forecast sizes to zero (Tab. 4), and it will stay that way for this question no matter how it resolves, because one observation is one observation. The next paper is not a better forecast of this question; it is the thirtieth forecast of its kind, the point at which a track record, rather than a single brave call, can finally be scored for skill.

9. Data Availability

All datasets are released under CC BY 4.0 and inventoried in data/manifest.yaml. The analysis is reproduced by notebook.ipynb, which loads the committed CSVs and regenerates every figure and the scoring scenarios with no network access. The datasets are: standings (the top public companies by market capitalization at the freeze); forecast-vs-market (the ATOL forecast and the devigged Polymarket mid per candidate); throne-timeline (the identity and approximate year-end market capitalization of the world’s largest company, 2011 to 2026); and scoring-scenarios (the Brier score for the forecaster and the market under each resolution).

The market mid is from the Polymarket market “Largest Company end of December 2026,” retrieved 22 June 2026; Polymarket data is the property of Polymarket and is used here at the scale of the retrieved snapshot, not redistributed in bulk. The standings are from public market-capitalization aggregators (companiesmarketcap.com and reputable financial coverage, June 2026). The historical leader record is from the public market-capitalization record (companiesmarketcap.com, 2026; Wikipedia, 2026); year-end leader identities are well documented and the year-end capitalizations are approximate.

10. References

See references.yaml. Inline citations resolve there: Brier (1950) and Murphy (1973) on scoring and its decomposition, Gneiting and Raftery (2007) on proper scoring rules, Wolfers and Zitzewitz (2004) on prediction-market efficiency, Tetlock and Gardner (2015) on pre-registration and forecasting discipline, the forecasting method (ATOL Research, 2026), the public market-capitalization record (companiesmarketcap.com, 2026; Wikipedia, 2026) and CNBC (2026) on the standings, Polymarket (2026) on the market mid, and ATOL Research (2026) for MKT-004, MKT-005, and MKT-006.