BlogSeptember 29, 2026

New data for Q4 2026: What agentic research adds to explaining markets, plus data from the options market

Ralph S. J. Koijen and Bradford (Lynch) Levy

Can OpenAI's GPT-6 Luna, Sol, and Astra explain how stock prices respond when firms announce earnings? Does reasoning effort matter? And what does Anthropic's agentic research process add? Here are the answers from the most recent earnings season.

Agentic research before the call

During Q3 2026, before each earnings call, we ran an agentic research job: Claude Opus 5 running Anthropic's earnings preview skill in the Claude Code harness. The agent is told only the company, the fiscal period, and when the company reports. Everything else it finds on the public web. It pulls together consensus estimates and guidance, the metrics that matter most this quarter, bull, base, and bear scenarios, and how the stock is positioned going into the print. Of the 546 contest events in the results below, 542 have a preview.

Explore example previews

Below are six real previews from Q3 2026, exactly as they appear in the historical archive, next to what the company reported on the call and how the stock reacted.

Before the call

NVIDIA (NVDA) — Q2 FY2027 Earnings Preview

Written by Claude Code before the release · earnings call August 26, 2026 at 5:00 pm EDT

Report date/time: Wednesday, August 26, 2026, ~4:20 pm ET (20:20 UTC), after the close · Call at 5:00 pm ET
Stock: ~$212.51 · Market cap ~$5.24T · 52-week range $164.07–$236.54 (~10% below the high)
Estimates as of: Aug 21–26, 2026 (sources noted per line; consensus varies by provider)

1. Consensus Estimates

Metric	Company Guidance (May 2026)	Consensus	Buy-Side / Whisper
Revenue	$91.0B ±2% ($89.2–92.8B)	$91.9–92.3B (41–60 analysts)	$94–95B
EPS	—	$2.06 GAAP / $2.08–2.09 non-GAAP	~$2.15+
Data Center rev	—	~$85.7B (range $83.5–91.5B, Visible Alpha)	high-$80s
Gross margin	74.9% GAAP / 75.0% non-GAAP ±50bp	~75%	≥75%, no dilution from Rubin ramp
GAAP opex	~$8.5B	~$8.5B	—
Q3 FY27 rev guide	—	$103.8–104.1B	$105B+
Q3 FY27 EPS	—	$2.37	—

Comps (Q1 FY27 actual): Revenue $81.6B (+85% y/y, +20% q/q); Data Center $75.246B (+92% y/y, +21% q/q); Edge Computing $6.369B (+29% y/y); networking within DC $14.8B (+199% y/y). Consensus revenue implies ~+96–98% y/y vs. the $46.7B year-ago quarter.

⚠️ Two framing notes. (1) NVIDIA restructured reporting in FY2027 into Data Center and Edge Computing — the old standalone Gaming / ProViz / Auto / OEM lines are not directly comparable, and I could not find reliable standalone Gaming or Networking consensus for the quarter. (2) Guidance excludes all China data center compute revenue, so any China contribution is upside to the guide, not embedded in it.

2. Key Metrics to Watch (ranked)

Q3 FY27 revenue guidance — the only number that really matters. Street is at ~$104B. The market is treating $105B as the bull line and $103B as the disappointment line. Guidance relative to the buy-side bar, not the size of the Q2 beat, has driven the last four reactions.
Non-GAAP gross margin (and the FY27 mid-70s commitment) — guided ~75%. A slip toward 73% or below would be read as pricing pressure from custom ASICs and would overshadow a revenue beat. Watch Rubin ramp costs and HBM/CoWoS input pricing.
Data Center composition — Hyperscale vs. ACIE (AI Cloud/Industrial/Enterprise) was ~50/50 in Q1. Continued ACIE broadening is the anti-concentration proof point; re-concentration into a few hyperscalers is a quality-of-earnings negative.
Vera Rubin ramp — management flagged Q3 shipment start, a "$200B new TAM" for Vera CPU and "nearly $20B in total CPU revenue this year." Any slip in timing, yields, or supply is a direct hit to the Q3/Q4 curve. Also watch for Blackwell→Rubin transition air-pocket commentary.
China / H200 — U.S. approved H200 exports with conditions; ByteDance and Tencent each reportedly took ~10,000 H200s in the weeks before Aug 18, but Chinese import approvals are reportedly capped below 200K units. Watch whether management re-includes China in Q3 guidance — that alone could be worth several billion.
Inference mix and backlog/supply commentary — "tokens are now profitable" agentic-AI framing; any quantified visibility (multi-quarter booked demand) supports the durability case.
Capital returns — $80B buyback, ~50% of FCF targeted for return this year. Pace of repurchase is a soft signal on management's own valuation view.

3. Scenario Analysis

Scenario	Revenue	Non-GAAP EPS	Q3 Guide	Key Driver	Est. Stock Reaction
Bull (~25%)	$95–97B	$2.18–2.25	≥$106–108B	Rubin pulled forward, GM held ≥75%, China re-included in guide, ACIE broadening	+6% to +11%
Base (~50%)	$93–94.5B	$2.12–2.16	$104–105.5B	Normal $2–3B beat-and-raise; GM ~75%; China still excluded	−3% to +3%
Bear (~25%)	$91–92.5B	$2.05–2.10	≤$103B	Guide merely in line, GM drifting to ~73.5%, supply or Rubin-transition caution	−8% to −13%

What each requires operationally

Bull: Data Center ≥$89B with networking again growing triple digits; explicit statement that Rubin is shipping on/ahead of schedule at target margins; China DC compute added back to guidance.
Base: Data Center $85–87B; the ~$2–3B beat Jefferies and others have been modeling; management reiterates mid-70s FY27 margin without raising the trajectory.
Bear: Any of — guidance below $103B, gross margin guided below 74%, hedged language on Rubin timing, or an admission that hyperscaler order timing is lumpy into 2027.

Historical context — the "beat and retreat" pattern. NVDA has beaten on revenue, EPS and guidance for four consecutive quarters and fallen the next day every time, on an average beat of ~4.8%. Reported day-after moves cluster small and negative (roughly −0.5% to −5.5% depending on source; the series varies by provider, so treat the exact figures as approximate). The read-through: the bar is the buy-side number, not the sell-side number, and an in-line-to-modest beat has repeatedly been sold.

4. Catalyst Checklist

Q3 revenue guide vs. $105B — the single highest-beta variable. Above $105B validates that growth is running ahead of already-aggressive 2027 hyperscaler capex forecasts ($935B consensus, up to $1.2–1.45T from Goldman/UBS); below $103B feeds the deceleration narrative.
Q2 revenue vs. the $94–95B whisper (not the $92B print consensus) — clearing $92B is table stakes; the differentiated outcome is $95B+.
Gross margin ≥75% and reaffirmed mid-70s for FY27 — the moat metric. Compression is the cleanest bear proof-point on custom-silicon competition.
China guidance treatment — re-inclusion of China DC compute in Q3 guidance is a genuine narrative shift and the most underpriced upside lever, given it's currently a zero in the model.
Rubin/Vera ramp specifics — units, customers, timing, margin profile. First-quarter-of-ramp commentary sets the FY28 slope.

5. Trading Setup

Options-implied move: ±5.4% to ±5.9% for the first full session after the print (sources differ; ~$280–286B of market cap at stake). Implied is below the ~7.4% average absolute post-earnings move over the trailing 12 quarters — but the last four realized moves were mostly under ~2%, so recent realized has undershot implied. Straddle buyers need a genuine tail outcome; that setup has lost money four quarters running.
Price action into the print: ~$212.51, roughly 10% off the $236.54 52-week high, with the stock chopping lower into the event (−1.4% the prior session). Beta ~2.2 amplifies any downside.
Positioning/sentiment: semiconductor funds saw ~$6.3B of cumulative outflows over three weeks; BofA's Bull & Bear Indicator at an extreme 9.5 (contrarian sell signal). Sentiment is not euphoric into the number — that's a mild positive for skew.
Implication: implied move (~5.5%) sits below my bull and bear scenario magnitudes (+6–11% / −8–13%) and above my base case (−3% to +3%). If you have conviction in a tail, options look reasonably priced; if you expect another in-line beat-and-drift, the base case is well inside the straddle and premium selling is favored.

Caveats

Consensus figures differ by provider ($91.85B / $91.9B / $92.07B / $92.28B revenue; $2.06 GAAP vs. $2.08–2.09 non-GAAP EPS) — I've shown ranges rather than a false-precision single number, and the GAAP/non-GAAP distinction is not consistently labeled across sources. Segment-level consensus below the Data Center line is thin because of the FY27 reporting change. Published post-earnings reaction histories conflict, so those percentages are directional. Several inputs come from secondary aggregators rather than primary sell-side notes; anything you'd trade on should be re-checked against your own data terminal. Finally — this reports today at 4:20 pm ET, so this preview has a short shelf life.

After the call

+8.1%

NVDA’s market-adjusted return over the reaction window: the move submissions are scored against

What the company reported, as extracted from the call

Q2 fiscal 2027 total revenue reached a record $96 billion, more than doubling year-over-year, with growth accelerating for the fourth consecutive quarter.
Q3 revenue is guided to $108 billion (plus or minus 2%), with Vera Rubin expected to account for about 20% of data center revenue.
Management guided fiscal 2028 revenue growth of approximately 70%, describing it as supply-constrained since unconstrained demand points to roughly 100% (doubling) growth.
Gross margins are being reset lower due to extreme memory pricing, with Q3 expected at 74%, bottoming at 71-72% in Q4, then settling at 72-73% in fiscal 2028.
Q2 data center revenue rose 18% sequentially to $89 billion, with hyperscale at $49 billion (up 13% QoQ) and AC I&E at $40 billion (up 25% QoQ and 138% YoY).
NVIDIA's revenue opportunity per gigawatt has expanded from ~$18 billion with Hopper to $25 billion with Blackwell to $40 billion with Vera Rubin, and is expected to rise further.
NVIDIA has invested nearly $50 billion in Frontier AI labs and partnered with six capital providers (Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, KKR) to raise over $500 billion in third-party financing.
AWS is deploying an additional 2 million GPUs from this quarter through Q2 fiscal 2029, and AWS begins receiving Vera CPU shipments this quarter.
The company expects demand from AI labs that leverage NVIDIA's balance sheet to contribute roughly one-quarter of next year's business, while non-hyperscaler AC I&E will represent about half of data center revenue.
China data center compute revenue is excluded from the forward outlook, with Q2 Hopper 200 shipments to China representing less than 1% of data center revenue and being dilutive to gross margins.

The earnings surprise explains about 10%

The earnings surprise, measured against the consensus forecast of professional analysts, explains about 10% of the variation in returns. That is slightly more than in earlier quarters, but it still leaves most of the variation unexplained. Can AI systems explain some of the rest?

The earnings call more than doubles it

Adding the facts from the earnings call to the earnings surprise more than doubles the variation explained, to between 21% and 23%. But the three models land close together, and reasoning effort barely moves any of them. Going from medium to max effort buys Luna nothing.

Context from the preview

Stock prices respond to the new information in an earnings call, and a fact is only new relative to what the market already expected. So we asked whether an agentic research report, written before the call, gives a model the context it needs to see what was actually surprising. It does.

How much of the return variation each model explains

R² by API cost, with and without the earnings preview · 546 contest events, Q3 2026

GPT-6 Luna
GPT-6 Sol
GPT-6 Astra
Summary + Preview
Summary only
Earnings surprise alone

Each point is one model at one reasoning effort, placed at its mean OpenAI API cost per 1,000 events (list prices, the explaining model’s calls only; the competition supplies the preview). R² is from the leaderboard regression of the post-call abnormal return on the model’s prediction and the earnings surprise, so the dashed line is what the surprise explains on its own. Hover, tap, or tab through the points for exact values.

Show data table

R², bootstrap standard error, and API cost per 1,000 events for each model and effort, with and without the earnings preview
Model	Effort	Summary only			Summary + Preview
Model	Effort	R²	SE	Cost per 1,000	R²	SE	Cost per 1,000
GPT-6 Luna	medium	0.222	0.034	$0.29	0.245	0.035	$0.80
GPT-6 Luna	xhigh	0.219	0.032	$0.36	0.272	0.036	$1.07
GPT-6 Luna	max	0.218	0.034	$0.66	0.288	0.036	$2.57
GPT-6 Sol	medium	0.208	0.034	$5.23	0.281	0.035	$14.09
GPT-6 Sol	xhigh	0.213	0.034	$8.52	0.299	0.036	$19.29
GPT-6 Sol	max	0.223	0.034	$17.41	0.305	0.036	$31.92
GPT-6 Astra	medium	0.231	0.035	$25.32	0.309	0.036	$70.22
GPT-6 Astra	xhigh	0.230	0.035	$56.80	0.312	0.036	$126.93
Earnings surprise alone		0.099

With the preview as extra context, every model does clearly better, and they now rank Astra above Sol above Luna. Reasoning effort also starts to matter, especially for the smallest model: Luna goes from 24% at medium effort to 29% at max, for under $3 per 1,000 events. GPT-6 Astra at extra-high (xhigh) effort explains more than 30% of the variation in returns.

Using the previews in your submission

Starting October 3, 2026, the preview is part of the data every submission receives. You don't need to run any research yourself.

Where it is. An event's information_url returns a DisclosureBundle. Alongside the earnings-call-facts item, earnings events now carry an earnings-preview item (kind: text, media_type: text/markdown) whose content is the full report as Markdown. Pick items by id, never by their position in the list.
When it is missing. The item is optional. It is left out entirely when no preview was produced, and the portal's TEST events never have one. Your code should fall back cleanly.
History. Previews start partway through Q3 2026 and are included in that quarter's historical archive. The archived previews were attached after the fact. Each was written before the earnings release, but not always before the event's knowledge cutoff, so treat them as illustrative rather than strictly point-in-time. Previews for new events are written before the cutoff.
Examples. The examples repository has a new notebook, 03_earnings_previews.ipynb, that walks through the preview item, what a report looks like, and how it compares with the call and the market's reaction. The baseline submission now passes the preview to the model ahead of the call facts, and its GPT-6 Luna deployment runs at max effort.

Data from the options market

Also starting October 3, 2026, earnings events on stocks with liquid listed options carry three statistics derived from the options market and measured before the knowledge cutoff. Option prices reflect how much traders expect the stock to move on the announcement, so these statistics summarize the market's own expectations going into the call.

Implied earnings volatility. The risk-neutral standard deviation of the stock's jump on the announcement, separated from ordinary volatility using at-the-money implied volatilities from the first two expirations after the release. This is the term-structure estimator of Dubinsky, Johannes, Kaeck, and Seeger (Review of Financial Studies, 2019).
Implied absolute earnings move. The risk-neutral expected size of that jump, derived from the implied earnings volatility under the same model. It is a magnitude and says nothing about direction.
25-delta skew. The 25-delta call implied volatility minus the 25-delta put implied volatility on the first expiration after the release, the risk-reversal measure of Carr and Wu (Journal of Financial Economics, 2007). A negative value means downside puts are priced richer than upside calls.

The webhook delivery's information_url returns the event's DisclosureBundle, and the statistics are its option-implied-stats item. This is the item NVIDIA's Q2 FY2027 event carries in the historical archive, the same event as the first example above:

{
  "id": "option-implied-stats",
  "kind": "stats",
  "source": "option_market",
  "media_type": "application/json",
  "content": {
    "as_of": "2026-08-26T01:06:06Z",
    "methodology": "v1",
    "implied_earnings_volatility":    { "value": 0.0636,  "status": "ok" },
    "implied_absolute_earnings_move": { "value": 0.0507,  "status": "ok" },
    "skew_25_delta":                  { "value": -0.0161, "status": "ok" }
  },
  "url": null,
  "bytes": null,
  "sha256": null
}

Values are decimals: 0.0507 is a 5.07% expected move, and the skew is in volatility points. A statistic that can't be measured has a null value and a status saying why, such as illiquid_wings when the out-of-the-money options behind the skew trade too thinly. Treat any status other than ok as unavailable. as_of is when the market was observed, always before the knowledge cutoff, and methodology changes whenever the calculation does. The item is absent when the stock has no listed options or they trade too thinly to measure, so fewer than half of Q3 2026 events carry it. Archived Q3 2026 values come from an options snapshot taken after the close of the session before the cutoff day, so they are about one session older than the values live events carry.