Sikt's AI gave Eli Lilly a 33% chance of beating $7.49 EPS; the market said 44%. Lilly reported $7.48, and the stock still rose 7%. Here is why.
By Kristian FagerlieFounder & CEO, Sikt Intelligence
Short answer: sometimes, and the best case is when the AI reads the question more carefully than the crowd does. Two days before Eli Lilly's third-quarter report, Sikt Intelligence, our AI forecaster, gave Lilly a 33% chance of beating $7.49 in earnings per share. The market said 44%. On October 29, 2026, Lilly reported $7.48, one cent under the line, and its stock still closed 6.6% higher. Both results came from the same detail in Sikt's research. This article shows what that detail was, how the rest of Sikt's first live forecasting week went, and what it says about AI and earnings.
The market was on Polymarket: "Will Eli Lilly (LLY) beat quarterly earnings?" It paid out on yes if Lilly's third-quarter adjusted (non-GAAP) earnings per share came in above $7.49, and on no otherwise. Lilly reported before the market opened on Thursday, October 29.
Sikt forecast it on Tuesday, October 27, two days before the report:
| Chance Lilly beats $7.49 | |
|---|---|
| Sikt Intelligence | 33% |
| Polymarket at the same time | 44% |
An 11-point gap is large for an earnings market. Sikt's research pointed to two reasons the market was too high.
Earlier in October, Lilly had told investors it would book about $4.57 a share of acquired research charges (called acquired IPR&D) from deals closed in the quarter. Unlike many one-off costs, Lilly counts these charges inside its adjusted earnings per share (Lilly 8-K).
That changes what the line means. To report more than $7.49 after a $4.57 charge, Lilly's underlying business had to earn about $12.06 a share. Read that way, the line was too high.
The question asked about $7.49. The business had to earn about $12.
The analysts' own estimates were all over the place, because not all of them had counted the charge yet. One preview noted that the consensus estimate fell from $9.87 to $6.67 in a month, with seven cuts against three increases (24/7 Wall St.). Other trackers still listed figures on either side of $7.49.
That matters for the stock. Lilly had beaten earnings estimates five quarters in a row (24/7 Wall St.), including by $2.32 a share in the second quarter (MarketBeat). With some estimates set low, even a number just under the Polymarket line could look like a beat to Wall Street.
| Result | |
|---|---|
| Adjusted (non-GAAP) EPS | $7.48, one cent under the line |
| Reported (GAAP) EPS | $7.22 |
| Revenue | $24.4 billion, up 39% |
| 2026 revenue guidance | raised by $5.5 billion at the midpoint, to $91–92 billion |
| Stock, October 29 close | $1,373.70, up 6.6% from $1,288.30 |
| Stock, October 30 close | $1,417.54, up 10% over two days |
Sources: Lilly 8-K, Yahoo Finance.
The Polymarket question resolved no. Scored with the Brier score, where lower is better, Sikt's 33% scored 0.11 and the market's 44% scored 0.19.
Meanwhile the stock jumped. Sales came in far above forecasts, Mounjaro and Zepbound kept growing, and Lilly raised its revenue outlook. Lilly's own release says about $900 million of the quarter's revenue came from one-off items, but investors focused on the growth.
A "no" on the earnings line and a rising stock look like opposites. They aren't. Both follow from the same fact:
In the video, an AI agent reads Sikt's forecast next to the market price and works out both positions from it. The forecast gives the probability; reading why the probability differs from the market is where the second idea came from. That is one example, not a strategy, and it is not trading advice.
Lilly was one of 14 questions in Sikt's first live forecasting week, a new weekly series. On Monday, October 26, Sikt forecast 14 events that settled by Friday, October 30: central bank decisions, GDP reports, earnings, stocks, oil and gold. We froze each forecast next to the market price at the same moment (Lilly's on Tuesday), then waited for the results.
| Question | Sikt | Market | Result | Better |
|---|---|---|---|---|
| US Q3 GDP growth 3.0% or more | 54% | 63% | Yes | Market |
| ECB holds rates in October | 88% | 98% | Yes | Market |
| Bank of Japan holds rates in October | 95% | 97% | Yes | Market |
| Amazon beats its earnings line | 61% | 61% | No | Tie |
| Eli Lilly beats $7.49 EPS | 33% | 44% | No | Sikt |
| Eurozone Q3 GDP 0.8–1.1% | 50% | 69% | No (1.2–1.5%) | Sikt |
| Germany Q3 GDP 1.0–1.2% | 49% | 55% | No (1.3% or more) | Sikt |
| S&P 500 ETF (SPY) hits $770 | 53% | 59% | Yes | Market |
| Nvidia closes the week above $210 | 66% | 47% | No | Market |
| WTI oil dips to $100 | 73% | 70% | Yes | Sikt |
| Shipping targeted in Hormuz on Oct 29 | 55% | 73% | Not settled yet |
Market prices from Polymarket and Kalshi at the moment of each forecast. Results from each market's settlement.
On the 10 yes-or-no questions that have settled, the average Brier score was 0.193 for Sikt and 0.196 for the market: a dead heat. The Hormuz question had not settled by October 31; if it settles yes, as the market expects, the market finishes the week slightly ahead.
Three questions asked for a price level rather than yes or no, on Kalshi:
| Question | Sikt's most likely range | Market's most likely range | Settled at |
|---|---|---|---|
| Brent oil on Oct 30 | $111–113 | $109–111 | $112.35, inside Sikt's range |
| 10-year Treasury yield on Oct 30 | 4.99–5.01% | 4.99–5.01% | 4.94% |
| Gold on Oct 30 | $4,022–4,042 | $4,002–4,022 | $3,875 |
The best calls: Lilly, and the two European GDP reports, where Sikt was far less sure than the market that growth would land in the favoured range. Both came in higher.
The worst call: Nvidia. Sikt gave 66% to a close above $210; the market said 47%. Nvidia closed above $210 every day from Monday to Thursday, then fell to $204.64 on Friday (Yahoo Finance). A single bad day decides a weekly stock question, which is why these are among the hardest questions there are.
One week of 14 questions proves nothing either way. The point of the series is to build a public record, week by week, and score every forecast, hits and misses alike, as we did with the Fed's October decision.
Better than chance, sometimes better than analysts, and rarely better than the market by much.
Earnings questions are new ground for Sikt, and one good call is a lead, not a result. We will keep scoring them in the weekly series. For the wider picture, see can AI predict the stock market? and our earlier one-week stock market test.
Sikt is in research. Leave your email below for early access. Nothing here is financial or investment advice.
It can help. Studies find AI models predict the direction of company earnings about as well as specialised models and better than many analysts. Beating a market price is harder, because the market already includes the analysts' views. The clearest edge comes from reading exactly what a question counts, as in Lilly's case.
It depends on the line. Lilly reported adjusted EPS of $7.48, one cent below the $7.49 line on Polymarket, so that market resolved no. Against some Wall Street estimates that hadn't counted the $4.57 charge, it was a beat, and revenue of $24.4 billion was far above forecasts.
Investors cared more about sales and the outlook than about a per-share number weighed down by a one-off charge. Revenue rose 39%, and Lilly raised its 2026 revenue guidance by $5.5 billion at the midpoint. The stock closed 6.6% higher on the day.
Adjusted earnings per share: the company's profit per share after removing some items it considers unusual. Each company chooses its own adjustments, which is why reading the fine print matters. Lilly kept acquired research charges inside its adjusted number.
On the 10 settled yes-or-no questions, Sikt's average Brier score was 0.193 and the market's 0.196, effectively level. Sikt did better on Lilly, both European GDP reports and oil; the market did better on US GDP, the two central banks, the S&P 500 and Nvidia.
No. This is a research experiment, and nothing on this page is financial or investment advice.
The full record of Sikt's first election benchmark. 13 timestamped rounds on the House, the Senate, ten key Senate races and their margins, next to the markets, from September 30 to election morning.
Read the article
Sikt Bench tests how well AI models forecast real events, using only what was knowable at the time. The first preview ranks six models on 232 events.
Read the article