# Can AI predict the stock market? We tested Sikt for one week

> We sent Sikt's AI back to one week in October and bet $1,000 each time it disagreed with the market. 7 of 10 bets lost, and it still came out ahead.

Oct 12, 2026 · AI forecasting · Sikt Intelligence · https://www.siktintelligence.com/blog/ai-stock-market-test

**Short answer: not on most days, but it doesn't have to.** We sent Sikt, our AI forecaster, back to one ordinary week in October 2026 and asked it eight questions about stocks, gold and oil. Each time it disagreed sharply with the market, we put $1,000 on Sikt's side. Seven of the ten bets lost. Sikt still finished the week **$2,984 ahead**, because one cheap long shot paid eight to one. One week proves nothing, and this article shows exactly why. Here is what happened, what went wrong, and where an AI might really find an edge.

## The test: a time machine for an AI

The week was **Week 41: Monday, October 5, to Friday, October 9, 2026.** Sikt forecast eight prediction markets that all settled on Friday afternoon. Each one asked whether a price would finish the week above a line:

- **Stocks:** Nvidia, Tesla, Netflix, Amazon and SpaceX.
- **Commodities:** gold and oil.
- **The whole market:** the S&P 500.

The hard part of testing an AI on the past is making sure it can't peek at the future. So we turned back Sikt's clock. As far as Sikt knew, it was **Monday morning, 9:25 New York time**, five minutes before the stock market opened. Everything it could read, the news, the prices and the reports, came from a library saved before that moment, and anything newer was locked away. Afterwards we checked every answer to make sure nothing from the future had slipped through. Then we moved the clock to **Wednesday morning** and did it again.

That makes this a **backtest**: a test on a week that is already over, not a live record. We explain why that difference matters in [can AI predict the stock market?](/blog/can-ai-predict-the-stock-market), our guide to the research and its traps.

## The rule: $1,000 on every big disagreement

We set the betting rule before looking at any result:

- **Only bet when Sikt and the market are at least 5 points apart.** Smaller disagreements: no bet.
- **Put $1,000 on the side Sikt rates more likely,** at the market's price at that moment. If the market says 12% and Sikt says 19%, Sikt buys YES at 12 cents.
- **Each share pays $1 if it's right.** So $1,000 at 12 cents buys about 8,000 shares, while at 50 cents it buys 2,000. Cheap long shots pay the most when they come in.

Prices are the market's mid prices at the moment of the forecast, with no fees or spreads. Real trading costs would take a little off every bet; see [expected value in prediction markets](/blog/expected-value-prediction-markets).

## The scoreboard: 10 bets, 7 losses, still ahead

| | Bets | Wins | Profit or loss |
|---|---|---|---|
| Monday (four days out) | 4 | 1 | +$4,000 |
| Wednesday (two days out) | 6 | 2 | −$1,016 |
| **The week** | **10** | **3** | **+$2,984 on $10,000 staked** |

Sikt was right on only three of its ten bets and still made about 30% on the money it put down. How? Almost entirely through one trade.

### Every forecast, Monday (four days out)

| Market | Sikt | Market | Bet | Result | Return |
|---|---|---|---|---|---|
| Netflix above $70? | 19% | 12.5% | YES at 12.5¢ | closed above | **+$7,000** |
| Tesla above $370? | 49% | 64.5% | NO at 35.5¢ | closed above | −$1,000 |
| Oil at or above $91? | 40% | 50.5% | NO at 49.5¢ | settled above | −$1,000 |
| Amazon above $250? | 58% | 66% | NO at 34¢ | closed above | −$1,000 |
| Gold above $4,161.99? | 52% | 48.5% | no bet | closed above | – |
| SpaceX above $160? | 47% | 49.5% | no bet | closed above | – |
| Nvidia above $235? | 54% | 53% | no bet | closed below | – |
| S&P 500 at or above 7,725? | 51% | no price | no bet | closed above | – |

### Every forecast, Wednesday (two days out)

| Market | Sikt | Market | Bet | Result | Return |
|---|---|---|---|---|---|
| Nvidia above $235? | 62.5% | 72% | NO at 28¢ | closed below | **+$2,571** |
| S&P 500 at or above 7,725? | 78% | 71% | YES at 71¢ | closed above | **+$412** |
| SpaceX above $160? | 76% | 87.5% | NO at 12.5¢ | closed above | −$1,000 |
| Tesla above $370? | 69% | 79% | NO at 21¢ | closed above | −$1,000 |
| Netflix above $70? | 33% | 39% | NO at 61¢ | closed above | −$1,000 |
| Amazon above $250? | 71% | 76.5% | NO at 23.5¢ | closed above | −$1,000 |
| Gold above $4,161.99? | 19.95% | 15% | no bet | closed above | – |
| Oil at or above $91? | 43% | 42.7% | no bet | settled above | – |

The S&P 500's Wednesday market price is approximate, so the +$412 on that bet is too.

How did the forecasts compare with the market? On Wednesday, two days out, Sikt was slightly more accurate than the market overall. On Monday the market was more accurate. Four days out, almost every forecast, Sikt's and the market's alike, sat near 50/50: that far ahead, nobody sees much.

## How seven losses still made money: Netflix

It all comes down to Monday morning and Netflix. The stock was at **$67.11** and needed to climb about **4.3%** by Friday to finish above $70. The market saw that as a long shot, about one chance in eight (12.5%).

Sikt looked at how much Netflix normally moves, about 2% on an ordinary day, and concluded that a 4% week is unusual but far from rare: closer to **one in five (19%)**. When the market thinks something is a long shot, betting on it is cheap. Our $1,000 bought 8,000 shares at 12.5 cents, a ticket that would pay $8,000.

Friday came, and Netflix closed above $70. **One bet, $7,000 profit**, enough to cover all seven losses.

This is how a forecaster can be wrong most of the time and still come out ahead. If Sikt's 19% was right, each dollar spent at 12.5 cents was worth about $1.52 on average. Most weeks that bet loses everything; the week it wins, it pays eight to one. The same math works against you when the long shot is priced too *high*, which is the more common problem; see [the favorite-longshot bias](/blog/favorite-longshot-bias).

## The other good calls, and the one that got away

- **Nvidia, Wednesday: +$2,571.** Nvidia sat just 1.8% above its $235 line with three days to go. Futures were down before the open and bond yields were climbing. The market felt safe, at 72% that Nvidia would stay above; Sikt was less sure. Nvidia slipped below, and $1,000 turned into $3,571.
- **S&P 500, Wednesday: +$412.** Sikt was more confident than the market that the index would hold above 7,725 (78% against 71%). It did.
- **Gold, Wednesday: no bet.** Gold had dropped about 2% overnight, and the market gave a rebound only a 15% chance. Sikt said almost 20%. The gap was 4.95 points, a hair under our 5-point rule, so no bet. Gold bounced and closed above the line. At 15 cents, $1,000 would have become $6,667. Rules are rules.

## What went wrong

All seven losing bets had one thing in common: **each time, Sikt bet that a price would fall short, in a week when almost everything went up.** Seven of the eight markets finished above their lines. Three lessons stand out:

- **Fast-moving news.** On Monday, Tesla sat right on its line and Sikt called it a coin flip. But overnight, Elon Musk had talked about a chip deal, and the market priced the excitement before Sikt did. Tesla rose, and Sikt lost on Tesla twice.
- **Brand-new stocks.** SpaceX had only listed weeks earlier, and its wild first weeks of trading made it look riskier than it now was. Sikt expected a bigger slide than came. The stock held steady.
- **Leaning against a rally.** Amazon (twice), oil and Netflix on Wednesday went the same way: Sikt was more cautious than the market, and the market was right.

## Where an AI could actually have an edge

Not by being right more often, but by seeing what the market hasn't put together yet. None of these is proven. They are where we are looking next:

1. **Hidden chains.** On Thursday, Sikt connected three separate news stories: a hurricane heading for the Gulf Coast on the day oil settled, missiles aimed at a Saudi oil port, and talk of a ban on American diesel exports. On its own, each was background noise; together, they pointed to higher oil. Sikt raised its forecast, and oil settled above the line. (This was outside the Monday and Wednesday bets.)
2. **Connections between markets.** Oil doesn't just move oil. In a separate earnings test, Sikt found that Delta Air Lines had planned for jet fuel at about $3.15 a gallon, while fuel averaged closer to $3.83 during the quarter. The market called Delta's earnings a coin flip; Sikt leaned toward a miss, and Delta missed.
3. **Speed.** Markets trade around the clock, but people sleep. When news breaks at 3 a.m., the first to re-read every market has an advantage, and an AI can run all night.
4. **Breadth.** There are thousands of markets, and most are barely watched. An AI can follow every one of them at the same time.
5. **The fine print.** Which price settles the bet? What counts as a tie? What's on the calendar before Friday: a share lock-up ending, a Fed release? Small details like these tend to be priced in late.
6. **Memory.** How did thousands of similar questions turn out before? Most companies beat their earnings forecasts, for example. A good forecaster starts from history, not the headline; see [base rates in forecasting](/blog/base-rates-forecasting).

## A custom question: will Adobe's stock drop 10%?

The second half of the video tries something new. Instead of forecasting an existing market, we asked Sikt a question of our own, one with no market price behind it.

**The background.** In early October, **ArtCraft** drew wide attention: a set of seven free, open-source apps modeled on Adobe's: PhotoCraft for Photoshop, FilmCraft for Premiere Pro, LightCraft for Lightroom, and others. Its developer built them with the help of AI coding tools, and they run locally without a subscription. The apps are still in early alpha, and press coverage says they are not finished replacements for Adobe's tools yet ([Generation Amiga](https://www.generationamiga.com/2026/10/07/artcraft-could-be-the-most-ambitious-open-source-adobe-alternative-yet/), [PhotographyTalk](https://www.photographytalk.com/free-adobe-alternatives-artcraft/)). It raises a fair question for investors: what happens to software companies when AI makes their products easy to copy?

**The question.** Will Adobe's stock close at least **10% below its October 9 close**, about **$218**, on any trading day from **October 12 to October 31, 2026**? Sikt also suggested a sharper wording and wrote down how it resolves, so the answer will be clear either way.

**Sikt's forecast: 27%**, a bit better than one chance in four. What it weighed:

- **For a drop:** Adobe had been reported down about 31% over the past year while the S&P 500 gained about 17%, with two sharp one-day falls in September, of 6.7% and 4.5%. ArtCraft adds evidence that AI could pressure Adobe's business.
- **Against a drop:** Adobe's latest quarter still showed revenue up 13% to $6.76 billion, with AI-first recurring revenue up more than 150% ([Pulse 2.0](https://pulse2.com/adobe-q3-revenue-rises-13-to-record-6-76-billion-and-company-raises-fy2026-targets/)). The company also has large cash flow and buyback capacity. ArtCraft is early alpha, and reaching real parity with Adobe could take months, if it happens at all.
- **The main uncertainty:** whether apps like these will matter for Adobe's business, or its share price, within three weeks.

This forecast is live: it resolves on October 31, and we will report the answer, whichever way it goes. The video also asks a second custom question, about whether an AI lab will solve another of the Millennium Prize math problems this year.

## What this test does and doesn't show

- **One week is not proof.** Ten bets is far too few to say anything about skill, and the outcomes were tied together: seven of eight markets went the same way.
- **One long shot decided the week.** Without Netflix, the other nine bets lost $4,016 together. A single long shot can make or break a week.
- **A few days out isn't Sikt's home ground.** Sikt is built for questions that resolve weeks or months away, where research has time to matter. Four days and two days before a weekly stock price settles is close to the hardest test there is.
- **It is a backtest.** The real test is forecasts made live, week after week, and graded when the answers arrive. That is what we are doing next.

## Where Sikt fits

[Sikt Intelligence](/) puts an honest probability on real-world events and shows it next to the market's price. It reads the news and data that exist today, checks every source, and has independent AI forecasters agree on one number. Every forecast is graded when the answer arrives, misses included, and the track record decides how much to trust it. You can follow our live election forecasts in the [Sikt Midterm Bench](/midterms).

Sikt is in research. Leave your email below for early access. Nothing here is financial or investment advice.

## Key takeaways

- In a one-week backtest (October 5–9, 2026), Sikt forecast eight stock, gold and oil markets four days and two days before they settled.
- With $1,000 on every disagreement of 5 points or more, 7 of 10 bets lost, yet the week ended $2,984 ahead.
- One underpriced long shot, Netflix at 12.5%, paid eight to one and covered every loss.
- The losses all bet against a rally; fast news (Tesla) and a newly listed stock (SpaceX) were Sikt's weak spots.
- On a custom question, Sikt gives Adobe's stock a 27% chance of closing 10% below its October 9 level by October 31.
- One week proves nothing: the real test is live forecasts over many weeks.

## FAQ

### Can AI predict the stock market?

Not reliably over a few days: four days out, Sikt's forecasts and the market's sat near 50/50. But an AI doesn't have to be right most of the time to be useful. In our one-week test, Sikt was right on 3 of 10 bets and still came out ahead, because it found one long shot the market had priced too low. One week is far too little to prove an edge.

### How did Sikt make money while losing 7 of 10 bets?

Through the payout on a cheap long shot. Sikt put $1,000 on Netflix finishing the week above $70 at 12.5 cents a share. When it did, the bet paid $8,000, a $7,000 profit that covered all seven $1,000 losses.

### Was this real money?

No. It was a backtest: Sikt forecast a week that had already happened, using only information from before each forecast, and we calculated what $1,000 bets at the market's prices would have returned. Fees and spreads were not included.

### What is a custom question?

A question you write yourself, with no prediction market behind it, such as whether Adobe's stock will drop 10% by the end of October. Sikt sets out how the question resolves and gives its probability, which is then checked against what happens.

### Is this investment advice?

No. This is a research experiment, not a trading strategy, and nothing on this page is financial or investment advice.

## Sources

- Sikt Intelligence: [Can AI Predict the Stock Market?](https://www.youtube.com/watch?v=W2E-rN2sgs0) (video, October 12, 2026), and the Week 41 backtest results shown in it
- Generation Amiga: [ArtCraft could be the most ambitious open-source Adobe alternative yet](https://www.generationamiga.com/2026/10/07/artcraft-could-be-the-most-ambitious-open-source-adobe-alternative-yet/) (October 7, 2026)
- PhotographyTalk: [ArtCraft Rebuilds Photoshop, Lightroom and Premiere as Free Adobe Alternatives](https://www.photographytalk.com/free-adobe-alternatives-artcraft/)
- Pulse 2.0: [Adobe: Q3 Revenue Rises 13% To Record $6.76 Billion](https://pulse2.com/adobe-q3-revenue-rises-13-to-record-6-76-billion-and-company-raises-fy2026-targets/)