Seven Misses, One Payday: Testing a Probability Engine

Think of Sikt Intelligence as a careful friend who reads the whole morning paper, weighs the evidence in each story, and tells you what chance to give it. The Oslo startup feeds world news into AI models and assigns honest probabilities to big events. Then it places each probability next to the market price. When the two differ by a lot, a signal is born. What hooked me is that discipline: not prediction, but probability set against price.

The trial setup works like a time setup. The forecasting model's clock went back to Monday, October 5, 9:25 New York time; every story, price, and report it could read came from an archive frozen before that moment. Eight familiar questions: a broad US index, a leading AI component maker, the ounce, oil, Tesla, Netflix, Amazon, and SpaceX. On Wednesday morning the clock reset and the tickets refreshed. The rule was plain: place $1,000 only on a wide gap between the AI and the market, keep hands off small gaps.

Ten tickets, seven misses, still up

The result looks bad at first: 7 of 10 tickets lost. Yet the book closed the week about $3,000 up. Netflix was the secret. The share had to climb about 4% by Friday; the market gave it a 1-in-8 chance while Sikt, studying usual swings, said 1-in-5. With the odds seen as long, the ticket was cheap: $1,000 bought a ticket paying $8,000. Friday delivered, one ticket earned $7,000 and wiped every loss. On Kalshi-style markets the ticket price simply is the probability; a wrong price means a cheap ticket.

There was also the one that slipped away. The ounce had dropped overnight and looked set to bounce; Sikt saw the bounce as more likely than most, but the gap sat under our rule, so no ticket. That ticket would have turned $1,000 into almost $7,000. The seven misses shared one trait: each backed the short side in a week when nearly all rose. Tesla sat on its line Monday and the model called it a coin flip; yet overnight talk of a component deal lifted market spirits and the share climbed. SpaceX, a fresh and jumpy listing the model read as risky, stayed calm.

Where an AI could edge ahead

What excites me most are six possible edges. First, hidden chains: on Thursday the model joined three separate stories — a Gulf Coast hurricane warning, rockets pointed at a Saudi crude harbor, and chatter over blocking US diesel sales. Noise apart, together they pointed oil upward, and crude finished above its mark. Second, links across markets: jet fuel flowed near $3.83 against the airline's $3.15 plan; the market called the quarterly results a coin flip, the model leaned toward a miss, and the results truly missed.

Third, pace: markets run all night while people log off; news landing at 3:00 in the morning rewards whoever reads first, and an AI stays up every night. Fourth, scope: thousands of forecasting markets exist and few watch them; one setup can track them all at once. Fifth, small print: which price counts, how ties settle. Sixth, memory: most firms beat expectations, so a sound forecaster starts from history, not headlines. A GoodJudgment comparison across 25 central-bank meetings backs this: human superforecasters beat Polymarket on 76% of days.

Custom asks: Adobe and the math prize

Then came the fun part: custom asks. Cloned software built with AI by reverse methods keeps spreading; the ArtCraft pack with Premiere, Photoshop, and After Effects lookalikes draws attention. The ask: does Adobe fall 10% before November 1, 2026? Sikt said 20%. Yahoo records show Adobe lost 26% across 2026 as HSBC cut its target from $388 to $302; yet sales broke records, up 13% to $6.62B in one quarter and $6.76B in the next. TechCrunch records show Canva crossing $4B in sales. Fear and records side by side. An AstralCodexten story fits here: one founder's AI turned $35 into $2M on Kalshi in seven months — but capacity runs out, and when the wrong prices vanish, the party ends.

The second custom ask cut deeper: do OpenAI or Anthropic settle another millennium problem beyond Navier-Stokes before year-end? Sikt said 8%. QuantaMagazine sums up the backdrop: on September 8, 2026, OpenAI announced its in-house setup had produced a proof finding finite-time singularity in fluid equations; 10,000 AI agents ran 88 hours, Lean checks took 17 more, and no prize was claimed. Debate continues and independent expert approval is pending. TechCrunch records note an unreleased Anthropic setup advanced on a major math problem. After the eight-question week, that 8% read to me like honest caution.

Fair to log the creator's own warning: the sample is tiny. One long shot carries or sinks a week; ten tickets prove no setup. These short horizons are not the setup's strong side either; forecasting days before settlement is its least favorite game. Early signs say distant tails beat the market more often. Sikt's midterm benchmark now nears the market by round three; Democrats lead the US midterm voting at 55 and 89. An honest frame: a small trial, no grand claims.

My takeaway: an AI need not beat the market daily; finding the single wrongly priced spot is enough. This trial taught me method over picks: write the odds honestly, set them beside the market price, and move only on a wide gap. I now follow the Oslo team's early access and midterm benchmark. The Adobe ask for November 1 and the year-end math ask sit in my notebook; both will truly test AI forecasting.