Sikt Intelligence: How Neural Networks Forecast Events and Bet on Prediction MarketsArticle · digest for November 1
Studio still life with brass analytical scales, financial reports, and a graphite Sikt prediction card on a wooden table.
Article4 sections≈ 6 min

Sikt Intelligence: How Neural Networks Forecast Events and Bet on Prediction Markets

If you've ever looked at decentralized prediction markets like Polymarket or followed odds on Kalshi, you've probably noticed strange price dynamics. All it takes is one alarming post from a media figure or a whale pouring a couple hundred thousand dollars into an unlikely outcome for the probability chart to shoot straight up. The crowd gives in to emotion, panic, and FOMO, turning collective intelligence into a chaotic rollercoaster.

One question has always fascinated me in this space: can you pit cold, continuous machine calculation against the emotional noise of the crowd? It's one thing to argue in the comments about central bank decisions or tech giants' quarterly earnings. It's another thing entirely to turn global information analysis into a mathematical model that spots statistical skews in odds and puts real capital on the line.

That is the exact problem tackled by the Oslo-based startup Sikt Intelligence. They are building an autonomous system — an AI superforecaster. It monitors the global information landscape around the clock, cross-references scattered evidence, and calculates objective event probabilities, competing directly against human market participants. Let's break down how this mechanism works, where the algorithm finds an edge, and what barriers it runs into.

AI Superforecaster Architecture: From News Noise to Bayesian Inference

Classic prediction markets run on binary contracts. On Polymarket, each outcome trades as YES or NO shares priced between zero and one dollar. If a contract on GDP growth exceeding three percent trades at 63 cents, it means the market consensus prices the event's probability at exactly 63%. If it wins, each share pays out one dollar; the losing side goes to zero.

Most participants guess outcomes based on intuition or fragments of analysis. Sikt Intelligence operates on a fundamentally different logic — as an impartial calibration agent:

  1. Continuous source parsing. The system pulls from hundreds of diverse feeds: official central bank releases (the US Federal Reserve, ECB, Bank of Japan), government statistical reports, industry databases, customs summaries, and international news wires.
  2. Factual extraction. Language models extract verifiable claims from the raw text. Clickbait headlines, emotional columnist takes, and coordinated social media campaigns are filtered out. If a source pushes unverified rumors, the algorithm lowers its weight.
  3. Bayesian probability updating. This is the mathematical core of the system. The model starts with a prior probability based on historical data. Each verified piece of evidence shifts the posterior probability according to its reliability.
  4. Comparing internal estimates with market odds. The system queries Polymarket and Kalshi APIs, matching its calculated percentage against the exchange contract price.
  5. Finding statistical edge. If the market prices an event at 63%, while the calibrated model gives only 54% with high confidence, the algorithm flags a 9% market inefficiency and generates a signal to buy NO shares.

The approach's main strength lies in processing the long tail of micro-facts. No human analyst can simultaneously track shipping logistics in the Strait of Hormuz, cross-check customs declarations, and pore over regional economic indices. The neural pipeline does all of this in parallel.

Field Trials: 14 Events and Betting Against the Crowd Consensus

In late October 2026, the team launched its first public pilot (Week 1 Pilot Test) with real capital across a pool of 14 diverse contracts with horizons ranging from a few days to one week — from October 26 to October 31, 2026.

They picked a broad mix of tasks to test the models' adaptability across domains:

  • Macroeconomics: estimating third-quarter US GDP growth (US GDP Growth Q3), Eurozone GDP dynamics, and German GDP reports.
  • Regulatory decisions: European Central Bank (ECB) and Bank of Japan (BOJ) policy meetings.
  • Geopolitics and logistics: probability of incidents and shipping closures in the Strait of Hormuz.
  • Corporate earnings: Amazon's quarterly results and pharmaceutical leader Eli Lilly's report on next-generation drugs.
  • Financial markets: NVIDIA's stock price by week's end, gold prices on October 30, US 10-Year Treasury Yield, the VTI index, and Brent crude oil prices.

The most telling case was the US GDP Q3 contract. On Polymarket, traders were buying up YES shares: the crowd believed growth would clear the 3% bar, bidding odds up to 63%. After digesting leading business activity indicators and retail sales figures, Sikt Intelligence's model put the probability at just 54%.

That 9-percentage-point gap let the system open a position against the crowd's euphoria. Similarly, on the Eli Lilly trade, cross-referencing clinical data and supply chains uncovered underestimated earnings potential, yielding a string of winning binary bets.

Where AI Beats Humans — and Where Reality Bites

AI's advantages are clear: the algorithm has no FOMO or gambling itch, holds no sympathy for particular politicians, and calibrates odds impartially. A model can analyze hundreds of markets in parallel, spotting inefficiencies where retail traders aren't even looking. We previously discussed how critical it is to manage your digital footprint and algorithmic assessments systematically in our piece on business reputation in the age of neural networks.

Still, purely algorithmic forecasting faces hard practical limits:

  • Latency. When major statistics drop, high-frequency arbitrage bots react in milliseconds. A large language model needs seconds to parse text and generate chain-of-thought reasoning, wiping out its edge on ultra-short timeframes.
  • Zero-footprint events ("black swans"). Sudden accidents, natural disasters, or behind-closed-doors regulatory moves cannot be predicted from public sources.
  • Slippage. On many niche Polymarket contracts, trading volume barely exceeds a few thousand dollars. Trying to enter a position with even 500 dollars noticeably moves the price, erasing expected profit.

The Prediction Market Ecosystem and the Future of Forecasting Agents

The methodological foundation of superforecasting was laid by Philip Tetlock's Good Judgment Project, which measured probabilistic accuracy using the Brier score. Today, AI labs have picked up the baton: Metaculus runs AI Benchmarks for Forecasting, while academic projects like AutoCast and ForecastQA test LLMs' ability to predict future outcomes.

While Polymarket scales its protocol on Polygon with USDC settlement and Kalshi courts institutional players under CFTC oversight, analytical engines like Sikt Intelligence act as a counterweight. By smoothing out crowd distortions, they turn prediction markets into an objective reality meter. Yet even the sharpest Bayesian pipeline remains just a probability calculator in a world where the future is never fully predetermined.

Want a deep diveVotes help us choose what to break down in detail. This material has 0 votes.
Sources · 2