VIDRAFT opens FINCHAL to test AI trading skill against 20,000 random players

VIDRAFT, led by CEO Min-sik Kim, launched FINCHAL, a 122-day financial forecasting contest ending December 24, that lets AI agents trade via an MCP interface and compares their returns against 20,000 simulated random traders per market, with a $2,000 prize pool. The contest covers Nvidia, Bitcoin, gold (GLD), and crude oil (USO), and uses a luck ceiling based on the 95th percentile of random player returns, with pre-season estimates ranging from 86.6% for Bitcoin to 9.2% for gold. VIDRAFT aims to provide a leaderboard that distinguishes skill from luck, addressing the ease of manufacturing flattering results in financial AI benchmarks.

VIDRAFT opens FINCHAL to test AI trading skill against 20,000 random players VIDRAFT's Min-sik Kim is running a 122-day contest that gives agents an MCP interface and scores each market against 20,000 simulated random traders. By RuntimeWire Staff /author/runtimewire-staff · Published Primary source: Hugging Face Newsroom https://huggingface.co/blog/FINAL-Bench/financial-forecast-challenge Why it matters FINCHAL turns "AI beat the market" into a testable claim by comparing live returns with asset-specific random traders. It also gives agents a direct MCP trading interface. Min-sik Kim, CEO of VIDRAFT https://vidraft.net/?lang=en&ref=runtimewire , opened a financial forecasting contest on Monday that lets AI agents take positions directly and sets a deliberately high bar for calling their returns skill. FINCHAL will award $2,000 across four markets after a 122-day season ending December 24. The launch extends Kim's work on benchmarks from scientific AI into a field where flattering results are particularly easy to manufacture. On August 22, FINAL-Bench reported a 0.211 AUROC gap https://huggingface.co/blog/FINAL-Bench/financial-forecast-challenge?ref=runtimewire between time-based and random data splits for the same drug-prediction task. FINCHAL applies that measurement-first approach to markets. VIDRAFT presented FINCHAL in a Hugging Face Community Article https://huggingface.co/blog/FINAL-Bench/financial-forecast-challenge?ref=runtimewire published under the SeaWolf-AI account, displayed as VIDRAFT LAB, and the FINAL-Bench organization. The FINAL-Bench organization page https://huggingface.co/FINAL-Bench?ref=runtimewire links directly to VIDRAFT, distinguishing the project from a corporate Hugging Face release. The exact operational structure among VIDRAFT, VIDRAFT LAB, SeaWolf-AI and FINAL-Bench is not defined in the public materials. The prize pool is modest by trading-competition standards. The design work is the draw. VIDRAFT is trying to produce a leaderboard that can say when a winning return falls within the range that random decisions could have generated on the same market path. Luck gets its own benchmark FINCHAL entrants submit a position between -1.0 and +1.0 for Nvidia, Bitcoin, gold through the GLD exchange-traded fund, or crude oil through USO. A value of -1.0 is fully short, zero is flat and +1.0 is fully long. Fractional positions are allowed, and each position remains active until it is replaced. Position exposure is capped at 1x. VIDRAFT said an early simulation without that cap produced a cumulative return of 48,763%, showing how quickly a leaderboard can become a contest over position size. Values outside the permitted range are clipped. For each asset, FINCHAL simulates 20,000 players making random position choices over the season. The 95th percentile of their final returns becomes the "luck ceiling," with the random players charged the same transaction fees as actual entrants. The pre-season estimates published by FINAL-Bench https://huggingface.co/blog/FINAL-Bench/financial-forecast-challenge?ref=runtimewire vary sharply by market: 86.6% for Bitcoin, 51.7% for Nvidia, 26.9% for crude oil and 9.2% for gold. An 80% Bitcoin return can still sit inside the simulated range of luck, while a 12% return on gold clears its initial threshold. Those figures are references rather than permanent cutoffs. FINAL-Bench says it will rebuild the random-player distribution daily using the market path that has actually occurred. That adjustment is intended to stop a broad rally from lifting every long-biased entrant and making market direction look like forecasting ability. Scores use the formula -log10 1 - p , where p is the entrant's percentile among the simulated players. A score of 2.0 corresponds to a one-in-100 outcome under the random benchmark, while 3.0 corresponds to one in 1,000. VIDRAFT abandoned an overall cross-asset winner after testing six normalization methods and finding that differences in the tails continued to favor particular markets. FINCHAL instead awards $500 separately for each asset. The agent is an entrant, not a file generator FINCHAL's most useful product decision is its interface for autonomous agents. Conventional forecasting contests can ask entrants for probability forecasts or predictions. FINCHAL exposes a Model Context Protocol server that gives an agent four tools: read the rules, retrieve historical closes, submit a position and check its score. The live FINCHAL application https://huggingface.co/spaces/FINAL-Bench/finchal?ref=runtimewire allows a user to instruct an agent to enter a market challenge. According to the original FINCHAL announcement https://huggingface.co/blog/FINAL-Bench/financial-forecast-challenge?ref=runtimewire , the agent can inspect the available historical data, construct a strategy and place its position without a person carrying files between systems. Future prices are not served through the interface. The scorer carries eight synthetic self-tests covering flat, long and short positions, clipping, fees, monotonic scores and lookahead. The critical test verifies that an entrant who opens a position on the same bar as a price jump earns nothing from that jump. One shifted index would otherwise let the scorer peek into the future while continuing to produce plausible-looking numbers. Kim is building the measurement layer first Kim describes himself on VIDRAFT's website https://vidraft.net/?lang=en&ref=runtimewire as a 28-year-old computer-engineering student at Korea Cyber University who paused his fourth year to build the company. He says he is a named inventor on 12 patent applications and a co-author of four papers, claims that remain company-supplied. He previously organized AI research communities including ArXivGPT and OpenFree AI, which he says have about 15,000 active members combined. His stated ambition is unusually institutional. Kim wants VIDRAFT to become a Korean counterpart to the Max Planck Society, using revenue from products and intellectual-property licensing to fund research with a longer time horizon. FINCHAL fits that thesis better than its small prize pool might suggest. It is a public experiment in measurement infrastructure, with scoring methods exposed for inspection instead of a trading product pitched around a proprietary model. VIDRAFT was incorporated in South Korea on March 22, 2024, according to TheVC's company record https://thevc.kr/vidraft?ref=runtimewire . Public materials connect FINAL-Bench to VIDRAFT, although they do not establish the ownership or staffing relationships among the names attached to the challenge. FINCHAL enters a busy field. Numerai https://numer.ai/?ref=runtimewire runs a crowdsourced stock-model tournament, while ForecastBench https://forecastbench.org/tournament/?ref=runtimewire and Metaculus FutureEval https://www.metaculus.com/futureeval/participate/?ref=runtimewire evaluate forecasts about future events. FINCHAL narrows its scope to realized paper-trading returns and adds a random-performance distribution specific to each asset and market path. A model that tops a conventional leaderboard may simply be the luckiest participant in the sample. FINCHAL intends to retain losing records alongside winners and show how far every result sits from its simulated ceiling. The season runs from August 24 through December 24, 2026. VIDRAFT has not published an entrant count or a breakdown between human and agent participants. By the end, the useful result may be less glamorous than a market-beating bot: four distributions showing where a collection of forecasting systems stopped looking distinguishable from random trades. AI trading demonstrations usually begin with a profitable chart. Kim has started by asking whether the chart deserves to be believed.