Fin-Analyst at FinMMEval 2026 Task 3: A Live Hybrid Trading Agent with LLM Specialists and Rule-Based Signals
Fin-Analyst, a hybrid trading agent featuring an eight-specialist LLM pipeline for Tesla and a rule-based signal system for Bitcoin, achieved the top ranking on the FinMMEval 2026 Task 3 leaderboard with a 13.51% return, demonstrating the superiority of event-driven LLM analysis over memoryless models and fixed thresholds while highlighting the volatility sensitivity of short-term live trading evaluations.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine you are running a tiny, high-speed trading shop in a chaotic digital marketplace. You have two customers: one is Tesla (TSLA), a stock that moves like a rollercoaster, and the other is Bitcoin (BTC), a digital coin that swings like a pendulum. Your goal? To guess if they will go up or down tomorrow and make a profit.
Enter Fin-Analyst, a new robot trader built for a live contest called FinMMEval 2026. Instead of just one brain, Fin-Analyst is a team of specialists working together, but it treats its two customers very differently.
The Two-Tiered Team
For Tesla (The LLM Squad):
Think of Tesla as a celebrity who is constantly in the news. To predict its next move, Fin-Analyst hires eight different experts, each with a specific job:
- The News Junkie: Reads today's headlines.
- The Legal Eagle: Scans official government filings (like the 8-K, 10-Q, and 10-K forms).
- The Accountant: Checks the company's financial health.
- The Wall Street Watcher: Reads what other analysts are predicting.
- The Tech Guru: Looks at charts and technical patterns.
- The Social Butterfly: Reads what people are shouting on social media.
- The Event Tracker: Watches for sudden, price-sensitive announcements.
- The Quarterly Reporter: Focuses on seasonal financial reports.
All eight of these experts talk to a Meta-Agent (the team captain). The captain listens to everyone, but if today's news is super loud and confident, the captain might ignore the others and just follow the news. This whole team is powered by a Large Language Model (LLM), which is like a super-smart robot that can read and understand text.
For Bitcoin (The Rule-Based Trio):
Bitcoin is treated differently. Instead of a team of eight, Fin-Analyst uses a simple, three-person voting booth with strict rules:
- The Trend Spotter: Checks if the price went up or down recently.
- The Fear & Greed Meter: Looks at a famous "mood index" for crypto. If people are "Extreme Greed," the robot sells; if they are "Extreme Fear," it buys.
- The Momentum Coach: Follows a label telling if the market is bullish or bearish.
If two people vote "Buy" and one says "Sell," the robot buys. If they tie, the "Fear & Greed" meter breaks the tie. This part of the system doesn't use a fancy AI brain; it just follows a checklist.
The Live Race Results
The team tested this system in a live, real-money simulation (not just a video game) from May to June 2026. Here is what happened:
- Tesla: The eight-expert team crushed it. They made a +13.51% return, beating the "Buy-and-Hold" strategy (which just sits and does nothing) by a huge margin of +28.33 points. They were so good they ranked 1st place out of all the agents in the contest. Their "Sharpe Ratio" (a score for how much risk they took for their reward) was a massive 4.10, and they won 88% of their trades.
- Bitcoin: The simple rule-based team did okay, but not amazing. They ended up with a -5.30% return, which sounds bad, but remember: the "Buy-and-Hold" strategy for Bitcoin lost about 26% during the same time! So, the robot actually saved its owner from a huge crash, finishing +17.63 points ahead of the passive strategy. However, it ranked 13th overall.
The Plot Twist:
Here is the tricky part. When the researchers first checked the scoreboard in early June, the Bitcoin robot was leading, and the Tesla team was barely breaking even. But by the end of the contest in late June, the rankings flipped completely. The Tesla team caught a late rally and won gold, while the Bitcoin robot's lead faded away. This suggests that in short, wild markets, rankings can change fast, and you can't be sure who is truly the best just by looking at a snapshot.
What Worked and What Didn't
The researchers ran a "surgery" on the Tesla team to see which expert was the most important. They turned off each specialist one by one:
- The Legal Eagle (8-K filings) was the MVP. When they removed this expert, the team's performance dropped the most. This suggests that sudden, official company announcements are the biggest drivers of Tesla's price.
- The Quarterly Reporter (10-K) and the Accountant actually hurt the team slightly. It seems that looking at annual reports every single day adds too much noise and confuses the robot.
- The Daily News was also critical. Without it, the team struggled.
For Bitcoin, the simple rules worked well when the market was crashing (the robot knew to sell early), but they failed when the market was just moving sideways. The robot kept flipping its vote back and forth on tiny, random price wiggles, losing money on "noise."
The Robot's Weaknesses
Even though the Tesla team won, the authors are honest about the flaws:
- No Memory: The robot has amnesia. Every morning, it wakes up with no memory of yesterday's mistakes. If it guessed wrong on Monday, it might guess wrong again on Tuesday, repeating the same error for days.
- Rigid Rules: The Bitcoin robot uses fixed numbers (like "sell if the price drops 0.5%"). In a crazy market, these numbers don't adapt, causing the robot to trade too much on random noise.
- Small Brain: They used a smaller, cheaper AI model (gpt-4o-mini) to save money. They suspect a bigger, smarter model could have done even better.
The Verdict
The paper suggests that using a team of AI specialists to read news and filings is a powerful way to trade stocks like Tesla, especially when you focus on breaking news and official filings. However, for Bitcoin, simple rules can protect you from big crashes, but they struggle in calm markets.
The authors conclude that this is just the beginning. To get even better, future robots will need to remember their past mistakes, use bigger brains, and perhaps have the different experts debate each other before making a decision. For now, Fin-Analyst proves that a hybrid approach—mixing smart AI for stocks and simple rules for crypto—can outperform just sitting still, even in a short, volatile race.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.