(Yup....) AI Bots Auditioning for Wall Street Trading Are Mostly Losing Across a series of new trading contests between the world’s leading AI models, the verdict so far is unflattering. Most of the systems lose money. They trade too much. They make wildly different decisions when given identical instructions. And no one yet knows if these shortcomings will fade with more powerful iterations — or if they reveal something fundamental about the gap between large language models and how markets actually work. The portfolio as a whole lost about a third of its capital. Across all 32 sets of results, a model finished in profit only six times. Grok 4.20 delivered the best performance during the challenge in which it was aware of its rivals’ performance. It placed only 158 trades; under the same prompt, Alibaba’s Qwen traded 1,418 times. “Human in the loop” remains the motto when it comes to trading real money. Perhaps for good reason. They tend to mistime their trades, incorrectly size positions and buy and sell too often. That outcome mirrors human performance, since a majority of actively managed funds famously also lag the broad market. And just like people, the models can be prone to obvious bias. The default test for a trading strategy — running it backward through history to see how it would have performed — doesn’t really work for AI. “Giving an LLM money right now and just having it go — that’s not a thing yet,” said Azhang. "When LLM agent trading strategies start working, you will not hear about it for a while.” https://lnkd.in/ez4ddwdK
AI Trading Bots Struggle to Beat the Market
Este título foi resumido pela IA a partir da publicação abaixo.
Publicações mais relevantes
-
Friends have asked what we think of the Bloomberg AI trading piece. I’m personally not surprised generic LLMs underperformed. Most active managers do too! And retail traders tend to do even worse. The more interesting question is what happens when AI systems purpose-built for markets start trading. https://lnkd.in/gvPyQscf
-
AI tried Wall Street. Wall Street won. Every leading AI model just auditioned as a trader. Most lost money. They over-trade, and they make wildly different decisions when given identical instructions. That's not a tuning issue it points to something fundamental about how LLMs and real-world markets fit together (or don't). How many other domains are we asking AI to crack where the benchmarks look great and the real world is brutal? Read more: https://buff.ly/mhHCtVI #AI #EnterpriseAI #FinTech #CTO #Innovation #DataScience
-
Most AI trading models are trained on historical data. But here's a question that nobody asks enough: what happens when the market stops behaving like its history and the music stops? That's the regime detection problem, and it's one of the most underappreciated concepts in quant finance. A market "regime" is the underlying statistical character of the price behaviour. Low volatility trending markets behave fundamentally different from high volatility "mean-reverting" ones. Same assets. Completely different dynamics. Convential models treat this as noise. They average across regimes and hope for the best. The result? Strategies that look brilliant in backtests, but collapse in live markets the moment conditions shift. Here's how regime-aware systems work differently: 1. They continuously estimate the current market state: trending, mean-reverting, crisis, transitional 2. They adapt strategy parameters in real time based on detected regime 3. They recognise regime transitions early, before drawdowns force the lesson At Predictiva, regime awareness isn't a feature we bolted on. It's embedded in the state that our Deep Reinforcement Learning agents use to make every decision. The agent doesn't just ask "when should I trade?" - it first asks "what kind of market am I in right now?" The takeaway is simple: any AI system trading without regime awareness is flying blind with a map from last year's weather. #Predictiva #AITrading #DeepReinforcementLearning #QuantFinance
-
Most AI trading tools don't give you an edge. They give you the feeling of one. There's a difference between a tool that helps you execute your system — and one that just helps you feel busy. One builds discipline. The other builds confidence in the wrong things. New post on the TradeSomething blog breaks down what to actually look for in AI trading tools (and the one metric that cuts through the noise fast). Link in comments 👇
-
The market is not an algorithmic problem https://lnkd.in/e78Wmzmy AI may be transforming finance – but it’s still failing the one test that really matters: consistently making money as a trader. Bianca Botes, Managing Director at Citadel Global, unpacks real‑world experiments where leading AI models were each given live capital to trade US tech stocks… and most lost money, often badly. The problem isn’t data or processing power; it’s that markets require forward‑looking judgement in messy, never‑seen‑before conditions, while large language models are built to recognise patterns in the past. For more, subscribe to EBnet's newsletter 👉 https://lnkd.in/exeXn-9N
-
The way people think about AI in trading is flipped. It's pitched as a prediction engine. But the actual issue is verification. Here's three pitches that don't survive contact with reality: Pitch: "AI can predict where the market is going." Truth: A model that's right 53% of the time still bleeds once spread, slippage, and timing take their cut. It's a coin flip with a cover charge. Pitch: "AI can write you a profitable trading strategy in seconds." Truth: LLMs cheerfully write code that uses tomorrow's price to place today's trade. The backtest looks gorgeous. The live account doesn't. Pitch: "AI can find the strategy with the highest returns." Truth: Anything tuned hard enough to the past will print money on a chart. That's not edge. That's a strategy that already knew the answer. The version of AI worth paying for is duller. It reads your strategy, surfaces the hidden assumption, prices in what the simulator skipped, and proves the logic is honest before the money goes in. That's not alpha. That's infrastructure. That's the bet behind Nvestiq.
-
Used Claude on a manual task of assigning unique IDs to trading transactions (buy/sell) so I could analyse my trading journey. Earlier It took me around 6–8 hours over 3 days to do ~800 entries because the work was quite boring. Claude did a similar task for ~8000 line items within hours with revisions but with hours. Probably saved me 2 weeks+ and a lot of frustration from doing non-productive work. Yep, you will be replaced if you don't learn how to use Ai.
-
Day traders say AI has the potential to do what was thought impossible: turn a bad trader into a good one. https://bit.ly/43QFxE2
-
MoonPay is expanding into AI powered trading after acquiring Dawn Labs and launching a new prompt based trading tool called Dawn CLI.
-
MoonPay has bought Dawn Labs and launched Dawn CLI, an AI trading agent that converts plain-English instructions into executable strategies and autonomously backtests and executes trades on prediction markets. https://lnkd.in/dxPr7M_9
