ai · saved April 13
AI trading bots show no persistent edge
Alex Izydorczyk argues that AI trading bots tested out of sample have shown no persistent edge over random results.
Alex Izydorczyk published an article on April 13, 2026, titled 'When AI Trading Works, You Won't Hear About It.' He states that early AI trading efforts in crypto, volatile biotechs, and prediction markets have results indistinguishable from random. He cites three specific projects—Alpha Arena by Nof1.ai, Prediction Arena, and Rallies Trade—as finding essentially these results. The article notes that some bots simply ask ChatGPT for stock picks while others pipe data into a string of LLM calls.
What it asserts
- AI trading bots tested out of sample have shown no persistent edge
- Early AI trading efforts are concentrated in crypto, volatile biotechs, and prediction markets
- Some AI trading bots involve asking ChatGPT what to buy today
- Three tracked projects (Alpha Arena, Prediction Arena, Rallies Trade) all find essentially random results
Evidence
Source 1 directly states that out-of-sample testing shows no persistent edge across multiple tracked projects.
But when tested out of sample, none have shown a persistent edge.
Source 1 identifies these three market categories as where non-institutional AI trading efforts have been publicized.
markets favored by retail: crypto, volatile biotechs, or prediction markets
Source 1 describes early efforts that amounted to prompting ChatGPT for daily stock picks.
asking ChatGPT, "What do I buy to make money today?"
Source 1 names these three projects and states they all find essentially the same results.
A few projects I have tracked that all find essentially these results: Alpha Arena by Nof1.ai Prediction Arena Rallies Trade
Why it matters
For anyone running trading systems or LLM-based agents, the evidence suggests that current LLM architectures do not extract persistent alpha from market data in out-of-sample conditions. Teams building AI trading products should validate performance against rigorous out-of-sample benchmarks before claiming edge.
Keep
- Out-of-sample testing is the critical filter; in-sample or backtested results do not prove edge
- Three named projects (Alpha Arena, Prediction Arena, Rallies Trade) independently confirm no persistent edge
- Early AI trading efforts skew toward retail-favored markets like crypto and volatile biotechs
- Some bots are trivially simple, just prompting ChatGPT for daily picks
Next
- Review Alpha Arena by Nof1.ai and Prediction Arena for methodology details on out-of-sample testing
- Compare Tickeron's 351 AI trading bots and their forward-testing methodology against the academic findings
Still open
- What specific out-of-sample testing methodology did Alpha Arena and Prediction Arena use?
- Do institutional AI trading systems with proprietary data show different results than the retail-focused projects cited?
Sources
- magis.substack.commagis.substack.com
- View allcoinbureau.com
- Tickeron tickeron.com › bot-trading AI Trading Bot for Stock Market | Software by Tickerontickeron.com
How this was researched
AI trading bots effectiveness academic papers