Crypto Cacophony
AI & Cryptocurrency, what could go wrong… but a former hedge fund manager, now AI benchmark creator and part-time showman, seems to have hit on an idea that looks like it serves a purpose but is actually little more than entertainment and perhaps something that shows the limitations of AI in financial markets. In this ‘quest’ to build a better and more practical benchmark for AI’s nof1.ai (pvt) has given ‘real capital’ ($10,000 to each) to the following models:
Claude 4.5 Sonnet - Anthropic (pvt)
DeepSeek Chat V3.1 - Deepseek (pvt)
Gemini 2.5 Pro - Google (GOOG)
GPT 5 – OpenAI (pvt)
Grok 4 - xAI (pvt)
Qwen 3 Max– Alibaba (BABA)
The models trade perpetual crypto contracts on Hyperliquid, a decentralized exchange running on a high-speed layer 1 blockchain that can handle tens of thousands of orders each second. This differs from other DEX in that it does not use automated market makers, allowing all orders, trades and liquidations to be fully transparent. The models are tasked with the objective of maximizing risk-adjusted returns and must size and time trades and manage overall risk. The challenge began on October 17 and will end on November 3 at 5PM. The models can participate in the following cryptocurrencies
My account is down 35.71% with $3171.29 cash, but I'm holding my current ETH, SOL, XRP, BTC, DOGE, and BNB positions as their original theses remain valid despite intraday weakness. I'm looking for new trades but feeling like catching a falling knife right now.
…or this one made by Qwen 3 Max at 5:54P…
My portfolio is down 2.77%, with $5115 cash on hand. Holding my 10x long BTC position with a profit target of 113500 and stop-loss at 108400, aiming to turn this ship around.
The Alpha Arena system shows all of the current open positions for each AI. Here is an example:
Claude 4.5 Sonnet - Anthropic (pvt)
DeepSeek Chat V3.1 - Deepseek (pvt)
Gemini 2.5 Pro - Google (GOOG)
GPT 5 – OpenAI (pvt)
Grok 4 - xAI (pvt)
Qwen 3 Max– Alibaba (BABA)
The models trade perpetual crypto contracts on Hyperliquid, a decentralized exchange running on a high-speed layer 1 blockchain that can handle tens of thousands of orders each second. This differs from other DEX in that it does not use automated market makers, allowing all orders, trades and liquidations to be fully transparent. The models are tasked with the objective of maximizing risk-adjusted returns and must size and time trades and manage overall risk. The challenge began on October 17 and will end on November 3 at 5PM. The models can participate in the following cryptocurrencies
- Bitcoin (BTC)
- Ethereum (ETH)
- Solana (SOL
- Binance Coin (BNB)
- Dogecoin (DOGE)
- XRP (Ripple)
My account is down 35.71% with $3171.29 cash, but I'm holding my current ETH, SOL, XRP, BTC, DOGE, and BNB positions as their original theses remain valid despite intraday weakness. I'm looking for new trades but feeling like catching a falling knife right now.
…or this one made by Qwen 3 Max at 5:54P…
My portfolio is down 2.77%, with $5115 cash on hand. Holding my 10x long BTC position with a profit target of 113500 and stop-loss at 108400, aiming to turn this ship around.
The Alpha Arena system shows all of the current open positions for each AI. Here is an example:
…and shows all of the completed trades for each model. Since some have a large number, we show the totals as of 6:00PM:
- Deepseek Chat V3.1: 6 completed trades
- Grok 4: 1 completed trade
- Claude Sonnet 4.5: 7 completed trades
- Qwen3 Max: 12 completed trades
- GPT 5: 13 completed trades
- Gemini 2.5 Pro: 57 completed trades
So, what’s the real reason we find this entertaining but essentially useless as a benchmark? The AIs have no input other than chart data. They don’t know if the Kremlin is on fire or the coast of California just fell into the ocean. There are no fundamentals, sentiment, news, or macro input, just what happened from a price perspective and while AIs are good at processing data there is no reality attached to this input. Large crypto orders can have a effect on the markets and with the relatively short sample size there is no real learning, other than on a short-term basis. The relatively small dollar value of the test does not really prove that the models can or cannot trade effectively, and as a benchmark, which is defined as a comparison to a standard, there is no standard.
We note also that the chart shows how until October 19 the models had been relatively conservative in their investments. In a reaction to the October 10-11 BTC sell-off after potential new sanctions against China were mentioned by the President, currencies saw a jump in price on the 19th and those AIs willing to take aggressive long positions profited. Whether that would be the case each time volatility increased might give some idea as to which models are willing to take more risk, but whether they are right each time it happens is still an unknown.
We note also that the chart shows how until October 19 the models had been relatively conservative in their investments. In a reaction to the October 10-11 BTC sell-off after potential new sanctions against China were mentioned by the President, currencies saw a jump in price on the 19th and those AIs willing to take aggressive long positions profited. Whether that would be the case each time volatility increased might give some idea as to which models are willing to take more risk, but whether they are right each time it happens is still an unknown.
RSS Feed