By Frank, PANews

In less than ten days, the funds doubled.

When DeepSeek and Qwen3 achieved this feat in Nof1's AlphaZero AI live trading, their profit efficiency far surpassed that of most human traders. This forces us to confront the question: AI is transitioning from a "research tool" to a "frontline trader." How do they think? PANews conducted a comprehensive review of the past 10 days of trading by six major AI models in this competition, attempting to uncover the secrets of these AI traders' decision-making.

A pure technical duel without "information gap"

Before we begin our analysis, we must clarify a premise: the AI decision-making in this competition was conducted offline. All models passively received the same technical data (including current price, moving averages, MACD, RSI, open interest, funding rates, and 4-hour and 3-minute time series) and were unable to actively connect to the internet to obtain fundamental information.

This eliminates the interference of "information asymmetry" and makes this competition the ultimate test of the ancient proposition of "whether pure technical analysis can be profitable."

In terms of specific content, AI can obtain the following aspects:

1. Current market status of the currency: including current price information, 20-day moving average price, MACD data, RSI data, open interest data, funding rate, and intraday series (3-minute cycle) and long-term trend series (4-hour cycle) of some of the aforementioned data.

2. Account information and performance: including the overall performance of the current account, return rate, available funds, Sharpe ratio, etc. The real-time performance of the current position, the current take-profit, stop-loss and expiration conditions, etc.

DeepSeek: A Calm Trend Master and the Value of “Reviewing”

As of October 27th, DeepSeek's account peaked at $23,063, with a maximum profit of approximately 130%. This made it undoubtedly the best-performing model, and an analysis of its trading behavior reveals that this achievement was no accident.

First, in terms of trading frequency, DeepSeek exhibits the low-frequency style of a trend trader. Over a nine-day period, it completed 17 trades, the fewest of all models. Of these 17 trades, DeepSeek chose to go long 16 times and short 1 time, which coincides with the overall market rebound from the bottom during this period.

Of course, this direction choice is not accidental. DeepSeek has conducted a comprehensive analysis through indicators such as RSI and MACD, and has always believed that the overall market is bullish, so it chooses to go long firmly.

During the actual trading process, DeepSeek's initial orders were not smooth, with the first five orders all failing. However, the losses were relatively small, with the highest not exceeding 3.5%. Furthermore, the holding periods for these early orders were relatively short, with the shortest closing taking only eight minutes. As the market moved in its desired direction, DeepSeek's positions began to show a sustained performance.

DeepSeek's trading style suggests that it tends to set large take-profit margins and small stop-loss margins upon entry . For example, on October 27th, the average take-profit margin was 11.39%, the average stop-loss margin was -3.52%, and the profit-loss ratio was around 3.55. This suggests that DeepSeek's trading strategy favors small losses for large gains .

This is also true in actual results. According to PANews' analysis, DeepSeek's average profit-loss ratio in settled trades reached 6.71, the highest among all models. While its 41% win rate wasn't the highest (it ranked second), it still ranked first with a profit expectation of 2.76. This is the main reason why DeepSeek achieved the highest profit.

Furthermore, DeepSeek also ranked first in terms of holding time, with an average holding time of 2,952 minutes (approximately 49 hours). This makes it a true trend trader among the various models, embracing the principle of "letting the bullet fly"—the most crucial element of profitability in financial trading.

DeepSeek is relatively aggressive in position management, with an average leverage ratio of 2.23 per position. It often holds multiple positions simultaneously, resulting in a relatively high overall leverage ratio. For example, on October 27th, the total leverage ratio exceeded 3x. However, due to its strict stop-loss requirements, risks are always under control.

Overall, DeepSeek's strong trading performance is the result of a comprehensive strategy. When selecting positions, it uses only the most mainstream MACD and RSI indicators, without any special indicators. Instead, it strictly adheres to a reasonable profit-loss ratio and resolutely holds positions regardless of emotional influence.

PANews also discovered a unique detail. DeepSeek's thinking process for link analysis follows a similar pattern. This involves a lengthy, detailed process, which is then summarized into a trading decision . This characteristic, when applied to human traders, is more similar to those who focus on replaying their positions, and this replay is performed every three minutes .

This ability to review market trends appears to be useful even when applied to AI models. It ensures that every detail of every token and market signal is analyzed over and over again, without overlooking any. This is perhaps another area where human traders can best learn from.

Qwen3: A Radical Gambler

As of October 27th, Qwen3 was the second-best performing large-scale model. With a maximum account balance of $20,000 and a 100% profit rate, Qwen3's profitability was second only to DeepSeek. Qwen3 is characterized by high leverage and a high win rate . Its overall win rate reached 43.4%, ranking first among all models. Furthermore, its single position size reached $56,100 (with a leverage of 5.6x), also the highest among all models. While its profit expectations are not as high as DeepSeek's, its aggressive trading style has ensured its results closely match DeepSeek's.

Qwen3's trading style is relatively aggressive, with an average stop-loss of $491, the highest of all models. Its maximum single loss reached $2,232, also the highest. This means Qwen3 can tolerate larger losses, commonly known as holding a position. However, it falls short of DeepSeek in that even with these larger losses, it doesn't achieve higher returns. Qwen3's average profit is $1,547, less than DeepSeek's. This also results in a final profit-to-be-earned ratio of only 1.36, half that of DeepSeek.

Another characteristic of Qwen3 is his tendency to hold a single position at a time and place heavy bets on it. He often uses leverage of up to 25x (the maximum allowed in the competition) . This type of trading relies heavily on a high win rate, as every loss results in a significant drawdown.

In his decision-making process, Qwen3 appears to pay particular attention to the 20-day EMA (4-hour moving average), using it as his entry and exit signals. His thinking process also appears simple. He also displays a lack of patience in holding positions , with an average holding time of 10.5 hours, ranking only above Gemini.

Overall, while Qwen3's current profitability appears promising, it also carries significant risks. Excessive leverage, a reckless approach to opening positions, a single indicator, short holding periods, and a small profit-loss ratio all pose potential risks for Qwen3's future trading. As of press time on October 28th, Qwen3's funds had experienced a maximum drawdown of $16,600, representing a 26.8% drawdown from its peak.

Claude: A persistent multi-headed executor

While Claude is generally profitable, with a total account balance of approximately $12,500 as of October 27th, representing a profit of approximately 25%, this figure is impressive on its own, but it pales in comparison to DeepSeek and Qwen3.

In fact, Claude and DeepSeek have similar performance in terms of order frequency, position size, and win rate. They opened 21 orders in total, with a win rate of 38% and an average leverage ratio of 2.32.

The significant disparity may be due to its lower profit-loss ratio. While Claude's profit-loss ratio also performed well, reaching 2.1, it was still more than three times lower than DeepSeek's. Therefore, given these combined data, its expected profit is only 0.8 (a ratio less than 1 indicates long-term losses).

In addition, another notable feature of Claude is that he only trades in one direction for a period of time . Among the orders completed as of October 27, all of Claude's 21 orders were long.

Grok: Lost in the vortex of direction judgment

Grok performed well in its early days, even becoming the most profitable model at one point, with peak returns exceeding 50%. However, as trading time increased, Grok experienced significant drawdowns. By October 27th, funds had returned to around $10,000, ranking fourth among all models, with an overall return close to that of holding BTC spot.

In terms of trading habits, Grok also excels at low-frequency trading and long-term holding. With only 20 completed trades, the average holding period is 30.47 hours, just below DeepSeek. However, Grok's biggest issue may be its low win rate of only 20%, coupled with a profit-loss ratio of only 1.85. This results in a profit expectation of only 0.3. Regarding trade direction, Grok's 20 positions included 10 long and 10 short positions. Clearly, excessive shorting in this market situation significantly reduces the win rate. From this perspective, the Grok model still has issues with its ability to predict market trends.

Gemini: High-frequency "retail investors" wear themselves out in repeated sideways fluctuations

Gemini was the most frequently traded model, completing 165 trades as of October 27th. This excessively frequent order opening led to poor trading performance on the platform, with the lowest account balance dropping to around $3,800, resulting in a 62% loss rate. Of this, $1,095.78 was spent in fees alone.

High-frequency trading is characterized by an extremely low win rate (25%) and a profit-loss ratio of just 1.18, resulting in a combined profit expectation of only 0.3. Given these statistics, Gemini's trading is doomed to be a loss-making operation. Perhaps due to a lack of confidence in its decision-making, Gemini's average position size is also very small, with a leverage ratio of only 0.77 per position, and each position is held for only 7.5 hours.

The average stop-loss is only $81, and the average take-profit is $96. Gemini behaves more like a typical retail investor, exiting at the slightest profit and fleeing at the slightest loss. They repeatedly place orders amidst market fluctuations, constantly eroding their account capital.

GPT5: A "double kill" of low win rate and low profit-loss ratio

GPT5 is currently the lowest-ranked model, with overall performance and curves very similar to Gemini, with both models showing losses exceeding 60%. While not as frequent as Gemini, GPT5 still made 63 trades. Its profit-loss ratio was only 0.96, meaning an average profit of $0.96 per trade, with a corresponding stop-loss of $1. Furthermore, GPT5's trade win rate was a mere 20%, comparable to Grok.

In terms of position size, GPT5 and Gemini are very similar, with an average position leverage ratio of approximately 0.76, which seems very cautious.

The GPT5 and Gemini cases demonstrate that lower position risk doesn't necessarily lead to profitable accounts. Furthermore, with high-frequency trading, both win rates and profit-loss ratios are inherently unreliable. Furthermore, the opening prices for long positions in the same currency pairs in these two models are significantly higher than those of profitable models like DeepSeek, suggesting that their entry signals appear somewhat delayed.

Observation Summary: Two Types of Trading “Human Nature” Revealed by AI

Overall, analyzing AI trading behavior provides another opportunity to re-examine trading strategies. The model analysis of two extreme trading results, DeepSeek's highly profitable performance and Gemini and GPT5's significant losses, is particularly insightful.

1. Highly profitable model behaviors have the following characteristics: low frequency, long holding, large profit-loss ratio, and timely entry.

2. The loss-making model behavior has the following characteristics: high frequency, short-term, low profit-loss ratio, and late entry timing.

3. There's no direct correlation between profit and market information. In this AI trading competition, all models received the same information, making their information sources more limited than those of human traders. Yet, they were still able to achieve profits far exceeding those of most traders.

4. The length of the thought process seems to be the key factor in determining trading rigor. DeepSeek's decision-making process is the longest of all models. This process is more similar to that of human traders who are adept at reviewing and taking each decision seriously. The thought processes of poorly performing models, on the other hand, are very brief, more like the process of human decision-making based on intuition.

5. With the profitability of models like DeepSeek and Qwen3, many are discussing whether it's appropriate to directly copy these AI models. However, this approach seems undesirable. Even if some AI models currently demonstrate strong profitability, there seems to be a certain element of luck involved, namely, coincidentally following the broader market trend during a specific market period. Whether this advantage can be maintained once the market enters a new phase remains uncertain. Nevertheless, AI's trading execution capabilities are still worth studying.

In the end, who will win? PANews sent these data performances to multiple AI models, and they all chose DeepSeek because its profit expectations were most consistent with mathematical logic and its trading habits were the best.

Interestingly, almost all of their second-favored models chose themselves.