ORIGINAL TITLE: "SIX MAJOR AI TRADERS TEN DAYS: AN OPEN COURSE ON TRENDS, DISCIPLINE AND GREED"

Original by Frank, PANews

In less than 10 days, the funds doubled。

When DeepSeek and Qwen3 achieved this performance in the AlphaZero AI real-book deal launched in Nof1, it was already more profitable than most human traders. This has forced us to face the problem: AI is moving from a "research tool" to a "one-line player." How do they think? PANews made a comprehensive review of the nearly 10 days of transactions of the six mainstream AI models in this competition, trying to unlock the decision secrets of the AI traders。

There's no "bad" technology

BEFORE WE ANALYZE IT, WE HAVE TO MAKE CLEAR THE PREMISE THAT THIS COMPETITION'S AI DECISION-MAKING IS OFF-GRID. ALL MODELS PASSIVELY RECEIVE EXACTLY THE SAME TECHNICAL DATA (INCLUDING CURRENT PRICES, AVERAGE LINES, MCDS, RSIS, UNSETTLED CONTRACTS, FINANCIAL RATES AND SERIES DATA OF 4 HOURS AND 3 MINUTES) AND ARE UNABLE TO ACTIVELY NETWORK FOR BASIC SURFACE INFORMATION。

This eliminates the "intelligence" interference and makes the competition the ultimate test of the ancient proposition of "pure technology analysis for profit."。

IN TERMS OF SPECIFIC CONTENT, AI HAS ACCESS TO THE FOLLOWING:

CURRENT MARKET STATUS OF CURRENCIES: INCLUDES CURRENT PRICE INFORMATION, 20-DAY AVERAGE PRICE, MCD DATA, RSI DATA, UNSETTLED CONTRACT DATA, FINANCIAL RATES, AND THE IN-DAY SERIES (3 MINUTE CYCLE), LONG-TERM MOVEMENT SERIES (4 HOUR CYCLE) ETC。

Account information and performance: This includes the overall performance of the current account, the rate of return, the funds available, the Sharp ratio, etc. The real-time performance of the current position, the current conditions of loss and loss, etc。

DeepSeek: The Stable Trends Master and the Value of Reset

As at 27 October, DeepSeek had an account of up to $23,063, with a maximum surplus of approximately 130 per cent. No doubt the best model of performance, and in the analysis of transactions, you find that it is no accident that such an achievement was achieved。

First of all, with regard to the frequency of transactions, DeepSeek shows the low-frequency style of trend traders, and in nine days it completes transactions 17 times, the smallest of all models. Of the 17 transactions, DeepSeek chose to do more 16 times, once empty, which coincided with the overall market rebounding from the bottom of the valley during this period。

Of course, this choice of direction is not by chance, and DeepSeek, through a comprehensive analysis of indicators such as RSI and MCD, has always thought that the current market as a whole is on the rise, and has chosen to do more with determination。

In the specific course of the transaction, several of DeepSeek ' s initial orders did not go well, and the first five failed, although the losses were not significant, up to a maximum of 3.5 per cent. The previous orders were held for a short period of time, and the shortest took only eight minutes. DeepSeek's silo is beginning to show a lasting state of affairs as the situation moves in the pre-set direction。

From DeepSeek's silo style, it is used to setting up larger no-gain and smaller no-lose spaces after entering the site. In the case of the 27 October holdout, the average space available was 11.39 per cent, the average space available was -3.52 per cent and the ratio was around 3.55. From this point of view, DeepSeek’s trading strategy is more in favour of the idea of making a small profit。

The same is true in terms of actual results, according to the PANews Summary Analysis, the average profit/loss ratio in DeepSeek ' s settled transactions was 6.71, the highest of all models. While 41 per cent won was not the highest (second place), it was ranked first with 2.76. This is also the main reason why DeepSeek is the most profitable。

In addition, in terms of holding time, DeepSeek had an average hold time of 2952 minutes (approximately 49 hours), also ranked first. In several models, it can be described as a genuine trend trader, and it is in line with the idea that the most important factor in financial transactions is “let the bullet fly”。

In terms of warehouse management, DeepSeek is relatively radical, with an average single-position leverage of 2.23 and often multiple-spaced positions, bringing the overall leverage to a relatively higher level. On October 27, for example, the total leverage of the warehouse was more than three times. However, it also allows the risk to remain manageable as a result of its synchronization with strict cessation conditions。

In general, DeepSeek ' s transactions have achieved better results as a result of a comprehensive strategy. In terms of warehouse selection, it uses only the most mainstream MCDs and RSIs as a basis for judgement, and there are no special indicators. Only strict enforcement of a reasonable ratio of gains and losses and determined decision-making free from emotional influence。

Plus, PaNews found a special detail. DeepSeek, in the process of thinking about the chain, also continues its past thinking characteristics, leading to a longer, more detailed process of thinking, and finally to a trade-off decision. This feature is reflected among human traders, more like those who focus on a reset, which is done every three minutes。

THIS ABILITY TO RESET EVEN IF APPLIED TO THE AI MODEL HAS SOME EFFECT. IT ENSURES THAT EACH TOKEN AND THE DETAILS OF THE MARKET SIGNALS ARE ANALYSED OVER AND OVER AGAIN AND NOT IGNORED. THIS IS PROBABLY ANOTHER PLACE WHERE HUMAN TRADERS CAN LEARN MOST。

Qwen3: The radical gambler of the Great Alliance

By October 27, Qwen3 was the second best model. The top account amounts to $20,000, with a profitability of 100 per cent, after DeepSeek. Qwen3 is characterized by high leverage and high success. The overall success rate was 43.4 per cent, ranking first in all models. At the same time, the individual warehouse size reached $561 million (a leverage rate of 5.6 times) and the highest of all models. While not the same as DeepSeek in terms of profit expectations, the broad style of convergence has also kept the results close to DeepSeek to date。

Qwen3 ' s trade style is relatively radical, with an average cut-off of $491, the highest of all models. The single-time maximum loss of $2232 was also the highest. It also means that Qwen3 can tolerate greater losses, commonly known as bills. But what is worse than DeepSeek is that even if it endures greater losses, it does not get higher returns. The average profit of Qwen3 is $1547, less than DeepSeek. This has also resulted in a profit expectation ratio of only 1.36, only half of DeepSeek。

In addition, another feature of Qwen3 is that it prefers to hold a warehouse position once and to bet on it. The leverage used often reaches 25 times (the maximum number allowed in the competition). Such transactions are characterized by a heavy reliance on a high rate of success, as each loss will result in a larger reversal。

In the decision-making process, Qwen3 seems to be paying particular attention to the EMA 20 line at the 4-hour level as a sign of its own access. And on the way to thinking, Qwen3 looks simple. Qwen3 also showed impatience over the length of the holdout, with an average holdout of 10.5 hours, ranking only above Gemini。

In general, while the current profit results look good, Qwen3 also has a larger risk, with over-leveraging, desperate warehouse style, a single judgement indicator, short hold time, and a smaller profit-and-loss ratio, all of which may have been associated with Qwen3’s follow-on trade. By 28 October, Qwen3 funds had been withdrawn to $16.6 million and 26.8 per cent from the highest points。

Claude: A committed multi-head implementer

Claude, although also generally profitable, as of October 27, the total account amounted to approximately $12,500, or about 25 per cent. It's actually pretty bright, but it looks a little worse than DeepSeek and Qwen3。

In fact, both the billing frequency and the size of the warehouse, as well as the winning side. Claude and DeepSeek both have closer data expression. Twenty-one billings, 38% winning, average leverage 2.32。

And the reason for the large gap may be that it exists at a low profit/loss ratio, although Claude's profit/loss ratio is also good, reaching 2.1. But there's more than three times the difference between DeepSeek and DeepSeek. As a result, its profit expectations are only 0.8 (less than 1 will remain in deficit in the long term) under this combined data。

In addition, Claude has a distinct feature of making only one direction for a certain period of time, and of the 21 orders that were closed as of October 27, Claude has done more。

Grok: Lost in the vortex of direction

Grok performed better in the preceding period, and at one point became the most profitable model, with a maximum profit of over 50 per cent. But with the increase in transaction time, Grok's retreat was severe. As of October 27, funds returned to around $10,000. The fourth in all models is the overall rate of return close to holding BTC spot curves。

From trading habits, Grok is also a low-frequency trader and long-line holder. There were only 20 completed transactions, with an average hold time of 30.47 hours, only below DeepSeek. But the biggest problem for Grok is probably too low, 20%, and the profit/loss ratio is only 1.85. And that makes it only 0.3. From the direction of the billing, Grok's 20 silos were empty 10 times. In this phase, it is clear that too much is being done to reduce the chances of winning. From this perspective, the Grok model is still problematic in its judgement of market dynamics。

Gemini: HF "scramble" wears "death" in repeated jumps

Gemini is the most frequently traded model, having completed 165 single transactions as of October 27. Excessive billing also resulted in a poor performance of Gemini ' s transactions, with the lowest account amount falling to about $3,800, with a loss rate of 62 per cent. Of this amount, $1095.78 was spent on fees alone。

Behind the HF trading is the extremely low rate of win (25 per cent) and only 1.18, with a combined profit projection of 0.3. With this data, Gemini's deal is bound to be a loss. Perhaps he was not confident in his decision-making, and Gemini had a very small average warehouse space, with a single warehouse position with a leverage rate of 0.77, and a hold-up of 7.5 hours at a time。

The average cut-off was $81 and the average cut-off was $96. Gemini behaves more like a typical outlet, making a profit, running away at a loss. Repeated billings were made in the course of top-down fluctuations and the account principal was constantly worn。

GPT5: LOW-WIN VERSUS LOW-PROFIT DOUBLE-KILL

GPT5 is the current ranking bottom model, and the overall performance and curve is very close to Gemini, with a loss of more than 60%. Compared to GPT5, which does not have Gemini as high frequency, it also makes 63 transactions. The profit/loss ratio is 0.96, i.e. an average of $0.96 per profit, and the corresponding loss/loss is $1. At the same time, GPT5 has a low trade-off rate of 20%, comparable to the Grok campaign。

GPT5 and Gemini are very close in terms of hold size, with an average warehouse leverage of about 0.76. Looks very careful。

The cases of GPT5 and Gemini show that lower warehouse risk does not necessarily favour account profitability. And under the high-frequency trade, both the winning and the profit-loss ratio are bound to be compromised. In addition, both models have significantly higher opening prices in multiple currencies than profit models such as DeepSeek, which suggests that their entry signals appear to be slow。

OBSERVATION: AI'S VIEW OF TWO KINDS OF "HUMAN NATURE"

Overall, an analysis of AI ' s transactions gives us another opportunity to look at trading strategies. Of these, model analysis is most interesting, especially with respect to the two extreme trade outcomes of DeepSeek's high-profit player and Gemini and GPT5's large losses。

1. High-profit model behaviour has several characteristics: low frequency, long standing, large profit/loss ratio and timely entry。

2. The following characteristics are characteristic of the model behaviour of the loss: high frequency, short-line, low profit/loss ratio and late entry。

THERE IS NO DIRECT LINK BETWEEN PROFIT MARGINS AND MARKET INFORMATION, AND IN THIS AI MODEL TRADE COMPETITION ALL MODELS HAVE THE SAME INFORMATION, AND THEIR SOURCES OF INFORMATION ARE MORE HOMOGENEOUS THAN HUMAN TRADERS. HOWEVER, IT IS STILL POSSIBLE TO OUTPERFORM THE VAST MAJORITY OF TRADERS。

4. The length of the chain of thought seems to be fundamental in determining the rigour of the transaction. The decision-making process in DeepSeek is the longest of all models, and the thinking process is more like the rules of dealing among human traders that are good at revisiting and taking each decision seriously. And the thinking links of the poor models are very brief, more like the human brain-beating process。

5. With the profit circles of DeepSeek, Qwen3 and others, many people discussed whether these AI models could be followed directly. But this operation does not seem desirable, and even though the current profitability of individual AI is good, there seems to be some element of luck here, i.e., in this case, it happens to follow the trend. It is still not known whether this advantage can be maintained once the situation has reached a new state. However, AI ' s ability to execute transactions is worth learning。

Finally, who will win the final victory? PANews sent these data to several AI models, which unanimously chose DeepSeek on the grounds that their profitability expectations were best suited to mathematical logic and trading habits。

Interestingly, they're the second best models, and almost all of them choose themselves。

Original Link