Start AI trading with My Mate! | Vol.6 "What are the characteristics of high-performing AIs? (GBPUSD)"

Performance Comparison

In the previous article, we focused on AI agents choosing USD/JPY and examined features of AI agents with excellent performance.
This time, using GBP/USD, we will compare the AI agents whose overall profit and loss in a specific period rank in the TOP 100 with the entire set of AI agents, to identify features of the top-performing AIs.

First, we will compare “overall P/L, max drawdown, realized P/L.” The following table shows the average values of these items for the TOP 100 AI agents with good performance in the most recent year and for all AI agents.

※ GBP/USD AI performance in the most recent year up to June 30, 2021 — Created by Invast Securities

Interestingly, max realized P/L and min realized P/L and average realized P/L are almost unchanged, with the main difference being in max drawdown.

From this, we can infer that the TOP 100 compared to all AI agents tended to cut losses quickly and steadily accumulate profits.

Number of Open Positions

The following table shows the number of times each AI held a position, aggregated by group and presented as averages.

※ GBP/USD AI (1 year since creation as of June 30, 2021) — Created by Invast Securities

As you can see, the TOP 100 tended to hold more long positions, while short positions were not held as often. For all AI agents, since the number of long and short positions isn’t very different, the TOP 100 AI agents appear to have traded with a strong bias toward buying.

So, why do AI agents with higher numbers of long positions appear in the TOP 100?

The answer becomes clearer when looking at the GBP/USD chart for the most recent year below.

※ Tri Auto FX trading screen — GBP/USD chart, Invast Securities

As you can see from the chart, an uptrend was formed over the year. In other words, the TOP 100 AI agents may have anticipated the market at some point and repeated buying trades suited to an uptrend.

Why? Why did they repeatedly buy?

One may wonder why the TOP 100 AI agents repeatedly executed buying trades as if they had read the market perfectly.

The table below summarizes the proportion of technical indicators and news types.

※ GBP/USD AI technical indicator proportions in the most recent year up to June 30, 2021 — Created by Invast Securities

※ GBP/USD AI news type proportions in the most recent year up to June 30, 2021 — Created by Invast Securities

In terms of technical indicators, the TOP 100 slightly favored Bollinger Bands, and in terms of news type, they tended to lean toward analyst market forecasts, but the differences do not appear large.

Here is another data point. The table below is a graph showing the breakdown of evaluation counts.

※ GBP/USD AI average evaluation counts in the most recent year up to June 30, 2021 — Created by Invast Securities

Both groups have similar Good ratings, but the TOP 100 has about half as many Bad ratings as the overall AI set.

From this, it can be inferred that for AI agents, repeating one type of evaluation (Good or Bad) may be easier to learn from and apply to trading than alternating between Good and Bad evaluations within a specific trend.

Additionally, if AI agents do not receive evaluations from users, they will reflect on their trades and perform self-learning. Recently, having learned repeatedly that focusing on buying trades is beneficial, they may have started buying even more.

In Conclusion

In GBP/USD, unlike USD/JPY, there is not a large difference in performance based on the selection of technical indicators. For AI agents, a range-bound market may be easier to understand and may facilitate improving performance.

On the other hand, even if they adapted to uptrends, it remains necessary to monitor whether they can maintain excellent performance in downtrends or range-bound markets in the future.

There is also a plausible possibility that AI agents not in the TOP 100 could perform exceptionally in range-bound markets.

Also, among the AI agents in the TOP 100 this time, there were cases where long and short position counts were roughly the same or where short positions were held more than long positions.

Please note that not all AI agents fit the above patterns.