キャラクター

Continuous Learning & Nurturing AI -Beginner’s Guide-

Hello, I am Sameo.

MAiMATE’s AI performs “continuous learning” once a week in order to adapt to the recent market conditions. In the process of continuous learning, there is a function called “EDUCATE” that allows the creator to reflect his thoughts by evaluating the trades.

In this article, I would like to introduce these “continuous learning” and “EDUCATE” in detail.

Adapt to recent market conditions through continuous learning.

Continuous learning: A function that re-learns all AI agents that exist as of 10:00 a.m. every Saturday morning to incorporate the successes and reflections of the most recent trade results.

By running relearning every weekend, the AI agent updates itself to match the most recent market.

Continuous learning is a key feature of MAiMATE’s AI.

Immediately after birth, AI agents are not continuously learning.

In fact, the performance of the AI agent immediately after its birth is the result of continuous learning that has not yet been applied. Therefore, in order for the new AI agent to continue to perform well over the entire period of the past three years, it is necessary that the current trading method, which can easily reflect the most recent learning, is a universal one that can be used over the entire period of the past.

Learning to create AI agents

Eggs in the learning process

1.The AI agent is created according to the user’s settings and learns to trade on the past three years of data.
2.In this process, the AI agent learns sequentially from the past to the present, and just like human memory, the content learned in the past gradually fades away, while the content learned most recently is easily remembered.
3.Take an AI agent that has completed its learning, take it back three years, let it execute a trade without updating its trading method through continuous learning, and store the results.

MAiMATE’s AI learns for three years and then goes back to three years ago to trade as if it were the first time it encountered the market environment. It does not adjust parameters or perform fitting according to past market conditions as is often done when creating general automated trading programs.

Poor past performance of AI after birth is inevitable! Why continuous learning is necessary

Is there really such a thing as a universal, magical trading method that can continue to generate profits over any period of time and in any market environment…?
Let’s take a look at the performance of the top 9 AI agents (top 9 in terms of realized profit/loss in the last 1 year) for the entire period as of 12:00 on September 9, 2019.

As the above graph shows, it is difficult for an AI agent to keep winning with the same algorithm over the entire period, even if its realized profit/loss over the past year is excellent. This suggests that it is quite difficult to find a universal trading method that works for the entire period, and that continuous learning is necessary to keep revising the trading method as needed.

Is the AI with poor recent performance being dragged down by past successes?

Some users may think, “My AI agent is not as good as the last one! Isn’t it possible that the most recent learning is reflected more strongly? Isn’t the most recent learning reflected in the results? This can be said to be a situation where the AI agent has had a successful experience of finding a trading method that is very applicable to a certain period of time in the past, and the most recent learning content does not diminish that success experience. It’s very human, isn’t it?

Effects of continuous learning and EDUCATE

Next, we will check the effects of “continuous learning” and “EDUCATE”.
The trading results of the AI agent to be tested for the period from Friday, August 23, 2019 to Friday, August 30, 2019 are as follows

You can see that I incurred a larger loss on Friday, August 30. While I think the decision to cut my losses was unavoidable, I scolded myself for the settlement on August 30, based on the thought, “You could have made the decision a little earlier, you could have cut your losses as of August 28 or 29.

Here is an experiment to verify the effectiveness of continuous learning and education.

Experiment #1: Effects of continuous learning, reduced loss

The object of our research is to see how the trading results change depending on “continuous learning” and “bad news on August 30”. First, we will check the effect of the AI agent’s spontaneous continuous learning while ignoring the “good” and “bad” responses. We brought the AI agent back on August 23 after continuous learning with the data from the above period, and had it trade again. The reason for the slight difference in the valuation profit and loss figures is that it is difficult to simulate the exact same time and date, and the difference is not something that would overturn the results of the analysis.

Changes after continuous learning

As shown above, even if there is no particular “good” or “bad,” the AI agent spontaneously reflects on the magnitude of the loss and reduces the loss by accelerating the timing of the stop loss.

Experiment #2: Effects of education, no change at one time!

Next, let’s see how the AI agent changes when it is given a “no-no” on Friday, August 30, to encourage strong reflection.

After the first round of education

There is no change. In other words, the AI agent believes that the best decision is to cut its losses as of August 29th, although it has been prompted to reflect strongly.

Experiment #3: Persistent education, change after 7 times!

Still, “I want you to cut your losses on August 28! So I repeated the continuous learning for this period including “No” many times. As a result, the trade result after the seventh continuous study made me sell SRS a day earlier.

Changes after the seventh evaluation

Consistent education is the key to change in AI.

Education has a gradual but steady impact on AI.

How strongly to praise the master when he gives a “like” and how strongly to scold him when he gives a “no” are determined based on careful analysis. We are very conscious of the fact that a single “like” or “dislike” will not easily destroy the learning results that the AI agent has accumulated over a long period of time, so the above results are as expected. However, please understand that your “likes” and “dislikes” will definitely affect the AI agent’s behavior little by little, and it will become its personality.

You don’t have to force yourself to appreciate it.

If you don’t have a clear policy on which trades to praise or scold, one option is to leave the AI agent to learn on its own without forcing it to learn. It is better to leave it alone than to praise or scold it inconsistently.

People with education policies should educate AI.

If you already have a clear idea of your favorite trades, please cultivate them. For example, praise trades with a profit of 100 pips or more, and scold trades with a loss of 60 pips or more. As shown in the previous experiment, AI will always change its behavior. We have designed the system so that a single “good” or “bad” comment will not destroy the AI agent, so please do not be afraid to review and evaluate the AI agent’s trades at least once a week. Please do not be afraid to review and evaluate the AI agent’s trades once a week.

Training is simple.
“You shouldn’t do this kind of deal” is a scolding.
“That’s a great deal!” is praise.
That’s it.

Finally

Most general automated trading services have a fixed logic, and therefore, there are quite a few that can adapt to various environments over a long period of time.

MAiMATE is trying to overcome this problem of automated trading services with its “continuous learning” feature. EDUCATE” will make the service smarter and more unique by incorporating the experience of the creator. This is a feature that can only be realized by MAiMATE, which utilizes cutting-edge technology. We publish monthly reports on the overall AI performance, and as of April 2020, we can see that the continuous learning is working very well so far.

I think a platform with about 5,000 different reinforcement-learning trading AIs doing continuous learning every week is unique. It will be very interesting for us to see how it turns out in the future.

Thank you for reading!

RECOMMENDED

MAiMATE Administrator

We run the official blog of MAiMATE. It provides general information about MAiMATE, including news and announcements.

RECOMMENDED

Mai mate

Trading is fun! But sometimes it's lonely. Should I set stop loss now, or wait?
If there is a partner you could trust to relieve your anxiety... Your own AI partner to fight alongside you.
That's MAiMATE.

PAGE TOP マイメイトをひらく