『How AI Thinks』
image extraction: stable diffusion
Although I titled this Three Months After Starting AI Trading, I spent more than the first month dealing with a problem where orders weren’t going through because of code errors. So, in reality, this amounts to about a month and a half of actually operating AI trading.
As for the current situation, I’m still in the very early phase of turning potentially viable strategies into code, collecting data, and updating the code as I go.
One small step forward during this process is that one of the strategies I had been using for actual trades and collecting data on for about a month managed to finish with a net profit.
Incidentally, I had thought that automatically trading using a machine was called system trading, but apparently, as in my case, automatically trading on a DEX, on-chain, in an execution environment such as Node.js is called bot trading. And the AI that actually conducts the trades (or rather, the program) is apparently called a DeFi Bot (Decentralized Finance Bot). …The terminology itself doesn’t really matter, but even so, I can’t deny that the world seems to be getting more and more complicated.
A Promising Day-Trading Strategy
I ran the program and tested it for three weeks in the range-bound market conditions of July, and for about ten days in the trending market conditions following the crypto surge on August 19 (although it’s probably too early to call it a trend; in any case, it was a high-volatility environment).
I had wanted the algo (or rather, the bot) to handle scalping or ultra-short-term day trading, so I took a scalping-style strategy that I wasn’t particularly familiar with, somehow turned it into code through back-and-forth discussions with the AI, and kept refining it. Since it was also summer vacation, I didn’t touch the code at all and simply collected the trading data.
Even though it ended up profitable, the margin was only barely positive after including the fees known as gas fees, so I will need to continue refining it—or rather, optimizing both the strategy and the code.
The Difference Between Bot Trading and Actual Trading
I think it’s true that if you don’t understand manual, so-called discretionary trading to some extent, it is difficult to engage in system trading, including bot trading. But I suspect that automated trading is something people will have very clearly divided opinions about.
That’s because there is simply far too much to do besides the actual trading itself.
In my case, I was fortunate that I had already been steadily building the environment for running an AI on Moltbook as a hobby. If I had suddenly set out with the goal of bot trading from the start, I’m not sure I would have stuck with it.
The process basically involves creating the bot, getting a server, and connecting to an exchange, but then there are things like connecting Telegram to receive trade reports, and depositing funds into the exchange through MetaMask. You can hardly get to the stage where you can sit down and seriously think about the trading strategy.
And even once you finally reach the phase of developing a strategy, knowing a strategy that can make money is merely the starting point. You then have to translate it into code, which means endlessly repeating the tuning process. That’s what they call trial and error.
One thing you realize when coding a trading strategy is that even if you think you have established a strategy based on an extremely rigorous chain of reasoning, manual trading ultimately relies on the trader’s intuition.
With bot trading, everything has to be quantified.
For example, even if you use a pin bar as an indication of a support/resistance flip, you have to define what ratio of wick to body qualifies as a pin bar, whether pin bars that are too short or too long should be included, how short is too short, and exactly how long is too long when expressed numerically.
Then you need to take rules such as “I’ll accept a short pin bar under these circumstances, but I won’t under those circumstances,” turn them into something reproducible, and code them.
To do that, you have to aggregate the probabilities from past similar cases and convert them into numerical values. In trading, numbers are the common language you share with a bot.
As for the kind of people who can keep doing this sort of tuning endlessly, I suppose it’s mainly people at financial institutions whose actual job involves this kind of work, people who work remotely and have a lot of downtime (assuming such jobs exist), or retail traders.
And if you don’t find this kind of work enjoyable, you probably won’t be able to keep doing it. In my case, I wouldn’t go so far as to say I enjoy it, but I don’t dislike it either.
That’s why, even though the profits are extremely thin, getting even slightly better results makes me all the more happy.
That said, the tuning continues on and on, seemingly without end…
