The Story of How I Tested AI Swing Trading and Lost Miserably to 'Just Holding'

Introduction

Hello, I’m a sloth trying to improve my life through stocks.

Up until now, I have been running simulations where I have an AI predict whether a stock price will rise in 10 business days and trade every two weeks. This time, I am verifying something I have been curious about for a long time.

In the first place, what would have happened if I had just bought everything at the start and slept without doing anything?

For a sloth, this is the most interesting comparison. I also checked whether ‘AI predictions are actually better than picking with dice’.

Conclusion first

1. Lost miserably to ‘doing nothing’ by a margin of about 2x: The 10-year annual return was 12.7% for doing nothing versus 5.0% for the AI.

2. AI predictions are slightly better than dice. But you can’t call it ‘skill’: It beat about 90% of the trades chosen by dice, but the possibility of it being a coincidence remains.

3. The real cause of the loss was the ‘trading method’: Rather than the AI’s intelligence, the rules of leaving money idle, paying taxes every time, and going all-in on one stock were the causes of the loss.

From here, I will look at each one in detail.

Rules of this match

The following two are competing.

AI Swing

• What to do: Buy stocks the AI says ‘look like they will rise’ and sell them 10 business days later

• Number of trades: Hundreds of times over 10 years

• Taxes: About 20% every time a profit is made

Doing nothing (Buy & Hold)

• What to do: Buy all stocks in equal amounts on the first day and hold them until the end

• Number of trades: Only once to buy and once to sell

• Taxes: Only once when sold at the end

Here are the conditions I set.

• Target: 9 large-cap financial stocks (3 megabanks, Resona, Sumitomo Mitsui Trust, Nomura, SOMPO, ORIX, Aozora Bank)

• Period: Approximately 10 years from January 2016 to December 2025

• Initial Capital: 1 million yen

• All trades are executed at the ‘first price of the morning (opening price)’

Note that the ‘buy and hold’ side assumes buying in 1-share units. If I used 100-share units, some stocks would be impossible to buy once the 1 million yen was split among 9 companies.

Conclusion 1: Lost miserably to ‘buy and hold’ by a factor of about 2

Here are the results.

• Final Assets (after tax): AI Swing approx. 1.62 million yen / Buy and Hold approx. 3.30 million yen

• Annual Return: AI Swing 5.0% / Buy and Hold 12.7%

• Maximum Decline (Maximum Drawdown): AI Swing -42.7% / Buy and Hold -45.6%

• Volatility (Annualized Volatility): AI Swing 26.3% / Buy and Hold 26.2%

The ‘buy and hold’ strategy resulted in about double the returns. Moreover, the volatility was almost identical. I experienced the same amount of anxiety for less than half the money.

Out of 10 years, the AI only won twice: in 2018 and 2020. Both were years when the overall market fell, and the AI benefited from spending a lot of time in cash without buying. Conversely, from 2021 to 2024, when financial stocks were rising steadily, it could not keep up with ‘buy and hold’ at all.

Conclusion 2: AI predictions are slightly better than rolling dice. But it cannot be called real skill

What bothered me here was the question, ‘Did it lose because the AI’s predictions were bad?’ So, I had it compete against ‘random trading,’ where stocks are chosen by rolling dice instead of using AI.

The trading rules (selling after 10 business days, maximum 3 stocks, 100-share units, etc.) were exactly the same as the AI. The only difference was ‘which stocks to buy.’ Since luck is involved, I repeated the dice-rolling trades 500 times each.

• AI Swing: approx. 1.62 million yen (5.0% annual return)

• Dice Trading A (Same days as AI, but stocks chosen by dice): approx. 1.06 million yen (0.6% annual return)

• Dice Trading B (Both days and stocks chosen by dice): approx. 1.16 million yen (1.5% annual return)

*Dice trading results are the median of 500 trials

The AI beat dice trading A in about 90% of the 500 trials, and B in about 83%. It seems better than rolling dice, at least.

However, the dice beat the AI in the remaining 10-20%. Statistically, this hasn’t reached a level where it can be called ‘not a coincidence,’ and it’s not enough to boast about the AI’s ‘true ability’.

Furthermore, the AI’s buying criteria (a prediction probability of 56% or higher) was a value chosen to maximize performance over these same 10 years. Since it’s like deciding after seeing the answers, it’s more honest to assume its true ability is a bit lower.

Conclusion 3: The real cause of the loss was the ‘trading method’

This is the most important discovery this time. Let’s compare the annual returns.

• Buy and hold: 12.7%

• Dice trading A: 0.6%

• AI swing trading: 5.0%

Even though it’s the same 9 stocks, the moment you set a rule to ‘buy and sell every two weeks,’ the annual return drops by about 12 points with dice. The AI only recovered about 4 points of that. Even the top 5% of dice traders who were lucky only made about 2.21 million yen, which didn’t reach the 3.3 million yen from just holding.

In other words, the cause of the loss was not the AI’s predictions, but the trading method itself. Specifically, these three things:

1. One-third of the money was napping: On days when the AI didn’t say ‘buy,’ it remained as cash. On average, about 34% of the funds were not working, and I missed out on the large rise in financial stocks.

2. Taxes are incurred every time: Every time a profit is made, about 20% is deducted, and that amount doesn’t go into the next investment. Since ‘buy and hold’ only pays once at the end, compound interest works for the entire 10 years.

3. Going all-in on almost one stock: The current rule is ‘buy as much as possible of the initially selected stock.’ Even though the price movement is as volatile as ‘buy and hold’ which is diversified across 9 stocks, the return was less than half.

In sloth terms, it’s a state of ‘I worked hard to move, but I lost money for every bit I moved’.

Summary

Instead of using AI to trade every two weeks, I made about twice as much money by buying at the start and sleeping. However, the reason for the loss was not the AI’s brain, but how the money was moved.

I learned two things from this verification.

• Have a benchmark to compare against: If I had only looked at the AI’s performance, I would have ended with ‘5% annual return, not bad.’ Only by comparing it to ‘buy and hold’ and dice did I see its true ability.

• How you move money is more important than predictions: Rather than slightly improving predictions, it seems more effective to review idle funds, taxes, and concentrated investment.

Honestly, as a lazy person, the conclusion that ‘sleeping and waiting is the best’ leaves me with mixed feelings. However, it is also a fact that the AI did manage to beat a roll of the dice. I don’t think it’s time to give up just yet.

Next time, I plan to change the rules to divide funds equally among stocks instead of going all-in on one, and verify how close the AI can get to a ‘set it and forget it’ strategy.