What you will learn in this article
-
I worked with AI to analyze 5 years of price movements to see if there is a way to win at day trading.
-
Since I have limited time, I decided on simple conditions and a low frequency of trades. With that in mind, I arrived at one specific method.
-
First, I will verify it with demo trading (a record of trades without using real money). I will record the results every business day starting from September 24th.
Usually, I use a strategy of analyzing with AI to hold stocks with a high probability of winning for the long term to grow my profits. While I will continue doing that, I suddenly thought: Might day trading be faster?
Long-term holding takes years to show results. With day trading, you get results within the same day. If I can find even one ‘law for winning with high probability,’ I should be able to accumulate profits over a short period.
And I thought this was a task AI excels at. Searching through thousands of stocks and years of price movements to find a winning pattern. I couldn’t handle that on my own. I think of the conditions, and the AI does the testing. With that division of labor, I might find a winning method. That is why I started.
Premise: Limited time
I cannot watch the screen all day on weekdays. Therefore, day trading that requires constant buying and selling every day is impossible for me to begin with.
So, I decided on these two things first:
-
Keep conditions simple. Deciding whether to ‘do it or not’ based on the morning candidates should take only a few seconds.
-
Keep the number of trades low. Only trade on days that meet the conditions. Do nothing on days that don’t.
Before asking ‘can I win?’, I narrowed it down by asking ‘is this a format I can continue every day?’
I tested hundreds of combinations of ‘what kind of days lead to wins’
Here is how I researched it. Using the price movements of the past 5 years, I tested every possible scenario of ‘buying (or selling) a certain type of stock in the morning on a certain type of day and closing the position within the same day’ to see if there were any combinations that left a profit after deducting commissions.
I only used these materials:
対象 約3,750銘柄（1日の売買代金が20日平均で1億円以上・株価100円以上）
期間 2021年10月〜2026年9月17日（1,222営業日）
回数 のべ205万回ぶんの「その日・その銘柄」（同じ会社が日ごとに何度も入るので、重複ありの数です）
売買 朝の寄り付きで入って、その日の引けで手じまう（＝売買を終わらせる）
費用 往復0.3%を引く（手数料＋思ったより悪い値段で約定するぶん）I limited the conditions to things known by the previous day. The previous day’s price movement, volume, whether the closing price was closer to the high or low, the previous day’s movement of the Nikkei 225 or TOPIX (an index representing the entire Tokyo Stock Exchange), whether it was the day after an earnings announcement, the size of the trading value, and the market segment. Plus, the ‘distance from the previous day’s closing price’ known at the morning opening.
I tested 128 ways to combine these materials, 144 ranges for profit-taking and stop-loss (ending when a profit is made / ending when a loss is made), and also tried different industry sectors and market directions. There were hundreds of combinations in total. Since each test required recalculating 2.05 million trades, I left this work to the AI.
I arrived at one method
As a result, there was only one pattern where the average remained positive after deducting costs.
There is one condition determined by the price movement up to the previous day. On days that meet that condition, I decide whether to enter by looking at the morning price, and close the position by the market close. I will write about the details of the condition separately once I have accumulated enough records.
The figures for the past 5 years were as follows.
出た回数 1,821回ぶん・640銘柄・1,212営業日のうち769日（1日あたり2.4件）
勝率 54.4%
平均 +0.96%
真ん中 +0.70%From 2021 to 2026, even when looking at it year by year, all 6 years were on the profitable side. It is not a case of having a big hit in one specific year and losing in the rest.
I also verified whether I just happened to choose a good condition after seeing the results. I selected the best-performing condition using only the data from the first half (up to March 2024) and tested it with the data from the second half, and the performance remained good.
With this format, you can continue even if you have little time. In the morning, candidates are determined by the price movement up to the previous day. I look at the opening price, and if it meets the condition, I enter; if it does not, I do nothing. That is the only judgment required, and there is no need to watch it during the day.
Record every day for 1 month
I will first verify whether this method can truly win through demo trading (a record of trades without using actual money). I will record one month’s worth of data every business day starting from September 24th. This is because I still lack sufficient grounds to put in my own money right away.
資金 100万円（デモ）
建玉 1回に使う金額。100万円まで。資金が100万円増えるごとに100万円ずつ増やす
銘柄 1日1つ。条件に合ったうち、いちばん強く当てはまるもの
入る 朝の寄り付きの値段
出る 決めた幅に届いたらそこで。届かなければ引け
費用 往復0.3%を引く
期間 1か月（10月23日ごろまで）。途中で条件は動かさないCandidates will arrive on my smartphone every morning at 8:30, and I will record the results in the evening. I will write one post every day after the market closes for that day’s trades. I will write these four things:
-
Today’s stock
-
Trading results (demo)
-
Today’s profit/loss
-
Total so far
The pace is 2.4 items per day on days that meet the condition, so there will be days where there is “nothing today.” I will write on those days as well.
After one month, I will look at three things: whether the total is positive, how the worst day went, and what percentage of those trades were actually executable to decide whether to continue, stop, or change the conditions. I will also consider whether to put in actual money at that point.
Three things I verify myself
-
First, narrow it down by “is it a format I can continue?” Can I make the same judgment every day with the time I have available? Whether I can win or not comes after that
-
Count the costs. What percentage of the round-trip cost is the commission and price slippage for one trade? This time it was 0.3%, and that was the dividing line for whether it would remain profitable
-
Select in the first half, verify in the second half. If you choose the best one for the entire period, you will end up picking something that was just lucky. I select in the first half and see if it remains profitable in the second half
However, this part has not been verified
-
The period is 5 years from October 2021, which is a long period of a bull market. I do not know if it will be the same in a bear market
-
Profits were concentrated on a few days (if you remove the top 5% of days out of 1,821, the average becomes almost zero).
-
I have not been able to verify whether those trades could actually have been executed for those stocks at the time (due to regulations, etc.).
-
This assumes that morning market orders were always filled at the desired price. In reality, they will deviate depending on the depth of the order book.
I will verify and record every day whether this is truly a winning mindset. I will be writing every day starting from September 24th!
About the rules of this article: The details of the conditions (what I look at and at what price I enter) will be written separately after a month of records have been accumulated. All numbers are calculated based on past daily charts and are records of demo trading.
*This article is a verification of my own data and a record of demo trading, and does not recommend the buying or selling of any specific stock.
