Nice to meet you, I’m Peko! I usually use machine learning for Nikkei day trading! This very week (today, November 13th) is the peak of Japanese stock earnings announcements. With earnings for major stocks, starting with mega-banks, coming out almost every day, I see many posts on social media about ‘huge profits/losses from holding through earnings’ and ‘getting burned by jumping into intraday earnings’.
Actually, when I build stock models using machine learning, there is one thing I do first in the data preprocessing stage. That is the task of ‘excluding data from earnings announcement days‘ (or completely separating them from the usual market).
Why shouldn’t AI be trained with earnings days mixed in? And why do humans continue to lose in total when they hold through earnings or jump into intraday earnings like locusts? Today, I will delve into this from the perspective of statistics, not mentalism.
1. Earnings days change the ‘shape of the probability distribution’ into something else (the trap of bimodal distribution)
When you collect a lot of stock price movements from ordinary days with nothing happening and create a graph (histogram), it usually forms a ‘single peak’ shape where ‘around 0% (flat to slight movement)’ is the most frequent, and most data falls within plus or minus 1-2%, with the center being the highest.
Many of the technical indicators and machine learning models we use daily are built on the premise of this ‘usual single-peak distribution’.
However, when you collect only the days following earnings announcements and draw a graph, this shape completely collapses. The center (around 0%) is hollowed out, and it changes into a so-called ‘bimodal distribution‘ (a camel-hump distribution) with peaks at both ends of ‘a 10% surge’ and ‘a 10% plunge’, or a distribution with extremely wide tails.
As a basic principle of statistics, thinking about the ‘average value’ for this kind of camel-hump data is almost meaningless. For example, the average value of a game where ‘half is plus 10% and half is minus 10%’ is ‘0%’, but in reality, it almost never ends at 0%, right?
If you mix this earnings day data into a machine learning model as is, the AI’s calculations will be pulled by the intense outliers of ‘plus or minus 15%’ in just one day, and even the delicate predictions for ordinary days will all be ruined. That is how much earnings days are ‘a different game operating under different probability laws‘ than regular stock trading.
2. ‘Negative skewness’ and ‘volatility tax’ lurking in holding popular stocks through earnings
Furthermore, what was symbolic in this week’s earnings is the pattern where popular stocks that had been bought up due to expectations for performance, like Sanrio, are sold off significantly after earnings.
In statistical terms, this is a game with ‘negative skewness‘. Skewness refers to the way a distribution leans to the left or right. Stocks that have risen in price due to expectations before earnings and have accumulated a lot of margin buying are prone to the following extremely skewed conditions:
-
Good earnings as expected by the market: No movement because it’s priced in, or a slight minus due to ‘sell the fact’
-
A super surprise that exceeds expectations: Finally a plus 5% to 8% rise
-
Falling short of market expectations (consensus) even slightly: A 15% crash due to disappointment selling and the dumping of margin buying
In other words, the upside potential is limited, but the downside tail (tail risk) is abnormally long, creating a disadvantageous skew from the start.
Moreover, there is another mathematical trap here called ‘volatility drag‘ (asset reduction due to variance).
For example, suppose you hold through earnings twice and ‘the first time is a big win of plus 20%, and the second time is a big loss of minus 20%’. Looking at the arithmetic average, it looks like ‘plus or minus 0% (a draw)’. But since actual money moves by multiplication, after 1 million yen becomes 1.2 million yen, a 20% decrease makes it ‘960,000 yen’, and ‘even with a 50% win rate and an average of 0%, your funds have decreased by 4%‘.
Statistically, gambling where the range of movement (variance) per turn is too large has a structure where this ‘multiplicative reduction’ acts like a tax as you repeat it, and unless you maintain a considerably high win rate, your assets will be mathematically eroded.
3. The ‘clock bug’ and ‘negative autocorrelation’ where intraday earnings locusts catch the top
And a particularly common pattern recently is seeing ‘intraday earnings’ announced during trading hours, such as at 1:00 PM or 2:00 PM, jumping into the first large bullish candle on a 1-minute chart thinking ‘this is the initial move!’, only to find that it was the absolute peak.
Why does it become the final destination even though you intended to enter at the ‘initial move, only one minute after the announcement’? Two mechanisms often used in financial machine learning are involved here: ‘volume clock‘ and the statistical concept of ‘autocorrelation‘.
First, the ‘1-minute’ or ‘5-minute’ charts we usually look at are just charts mechanically divided by the hands of a clock. When dealing with markets in machine learning, we often discard these ‘time charts’ and convert them into ‘dollar bars‘ (amount-based charts), where one candle is drawn for every fixed amount of trading value.
When you apply this to the moment of an intraday earnings announcement, you can see exactly what is happening.
In the single minute immediately after an intraday earnings announcement, a tremendous amount of volume, equivalent to two hours or half a day of normal trading, surges all at once. In other words, even if it looks like ‘just one minute after the announcement, the first bullish candle has just appeared’ on a human clock, when viewed on a market clock (dollar bars), it is actually ‘the very end of the move, where trading equivalent to 100 candles has already passed and the trend is completely over‘.
Furthermore, this is compounded by what is known in statistics as ‘negative autocorrelation‘ (the property of being strongly pulled back in the opposite direction of the immediate preceding movement).
At the moment of an earnings announcement, market makers (algos providing liquidity) who usually line up limit orders on the board withdraw their orders all at once to avoid risk. Because market buy orders from chasers flood in while the board is thin and empty, the stock price momentarily jumps far beyond its true fair value (overshoot).
When you take statistics, such sudden spikes at the moment liquidity disappears have ‘strong negative autocorrelation‘, and there is a very high probability that it will be pulled back sharply toward the fair price (becoming a large bearish candle) several tens of seconds to several minutes later, once the board begins to return to normal.
Interestingly, the phenomenon where institutional buying enters stocks with genuine earnings surprises and they continue to rise gradually (positive autocorrelation called ‘post-earnings drift‘) appears slowly from the next day onward, once the panic of the day has settled. In other words, those who jump into intraday earnings are using the most wasteful strategy possible: diving into the first minute where they are statistically most likely to get ‘fully retraced (negative autocorrelation)’ and get burned, and then not touching it from the next day onward when the real edge appears.
Summary
The timeline during earnings season looks fun like a festival, but if you look behind the data, all these disadvantageous conditions are present.
-
The distribution becomes bimodal, making normal averages and intuition unreliable (bimodal distribution)
-
The higher the expectations for a popular stock, the greater the downward skew, and the intense price movement itself multiplies the erosion of capital (negative skewness and the variance trap)
-
Even if it is one minute later on the clock, it is already the end stage on a volume basis, and it is easily pulled back from the overshooting caused by a thin board (volume clock and negative autocorrelation)
Since this is a time zone that even AI avoids, considering it an ‘anomaly where normal laws do not apply,’ I think it is better to stop jumping into the 1-minute chart immediately after an announcement and wait until the panic has passed and the board and distribution have settled; in the long run, you are much more likely to keep your money.
I recently started a YouTube channel, so please check it out if you have some free time! https://www.youtube.com/@pecopekocandy
