Chapter 36: No Mechanics, Only Stats

A few days later, in the laboratory, Zhou Yun nodded with satisfaction as he looked at the screen, which displayed an average error of only 0.05% and a prediction accuracy rate as high as 80%. (This error rate already accounts for transaction costs.)

The so-called average error refers to the difference between the price predicted by the model and the actual price.

This was undoubtedly a staggering figure.

According to his estimates, even the world’s top quantitative models usually hover around a 0.1% error rate, with prediction accuracy of only about 60%.

As for why such seemingly mediocre accuracy and error rates can still allow top quantitative models to rake in huge profits?

It is quite simple: it is like flipping a coin. The model’s job is to guess whether the next second will go up or down. Suppose a top-tier model has a 55% probability of guessing correctly.

A model is not a human; it can make dozens or hundreds of guesses in a single second and execute trades automatically.

Every time it guesses correctly, it earns a very, very small amount of money—for example, on a 100-yuan item, a correct guess might yield only a few cents. If it guesses wrong, it immediately cuts its losses, also losing only a few cents.

In a single day, it can make hundreds of thousands or even millions of guesses, each with a 55% probability of success.

After deducting losses, the daily profit is a massive sum. This is what is known as "accumulating small gains into a fortune."

With an average error of 0.05% and an 80% prediction accuracy, his model has a much higher "winning edge," allowing for more precise control over the profit or loss of every trade.

To put it in one sentence: it is not about strategy; it is all about the numbers.

However, it was too early to celebrate. Although it performed well on the test set, a simulation is, after all, just a simulation.

Practice is the only criterion for testing truth.

He organized the new data from the last ten days and fed it into the model to begin retraining. With more than ten hours left before the US stock market opened, there was enough time for fine-tuning.

If the experimental results were successful that night, this preliminary research project could effectively be declared complete.

At 9:20 PM, for the first time ever, Zhou Yun did not return to his dormitory but stayed in the lab.

Although his fellow students were curious, seeing Zhou Yun’s serious expression, they felt it inappropriate to disturb him.

After checking the program one last time to ensure there were no issues, he leaned back in his chair, closed his eyes, and waited for the final ten minutes.

Time ticked away, second by second.

Buzz~ Buzz~

His phone vibrated in his pocket. It was the alarm he had set in advance.

Zhou Yun opened his eyes and subconsciously swallowed.

On the left side of the screen, real-time stock market data began to pulse. Two dots, one blue and one red, appeared simultaneously: the blue one was the actual market price, and the red one was the predicted price.

At the opening price, the two dots nearly overlapped—they were in perfect agreement!

The error compared to reality was less than 0.01%.

But if one observed closely, it would be clear that the red dot appeared about a second earlier than the blue one.

From this moment on, another function of the model began to operate, which Zhou Yun called "real-time calculation."

The model performs real-time fine-tuning based on live stock data. While existing technology can technically achieve this type of online learning, the brilliance of OracleNet lies in its ability to compress this time to the millisecond level. In a field as time-sensitive as the financial market, this is incredibly fast.

Coupled with the model's inherent predictive capabilities, even with this delay, the model's projected data still ran ahead of the real-time data.

As for why real-time calculation is necessary, the reason is simple: the market changes every moment, and relying solely on historical data can lead to certain deviations.

Assuming historical data covers everything up to time *t*, the model can predict the price at *t+1* very well, but predicting the price at *t+2* might result in minor errors, such as stock price fluctuations caused by sudden large transactions.

If the data between *t* and *t+1* can be used to fine-tune the model, these minor differences can be reduced or even eliminated.

As time passed, although the two lines developed some slight deviations in the middle, they remained nearly perfectly aligned.

At 11:00 PM, the two lines still hadn't diverged significantly.

At this point, Zhou Yun was the only one left in the lab.

He waited until 4:00 AM. Outside the window, the sky was beginning to brighten, and the lines stopped changing.

On another screen, the average error was displayed at only 0.049%.

The corners of his mouth curled upward uncontrollably—it was a success!

Just as he had anticipated, the results on the test set were solid. While it might not yield such perfect results every single time, it wouldn't deviate much, and at the very least, it would significantly outperform existing top-tier quantitative models.

Moreover, AI prediction does not completely replace humans; quantitative models are just one part of quantitative trading.

How to use the model is ultimately a human decision.

With the predictions from OracleNet, it was like taking a test with the answer key in hand. The final score might fluctuate slightly depending on the grader's mood, but it was impossible for it to be poor—even if the answer was known only one second in advance.

He leaned back, took a deep breath, and let it out slowly. His heart was pounding violently, though he couldn't tell if it was from staying up all night or from sheer excitement.

The computer screen remained bright, the once-pulsing lines now frozen in place.

Looking at the beautiful curve, even after an all-nighter, Zhou Yun felt little fatigue.

He dragged the mouse and saved the trained model to the server.

One successful result could only prove so much; he needed more experimental data.

Over the next few days, Zhou Yun essentially flipped his sleep schedule, staring at experiments all night and resting during the day.

His classmates and roommates began to wonder if he had secretly gone off to practice some form of cultivation.

For five consecutive days, the experimental results showed no major differences.

The excellence of OracleNet was indisputable.

At 6:00 AM, with the lab still empty, Zhou Yun looked at the experimental data on the screen and made a call.

The call went through, and after three rings, it was quickly answered.

A slightly hoarse male voice came from the other end: "Hello?"

"The research has made some progress, and the results are quite good. I think we are ready for the final inspection."

"Really?" The voice on the other end became much more alert, though still tinged with skepticism, as it was still long before the agreed-upon deadline.

"Yes, you can come for the inspection whenever you like."

"Hiss..." A faint intake of breath came from the phone. "Alright. It’s not convenient to discuss over the phone. I’ll bring people over tomorrow, and we can talk in person."

"Fine, I’ll wait for you."

The call ended.

Zhou Yun stretched, rotating his slightly sore neck.

Come to think of it, since his rebirth, he hadn't experienced such high-intensity work. While waiting for them to arrive, he could finally get some proper rest.