Chapter 16: Reversing the Celestial Force

After some thought, Zhou Yun logged onto the server and created a new folder.

He had originally assumed that his server account would be revoked after the previous project ended. He even specifically asked about it, since renting eight H100 GPUs was no small expense.

To his surprise, Deng Yonghua waved it off and told him to keep using it for now. Once the group purchased their own servers, they’d address it.

Zhou Yun couldn’t very well refuse.

Since the professor said to continue using it, then he would.

He had a clear idea of the technical side and had already thought of a few promising approaches. Whether they would work, however, still needed to be tested.

To run experiments, the most important thing was data. To get data, he needed to finalize which stock to predict so he could scrape the necessary information online.

“What stock should I choose?” Zhou Yun’s hands hovered over the keyboard but hesitated to type.

He wasn’t very familiar with stocks; mainly, he had never delved deeply into the subject. To him, the stock market was essentially a gambling den, a probabilistic game he generally disliked.

“Zhou Yun, what are you thinking about? Planning to buy stocks too?” Qiu Yan, who overheard the word “stock,” asked.

“Hmm, it’s not that I want to buy. I’m thinking—there are many quantitative firms now, but essentially, it’s all deep learning models. I wonder if I can pick a stock to predict with a model. If it works, maybe I could even make some money.”

Qiu Yan’s eyes lit up instantly. “That’s a great idea! If it really works, you might not make a fortune, but at least you won’t lose so badly.”

“Yeah! But it must be pretty hard to pull off, right?” Shen Rui felt it wouldn’t be so simple.

Zhou Yun nodded. “I have some technical ideas, but they need time to verify. The problem now is, I don’t know which stock to choose.”

“That definitely requires some thought. I have a few ideas that might help,” Yu Tianrui said, tapping the table thoughtfully. If anyone in the lab knew the stock market best, it was him.

He had turned ten thousand yuan into over thirty thousand in a year—a remarkable feat for an ordinary person, with an annualized return of 300%.

“Go ahead, senior,” Zhou Yun said humbly, always eager to learn in unfamiliar fields.

“Actually, I don’t recommend single stocks. Their movements are influenced by many complex factors—earnings surprises, executive changes, product failures, shareholder sell-offs—all unpredictable events. Not to mention that smaller stocks might be manipulated by big players. So, compared to stocks, I suggest ETFs.

ETFs’ trends are more driven by systemic factors, meaning individual influence is small—though there are exceptions. This makes technical indicators more reliable and predictions easier. Some experts can even roughly forecast their trends. I think a deep learning model could do even better.

Also, ETFs usually have low volatility; it’s rare to see a ten percent swing in a single day. So even if you lose money later, the losses can be controlled within a reasonable range.

I recommend two ETFs that I hold heavily: one is gold, the other is Hengke. Both support T+0 trading—you can buy and sell on the same day.

But the final choice needs your own research. I could be wrong; just take it as a reference.”

Zhou Yun jotted down notes while listening. “Thanks, senior. I’ll look into it myself.”

“No problem. These things are easy to understand even without being told.”

Though he had a general idea from the senior, Zhou Yun still did some research on his own.

He spent the next week studying the stock market but with little money, he could only trade on a virtual account.

The results were spectacular.

-18%.

An eighteen percent loss—an unsurprising bloodbath. Seventeen percent of the losses came from single stocks. Every time he bought, he thought the price wouldn’t fall further, but it did the next day without fail.

After research and practice, he finally decided to pick the gold ETF.

Mainly because he recalled that gold seemed to be on a steady upward trend in the future; even if the prediction wasn’t accurate, he shouldn’t lose money.

When the lab seniors learned Zhou Yun had a new idea, they didn’t hesitate to join, and Zhou Yun naturally welcomed them.

Although they couldn’t help much with the core model architecture or algorithms, data collection, cleaning, and labeling were easy tasks for them.

Usually, such chores fall to first-year graduate students, since data processing is nothing but a hassle.

This reversal was a bit ironic, but none of the seniors complained.

They all knew their own levels; Zhou Yun had just arrived and could already publish at NeurIPS—he was undoubtedly a big shot in their eyes. Helping a big shot with chores to earn a second or third authorship on a top conference paper was something no fool would refuse. While it might not help much with job hunting, it certainly aided scholarship applications.

As for age, in this day and age, it’s all about respecting those who are competent.

Deng Yonghua learned about this from the group meeting and did not object. He only reminded everyone to focus on their own primary research work first. Otherwise, if graduation itself becomes problematic, having a second or third authorship wouldn’t help at all.

Time quickly moved to July.

In the lab, the seniors gathered behind Zhou Yun, all eyes fixed on the screen.

Two interfaces were displayed: one was simple, showing only a single line graph; the other was trading software displaying the real-time gold ETF price.

At a glance, the two lines looked identical, but on closer inspection, many details diverged entirely.

This was the result of over two months of Zhou Yun’s team’s research—a brand-new architecture for time series prediction.

“There are still some issues, but this is the best we can do for now,” Zhou Yun sighed, staring at the two screens.

“I think it’s already excellent. An error under 0.1%—isn’t that enough?”

“Yeah, other models we ran had errors above 1%.”

Though they said so, Zhou Yun still felt the model had a lot of room for improvement. The main bottleneck was resources.

Even with eight H100 GPUs, there was a limit. To push predictions further, more data was needed—not just historical prices, but also news and information related to gold from around the world, in all formats: video, text, audio, images.

This involved another domain—a multimodal large model.

And once the words “large model” come up, eight H100 GPUs are far from enough, because training—not just inference—is required.