Chapter 15: A New Direction in Research

Page 1 of 3

Cambridge, Massachusetts, United States. Across the Charles River, a cluster of red-brick and gray-stone buildings emerged and vanished in the morning mist.

This was the world-renowned Massachusetts Institute of Technology.

Inside one of the university’s classic red-brick buildings, a middle-aged man sat before a computer.

Kaiming He—an undisputed giant in the field of computer science.

His ResNet had become a foundational achievement in computer vision. Even papers published in such universally acclaimed journals as Nature and Science were outshone by ResNet; in terms of citations alone, it had surpassed the astonishing figure of one hundred thousand. It was, without question, an absurdly impressive number.

This also revealed how, in computer science, many traditional elite journals—including Nature, Science, and Cell—had gradually fallen out of favor.

And even if one removed ResNet, that epoch-making work, from his list of accomplishments, he would still unquestionably rank among the very best researchers in the field.

Although he was currently an associate professor in the Department of Electrical Engineering and Computer Science at MIT, he remained a Chinese citizen. He was also serving as a Senior Area Chair, or SAC, for NeurIPS.

His primary responsibility was to oversee the work of Area Chairs and reviewers, offering his own judgment whenever disagreements arose between them.

Normally, he rarely read papers himself. He would usually skim one, then assign it to suitable reviewers below him.

Yet the paper displayed on his screen, titled *AgileEdge: Adaptive Co-Optimization for Pervasive Low-Latency Edge AI*, had piqued his interest.

About an hour later, Professor He leaned back in his chair and let out a long breath. A trace of approval showed in his eyes. Although he had not read the paper in depth, he had already formed a favorable impression.

“Interesting work. I wonder who the authors are.”

Returning to the SAC management interface, he assigned the paper to four appropriate reviewers.

Even after delegating the paper to his reviewers, Kaiming He decided to examine it more thoroughly himself.

He poured himself a cup of coffee, took a delicate sip, reopened the paper, and began reading it carefully.

Far away on the other side of the Pacific, however, Zhou Yun had no idea how his paper was faring. At the moment, he was facing a more immediate problem—he was almost out of money.

Buying a road bike had consumed half his savings, and his living expenses for the past month had taken another third.

Now Zhou Yun had only a thousand yuan left. Unless he asked his family for help, that little sum would probably barely last until his next payment for contract work.

Supporting his head with one hand, Zhou Yun stared blankly at the computer screen.

Page 1 of 3

Page 2 of 3

“How am I supposed to make money?”

There were, in fact, quite a few ways. He could tutor students or take on some part-time work.

But that would be too troublesome. He was short of money, not starving to death. Rather than waste time on such things, he might as well ask his family for money.

“Is there any way to conduct research and make money at the same time?”

That would be the ideal solution.

Just then, a scream came from Qiu Yan beside him.

“Ah! Why did it fall again? If I’d known, I never would have bought this stock.”

Qiu Yan slumped lifelessly against his chair, his phone casting a sickly green glow across his face.

“Who told you to keep buying individual stocks? Unlike you, I only buy bank stocks. Even if they fall, they won’t lose that much.”

“Stop lecturing me, stop lecturing me! I’m never touching individual stocks again!”

Across from them, Shen Rui let out a cold laugh. “Didn’t you say that last time? And what happened? You bought them again.”

Qiu Yan curled his lips, looking somewhat sheepish. “It had already fallen this far. Was I supposed to pass up the chance to buy the dip? Who could have guessed it would keep falling?”

“So this is your version of ‘Be fearful when others are greedy, and greedy when others are fearful’? Do you think you’re Warren Buffett?” Yu Tianrui mocked mercilessly from nearby.

Faced with the ridicule of his fellow students, Qiu Yan had nothing to say.

“Don’t just talk about me. How did you do today?”

“Heh heh, you asked for it. I wasn’t even going to say anything. How did you know the stock I bought today hit its daily limit?”

As he spoke, he raised his phone. Its screen glowed bright red, forming a stark contrast with the green light shining from Qiu Yan’s phone.

“Don’t say another word. One of my friends is feeling a little unwell.”

Listening to the discussion among his senior colleagues, Zhou Yun’s eyes suddenly lit up.

Stocks plus deep-learning models—wasn’t that a way to make money while conducting research?

Moreover, this approach was already highly developed. It was known to everyone as quantitative trading.

At its core, quantitative trading relied on data and algorithms. By building mathematical models, it identified statistical patterns in the market and carried out automated trades based on those patterns.

Page 2 of 3

Page 3 of 3

Put simply, quantitative trading strategies based on complex models and algorithms could be abstracted as a function, Y = f(X).

X represented all kinds of market data, economic indicators, and other information that might influence market movements. This could include historical prices, trading volume, order-book data, news sentiment, corporate financial reports, macroeconomic data, and more.

Y represented the model’s trading decision or prediction, such as a buy or sell signal, position size, predicted price direction, or expected volatility.

This was essentially the same as a neural network. At its core, a neural network attempted to fit or learn a complex nonlinear function from vast quantities of data.

That was why quantitative trading was increasingly making use of neural networks, especially deep-learning models.

Of course, such models could never be open-sourced. They were, after all, the tools on which their creators made a living.

The more Zhou Yun thought about it, the more feasible it seemed.

If he could truly make it work, it would be no exaggeration to say that he would never have to worry about money again.

But meals had to be eaten one bite at a time, and research had to proceed step by step.

Training a model that encompassed the entire stock market was far beyond what his current resources could support. But if he trained a model capable of predicting the movements of a single stock, Zhou Yun felt that he could at least give it a try.

Once he had settled on his next research direction, Zhou Yun immediately got to work.

In the field of artificial intelligence, this kind of stock-prediction task had a specialized name: TSF, or time-series forecasting.

There were already many TSF algorithms. But if one applied them directly to stock-market prediction, losing every last penny would be considered a fortunate outcome.

After all, if things were really that simple, the researchers who developed these algorithms would all have become billionaires long ago.

The main difficulties could essentially be summarized in two points.

The first was obtaining vast amounts of high-quality historical data. That part was not particularly difficult, merely troublesome. Considerable processing was still required to organize the data into a form suitable for training.

The second was the algorithm itself. Many papers made their algorithms sound astonishingly powerful, but those results had been achieved on specific datasets.

In reality, many algorithms performed completely differently when transferred to another dataset, or even when the distribution of data within the same dataset was altered slightly.

Page 3 of 3