What AI in Finance Really Means: Comparing Games, Cars, and Markets

Chapter 2 of Artificial Intelligence in Finance is technically labeled a “Preface,” but it does a lot more than set the stage. Hilpisch opens with a quote from Robert Shiller asking whether financial markets will ever become truly perfect, with every asset priced correctly. It is a big question. And honestly, the way the chapter frames AI in finance around that question is what makes it interesting.

AI Does the Same Thing Everywhere (Sort Of)

Here is the thing that clicked for me reading this chapter. Whether you are talking about AI playing Atari games, driving a car, or trading stocks, the pattern is basically the same: you have an agent, it has a goal, it learns through trial and error, and it gets rewarded when it does well.

Hilpisch lays this out in a comparison table, and it is one of those moments where a simple table does more work than five pages of text.

When you see it laid out like this, the similarities are obvious. But the differences matter more.

With arcade games, there is zero risk. The AI plays in a perfect virtual environment. It can fail a million times and nothing happens. Nobody gets hurt. Nothing breaks. It is the ideal training setup.

Self-driving cars are the opposite extreme. Sure, you can start training in a virtual environment (Hilpisch mentions GTA as an example, which is kind of funny). But at some point, the car has to drive on real streets with real people. The jump from virtual to physical is the hard part. And the risks are serious.

Finance Sits in the Sweet Spot

So here is what makes finance interesting for AI. Trading is closer to the arcade game scenario than the self-driving car one.

A trading bot can train entirely in a simulated market. No one gets hurt if it makes bad trades during training. The worst outcome is financial loss, which is bad, sure, but it is not “self-driving car plows into pedestrians” bad.

Hilpisch points out that the financial domain is “an ideal place to train, test, and deploy AI algorithms.” And I think he is right about that. The feedback loop is clear (you make money or you lose money), the environment is digital, and you can simulate as much as you want before going live.

But here is the problem. Markets are not arcade games. Arcade games have fixed rules. Pac-Man works the same way every time you play it. Markets change. Other players adapt. What worked last year might not work next month. The environment itself fights back.

The book mentions that some hedge funds already go all-in on machine learning. Voleon, for example, had over $6 billion under management at the end of 2019 and relied exclusively on ML. Their return that year? 7%. The S&P 500 did almost 30%.

That is a sobering number. If one of the most well-funded ML-driven hedge funds can underperform a simple index fund by that much, it tells you something about how hard this problem really is.

A Student with a Laptop Could Do This

One of the more interesting claims in this chapter is that an ambitious student with a notebook and internet access could apply AI to financial trading. Not because the problem is easy, but because the tools have become so accessible.

Online brokers provide historical and real-time data. They offer APIs for executing trades. Python libraries for machine learning are free. Cloud computing is cheap. The barrier to entry has dropped dramatically.

I find this both exciting and a little unsettling. Exciting because it means anyone can experiment. Unsettling because “anyone can experiment with AI trading” is the kind of sentence that usually ends with someone losing their savings.

How the Book Is Structured

Hilpisch lays out the five parts of the book, and the structure tells you a lot about his thinking.

Part I covers AI basics: supervised learning, neural networks, and even the concept of superintelligence. He is not saying superintelligence is coming tomorrow. He is saying that thinking about it gives you a useful framework for understanding what AI can and cannot do.

Part II is about traditional finance theory and how data-driven approaches are changing it. This is where the book starts building toward what Hilpisch calls an “AI-first approach to finance.” Basically, throwing out the old models and letting the data speak.

Part III gets into the heavy stuff: deep learning, neural networks (dense and recurrent), and reinforcement learning applied to finding statistical patterns in markets. This is where you try to find signals that predict future market movements.

Part IV is about turning those signals into actual trading strategies. Backtesting, risk management, deploying algorithms in the real world.

Part V zooms out to the big picture. What happens when AI dominates finance? Is a “financial singularity” possible, where AI agents control all aspects of markets?

The Chain: Statistical Inefficiencies to Economic Exploitation

The most important idea in this chapter is the logical chain that runs through the whole book.

First, you need to find statistical inefficiencies. These are patterns in market data that an AI can detect. If a neural network can predict where prices are going with better-than-random accuracy, that is a statistical inefficiency.

Then, you need to turn those predictions into economic profit through algorithmic trading. This is harder than it sounds. Transaction costs, slippage, timing, and risk management all eat into your edge.

And if you can do both consistently? You have basically disproven the Efficient Markets Hypothesis (EMH), one of the most fundamental ideas in finance. The EMH says that markets are so good at processing information that you cannot consistently beat them.

So the book is really making an argument: AI might be the tool that finally breaks the EMH. Or at least shows where it cracks.

My Take

I like that Hilpisch is honest in the author’s note. He admits the book is opinionated, based on personal experience, and sometimes lacking proper theoretical support. He even acknowledges that experts like Francois Chollet (creator of Keras) flat-out doubt that prediction in financial markets is possible.

That kind of honesty is refreshing. Too many books in this space promise you a money-printing machine. This one tells you upfront: here are the tools, here are the challenges, and some very smart people think this whole endeavor might be impossible.

The Voleon example alone is worth remembering every time someone tells you AI has “solved” trading. $6 billion in assets, a team of PhDs, and they still got beaten by buying the S&P 500 and going to the beach.

Still, the comparison table from this chapter sticks with me. Finance really does sit in that sweet spot where the training environment is virtual, the feedback is clear, and the risk, while real, is manageable. If AI is going to master any real-world domain beyond games, finance seems like a reasonable next candidate.

Whether anyone will actually pull it off consistently is a different question entirely.

This post is part of a series retelling and reviewing “Artificial Intelligence in Finance” by Yves Hilpisch (O’Reilly, 2020, ISBN 978-1-492-05543-3). Views and commentary are my own.

Previous: Artificial Intelligence in Finance: A Book Worth Reading

Next: AI Algorithms: Types of Data, Learning, and Problems