Can AI Predict the Future? How Deep Research Differs From Quant

“Can AI predict the future?” sounds like a question for a fortune teller, but it’s one people genuinely type into search engines — usually with money or a sports bet on the line. The honest answer is more interesting than yes or no: it depends entirely on what kind of prediction you mean, and which kind of system you point at it.

Simon Du — Reasoning Chief Scientist at Apodex and a former quantitative-finance researcher before his ML career — has lived on both sides of this. Apodex builds deep-research agents that gather and verify evidence to answer hard questions, and Du draws a sharp line between how a classical quant model forecasts and how a reasoning agent does. The distinction is the whole story.

Two completely different kinds of prediction

The version of forecasting most people picture is quant finance: predicting a number from market data. Du worked in it. “For finance, when you try to predict future, usually the most informative information is from the market data — the price and volume, the bar chart you see on Robinhood,” he says. A quant model takes that structured data and outputs a number — say, a stock’s next-day return.

“Quant model is more like a classical machine learning model,” he explains. “Basically you’re predicting a number.” Even news gets folded in only after it’s “transformed into some features, still numerical.” The whole pipeline runs on structured inputs and produces a structured output. Predict better than 50% and, as he puts it, “you can generally make some money.”

That is not what a reasoning agent does.

Why an AI agent acts more like a fundamental analyst

A large language model is a generative system that reasons in natural language. So when it forecasts, it doesn’t behave like a quant model — it behaves like the other kind of financial researcher.

“Large model is a generative model. It generates natural language with reasons,” Du says. “That’s more like a fundamental analyst — those researchers who analyze a stock. You gather all the information, and it may not even be quantitative. It can be some news, some words from the CEO, and you make some human judgment.”

This reframes “AI predicting the future” entirely. Forecasting a company’s outlook means gathering everything about its upstream and downstream sectors, its industry, its earnings trajectory — and reasoning over it. “If you think about it, this is exactly what deep research is about. You’re trying to gather all the information about this company, and then predict the future of this company.”

And it generalizes beyond finance. “You can also predict the outcome of a World Cup match,” Du notes — by gathering whether the players are strong, how good the coach is, whether anyone’s injured. Same machinery: gather evidence, reason, judge.

Where the AI has an edge over a human analyst

If the agent is doing fundamental-analyst work, the obvious question is whether it does it any better than a human analyst. Du’s answer is that on one axis it clearly does.

“Large model, in terms of search capability, is much stronger than human right now,” he says. “It can do multi-hop search, and you can run 24 hours, so it can analyze a lot more information than a human.” A human analyst reads a finite number of sources in a workday. An agent can run continuously and chase information across many hops.

But — and this is the load-bearing caveat — that edge only exists if you train for it. “You need to train the model to have this judgment, reasoning, analysis capability,” using expert analysis data to mimic how real analysts think, or reinforcement learning that rewards correct predictions. The raw model doesn’t forecast well; the trained reasoning does.

The practical takeaway

If you’re confused by “AI predicting the future,” the useful split is this. For tick-by-tick price moves from market data, a quant model is the right tool — that’s number prediction. For an open-ended outlook that depends on news, context, and judgment, an LLM-based deep-research agent fits, because that task is fundamental analysis, and fundamental analysis is deep research. One last caution from the demo in the conversation: a market-implied betting line (“USA 62%, Australia 19%”) is the market’s number, not the agent’s own forecast — don’t confuse pulling a figure with predicting it.

FAQ

Can AI actually predict the future?

For some tasks, usefully — but it’s better described as deep research than fortune-telling. An LLM agent gathers evidence about a company, event, or match and reasons to a judgment, the way a fundamental analyst does. For tick-by-tick market moves, a classical quant model predicting a number is the right tool instead.

What’s the difference between a quant model and an AI agent for prediction?

A quant model is a classical ML system that predicts a number (like next-day return) from structured market data such as price and volume. An AI reasoning agent is generative — it gathers unstructured information like news and CEO statements and reasons to a judgment, closer to a fundamental analyst than a number-cruncher.

How does AI predict a stock’s future?

By behaving like a fundamental analyst: gathering information about the company’s upstream and downstream sectors, industry, and earnings trajectory, then reasoning over it to form a view. Simon Du frames this as deep research — assembling all relevant evidence about a company to project its outlook — rather than predicting a price from market data alone.

Can AI predict sports outcomes like the World Cup?

The same deep-research approach applies: gather whether players are strong, how good the coach is, whether anyone is injured, and reason to a judgment. Du notes that forecasting a match is structurally like forecasting a company — both come down to gathering all the relevant information and making the right call.

Is an AI agent better than a human analyst at forecasting?

On search and breadth, yes — it can run 24 hours and do multi-hop research across far more sources than a person. But that advantage only exists if the model is trained for judgment and reasoning, using expert analysis data or reinforcement learning. The raw model doesn’t forecast well on its own.

Why is fundamental analysis the same as deep research?

Both involve gathering wide-ranging, often unstructured information — sector dynamics, news, statements — and reasoning to a forecast. Du points out that analyzing a company to predict its future is exactly the deep-research workflow: assemble all the evidence, then judge. That’s why a reasoning agent maps onto an analyst’s job, not a quant’s.

Does AI use market data to predict prices?

A quant model does — it transforms price, volume, and even news into numerical features and outputs a number. An LLM-based research agent works differently, reasoning over unstructured information in natural language. The two approaches suit different prediction problems: short-horizon price moves versus open-ended outlooks.

How do you train an AI to make good predictions?

By giving it judgment, not just data access. Apodex uses expert analysis data so the model can mimic how real analysts reason, and reinforcement learning that rewards correct predictions. Without this training, the search advantage is wasted — the model can gather information but won’t weigh it well.

Is a market-implied probability the same as an AI prediction?

No. A market-implied figure (like odds showing one team at 62%) reflects what a betting market is pricing, not an AI model’s own forecast. When an agent surfaces such a number, it’s reporting the market — distinct from the agent reasoning to its own prediction. The distinction matters when judging accuracy.