How Quants Actually Build Trading Strategies (and What AI Changes)

Most retail trading runs on vibes: “I’ve got a good feeling about Bitcoin.” Quants run on the opposite. They don’t draw trend lines or read indicators off a chart; they turn the market into numbers, model how those numbers move, and size bets by probability. I spent seven years in and around quant trading desks, and the gap between how a hedge fund operates and how a retail trader operates is mostly this one mindset shift, which AI is now putting within reach of far more people.

Here’s how quants actually build strategies, using a clean regime model as the example, and what changes now that an agent can do the heavy lifting.

Step one: quantify the feeling

The whole discipline starts by refusing to trade on intuition. Instead of “it’s windy today,” a quant wants a number: how strong, in which direction. In markets, one common way to do that is to define states. Sum the last 20 days of returns: 5% or more is a bull state, -5% or worse is a bear state, anything in between is sideways. Now every day in an asset’s history gets a label. It sounds elementary, but that labeling is what makes everything after it computable.

Step two: model how states move

The key assumption is the Markov property: tomorrow depends mostly on where you are today, not the whole path that got you here. Drive to New York from Arkansas or from Ohio and the route differs entirely, what matters is your current position. So quants count every historical transition, bull to bull, bull to bear, sideways to bull, and turn the tallies into a 3x3 transition matrix where each row (today’s state) sums to 100% across tomorrow’s possibilities.

Two things fall out of that matrix. The diagonal, bull-to-bull, bear-to-bear, captures persistence: states are “sticky,” which is the rigorous version of “the trend is your friend.” And to forecast further out, you simply raise the matrix to a power (square it for two days, cube it for three). Push that far enough and the probabilities converge into noise, the stationary distribution, which is itself a useful warning about how little signal lives in long-range prediction.

Step three: turn it into a bet

After all the matrix math, the trading signal is almost insultingly simple, and that’s the pattern with good quant work: complex machinery, simple output. Take the probability of a bull state tomorrow and subtract the probability of a bear state. A positive number says go long, a negative one says go short, and the magnitude sets position size. A +45% edge is a bigger bet than a +10% edge. Risk management isn’t a separate step; it falls out of the signal itself.

• Walk-forward testing. A naive backtest lets the strategy “learn from the future,” baking in data it couldn’t have known. Walk-forward testing recomputes the model day by day so it only ever uses information available at the time. It’s computationally brutal, which is exactly where modern compute and AI help.
• Let the data define the states. Picking “5% = bull” by hand is a subjective weak link in an otherwise mathematical system. More advanced approaches (hidden Markov models) let the data discover the regimes itself, then you check where the discovered states agree with your definitions for confirmation.

What AI actually changes

None of this is new, hedge funds have done it for decades. What’s new is the cost. The walk-forward computation that used to need a research team and serious infrastructure can now be expressed as a prompt or a reusable skill and run by an agent in minutes. That democratizes the mechanics. What it doesn’t democratize is the judgment: which features matter, how to avoid fooling yourself, when a number is signal versus noise. That’s the same lesson from every domain I work in, including building with Claude Code, the tool collapses the effort, but the operator’s judgment is still the edge.

This is the world I came from, seven years across quant trading and crypto before building AI products. See more about me or get in touch.