Learning Paradigms Compared

Think about three ways a new analyst could learn to read company filings. In the first, someone hands them a thousand filings each already stamped "fraud" or "clean" by an auditor, and the analyst studies the pattern separating the two piles. In the second, nobody stamps anything — the analyst just reads the filings and notices, unprompted, that they naturally cluster into a few recognizable styles. In the third, the analyst trades on hunches from filings and only learns days later whether a call made or lost money, with no filing ever labelled directly. Same skill, three different kinds of feedback. Machine learning has the same three shapes, and the name for each is the paradigm.

Supervised learning: learn the map from example to answer

Every training example comes as a pair $(x,y)$ — an input $x$ and a known correct output $y$. The model's job is to approximate $f$ such that $f(x)\approx y$, graded directly against $y$ at every step. This is the paradigm behind almost every classifier and regressor in day-to-day quant work: predicting a return, classifying a regime, forecasting a default. Its strength is precision — the feedback is unambiguous — its weakness is cost: every $y$ had to come from somewhere, usually expensive labelling or a future outcome that took real time to arrive.

Unsupervised learning: find structure with no answer key

Training examples are just $x$, with no $y$ at all. The model is not told what the "right" grouping or representation is; it is asked to discover regularities on its own — clustering similar instruments together, finding a lower-dimensional representation that preserves the important variation (as in PCA), or estimating the underlying probability distribution the data came from. There is no ground truth to check against, so success is judged indirectly: does the discovered structure hold up, is it useful for a downstream task, is it stable when re-run on new data. This is the paradigm behind clustering correlated assets into risk buckets or discovering latent factors nobody hand-specified.

Reinforcement learning: learn from delayed, evaluative feedback

An agent takes a sequence of actions inside an environment, observes a state after each one, and receives a reward — a scalar signal, not a labelled correct answer — often only after a long chain of actions, and only telling you how good the outcome was, never what the right action would have been. There is no $y$ to imitate directly; the agent has to work out, through trial and consequence, which actions tend to lead toward higher reward. This is the paradigm behind an execution algorithm learning to slice a large order, or a market-making agent learning to quote — it never sees a labelled "correct spread," only the P&L that resulted from the spread it chose.

This explorer is a supervised classifier — every point already carries a known label, and the model's job is to draw the boundary that separates them. Notice there is nothing analogous to feed an unsupervised or reinforcement learner here: an unsupervised algorithm would receive the same dots with their colors erased, and a reinforcement learner would receive no dots at all, only a running score after each guess.

Worked example 1: the same data, three ways

Take five days of a stock's return alongside its trading volume that day: $(2\%, 1.1x), (-1\%, 0.9x), (3\%, 2.0x), (-2\%, 1.8x), (0.5\%, 1.0x)$, where "volume" is expressed relative to average.

Supervised framing: if each day also carries a label "next-day up" or "next-day down," the model learns to predict that label from today's return and volume — direct, gradable, one right answer per row.

Unsupervised framing: strip the labels entirely and just hand over the $(return, volume)$ pairs. A clustering algorithm might notice two natural groups — $(2\%,1.1x)$ and $(0.5\%,1.0x)$ sit together as "calm days," while $(3\%,2.0x)$ and $(-2\%,1.8x)$ sit together as "high-volume, large-move days" — a grouping nobody specified in advance, discovered purely from the shape of the data.

Reinforcement framing: an agent sees only today's return and volume as its state, picks an action (buy, sell, hold), and three days later receives a single reward number — the realized P&L of that decision, contaminated by everything that happened in between. It has to work backward from that one delayed number to figure out which of its many decisions actually deserved credit.

Worked example 2: counting the feedback each paradigm needs

To train on 10,000 historical days, supervised learning needs 10,000 labels — known next-day outcomes, each requiring the future to have already happened. Unsupervised learning needs zero labels; the raw feature rows are the entire dataset. A reinforcement learner needs no labels either, but must actually act across those days and receive reward signals that are noisy and only weakly attributable to any single decision — so despite needing "less labelling," RL frequently needs far more total data, because each reward carries less information than a clean label did.

What this means in practice

Most production quant models are supervised, because a known past outcome usually exists and gives the cleanest training signal per example. Unsupervised methods show up upstream — feature discovery, dimensionality reduction, regime detection — feeding into a supervised model rather than replacing it. Reinforcement learning is reserved for genuinely sequential decision problems, where "the correct single action" at each moment isn't even well-defined, only the cumulative consequence of a sequence of actions is.