GMM and Soft Clustering: A Beginner's Guide

Learn what a Gaussian Mixture Model is, how it gives each stock a probability of belonging to a group, and why that is a fit estimate, not a forecast.

Category: Machine Learning

In plain English

Imagine sorting a bag of marbles into groups by colour. "Hard" sorting puts each marble in exactly one pile. "Soft" sorting is gentler: a blue-green marble might be 70% blue and 30% green, so it partly belongs to two piles at once. A Gaussian Mixture Model, or GMM, is a method that sorts things this soft way. SurgeFlow uses it to ask which stocks look similar today, giving each stock a chance of belonging to each group instead of forcing it into one box.

What it measures

A GMM assumes the data is made of a few overlapping "blobs" (each a bell-shaped cloud, called a Gaussian). For every stock it produces a posterior probability: how well that stock fits each blob, with the numbers adding up to 100%. A higher probability for one group means the stock sits closer to the centre of that group's cloud. This is a measure of FIT, not of any future price move. - Group A cloud = "high-volatility" stocks; Group B cloud = "steady" stocks. - A stock might come out 80% Group A, 20% Group B. - Another stock split 55%/45% sits on the fence between the two. - The 80% is confidence in the grouping, never an 80% chance of going up.

How SurgeFlow uses it

GMM is one of three clustering methods behind the ML "market map" tab (at /ml), which gives a daily, descriptive read of which stocks look alike. SurgeFlow runs KMeans, GMM, and a third method, then combines them into a single agreed-on group for each stock. The GMM's soft probability (its posterior) is blended into a cluster confidence score: roughly half from how much the methods agree and half from the GMM posterior. So a high confidence means the model is sure where a stock belongs, not that the stock is a good buy.

How to read it on the site

Open the ML tab (/ml) and pick a market. You will see cluster cards (each a group with a plain-language name like "momentum" or "deep-value") and a ticker table. Each stock shows a confidence value: higher means it sits firmly inside its group, lower means it straddles two groups and could shift between daily runs. Click a cluster card to filter the table to just that group's members. - High confidence: stock is a clear, central member of its group. - Low confidence: borderline; expect it may flip groups over time. - Cluster names are descriptions of shared traits, not recommendations.

What it cannot tell you

The most important caveat: GMM probabilities are the model's estimate of how well a stock FITS a statistical blob today, not the probability of any investment outcome. A 90% confidence does not mean a 90% chance of profit, or any chance of profit at all. Clustering describes structure; it does not predict returns and gives no buy or sell signal. The groups also depend on the model's assumptions, the features chosen, and the day's data, so they can change as conditions change. - Not a forecast: high confidence ≠ likely to rise. - Assumes data forms neat bell-shaped clouds, which is only an approximation. - Group names are shorthand for shared traits, not quality judgments. - SurgeFlow is research and education, never investment advice.