The Library

Concept family

Machine Learning

Learned models on market data — kernels, clustering, neural networks, and ensembles.

Machine learning enters technical analysis as an estimation toolkit rather than a black box: methods that fit themselves to the data instead of relying on fixed formulas. On charts these tend to be compact, fitted on the fly to the available history. Nearest-neighbor lookups, kernel smoothers, and k-means state detection are the most common approaches in the Library, with neural networks and probabilistic models in smaller numbers.

The pieces fit together as a pipeline. Learned models map engineered inputs to forecasts; feature, label, and calibration choices decide whether those forecasts mean anything; regime detection tells other tools what kind of market they are in; similarity search treats history as memory; and voting blends imperfect signals into one. The caveat throughout: markets are nonstationary, so anything fitted to the past can degrade out of sample.

Start here

Learned models

The models themselves, from kernel smoothers and Gaussian processes to support vector machines, neural networks, and reinforcement learning, each fitted to market data to produce an estimate, a classification, or a decision.

Features & pipeline

Everything around the model itself: engineering inputs, reducing dimensions, defining labels and horizons, updating the fit as new bars arrive, and calibrating raw outputs into usable probabilities.

Feature Engineering3Logistic Signal Calibration3Online/incremental Learning2Label Definition & Prediction Horizon1PCA1Feature SelectionModel OverfittingProbability Calibration CurvesTrain/validation Discipline

Regimes & clustering

Unsupervised methods that sort market conditions into a small set of states, such as trending, ranging, or volatile, so other tools can adapt to the current environment.

Similarity & memory

Forecasting by analogy: find the historical stretches most similar to the present and study what followed them.

kNN Analog Forecasting13Dynamic Time Warping SimilarityMatrix Profile

Ensembles

Several imperfect signals combined by vote or weight, on the reasoning that independent errors partly cancel.

Ensemble Voting of Signals3BaggingModel Stacking