Using Clustering for Trading Signals and Neural Network Features
Summary
The article explains two ways to use clustering results in a trading model. As a standalone method, a trained clustering model assigns historical market states to clusters, and labeled examples are used to estimate the frequency of buy, sell, or neutral outcomes after each cluster appears. It stresses that the clustering model is not retrained during this step, and suppresses probabilities for clusters with fewer than ten observations to avoid treating tiny samples as reliable.
As model inputs, cluster identifiers alone add little information beyond those statistics. The article instead proposes using distances to cluster centers, transformed so nearby clusters receive greater weight, then feeding the features to another model. The examples concern identifying a developing fractal pattern before it is fully formed. These are implementation approaches rather than evidence of profitable performance; the text emphasizes the need for sufficient, representative labeled data and compatible clustering and labeling periods.
Key ideas
- Cluster assignments can support outcome statistics for buy, sell, and neutral signals without retraining the clustering model.
- The method sets signal probabilities to zero when a cluster has fewer than ten labeled occurrences.
- Cluster identifiers alone may provide little beyond direct cluster-level statistics.
- Distances to cluster centers can be transformed to emphasize proximity before use as neural network inputs.
- The examples target probabilistic early identification of fractal patterns and do not establish profitability.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.