Unsupervised Methods for Discovering Patterns in Financial Time Series

Summary

The document asks which unsupervised learning methods can help discover patterns in manually selected financial time-series features, with the aim of making broad predictions about future price behavior. It gives examples ranging from predicting the next candle's color to whether the next day will reach a higher high. The answer offers a brief list of candidate methods: k-means clustering, principal component analysis, hidden Markov models, autoencoders, and outlier detection, with Chauvenet's criterion as an example.

These methods serve different purposes. Clustering can group similar observations, PCA can reduce feature dimensions, hidden Markov models can represent changing latent states, autoencoders can learn compressed representations, and outlier detection can flag unusual observations. The response is only an introductory inventory; it gives no implementation guidance, dataset, comparison, or evidence that any method predicts returns. Results would depend on feature and target choices, and discovering structure does not by itself establish predictive value.

Key ideas

• K-means can group observations with similar feature values.
• PCA can reduce the dimensionality of manually selected features.
• Hidden Markov models can represent time series as sequences of latent states.
• Autoencoders and outlier detection offer additional ways to identify structure or unusual observations.
• Unsupervised pattern discovery alone does not show that a pattern can predict future prices.

Tags

# unsupervised pattern discovery - methods?


# unsupervised pattern discovery - methods?












Given that I select features manually, what methods are available for pattern discovery with the purpose of time series prediction (footnote)?

I only stumbled upon hierarchical clustring ("bottom-up") and proprietary sofftware. This post says there are a lot such methods.

Footnote: prediction in wider sense is meant. The outcome to predict is chosen manually, too, from the trivial "color of the next candle stick" via "wiskers will be longer than the body" to "next day will set a higher high", just anything, you name it

## Answer by madilyn (score 2)

https://quant.stackexchange.com/a/37759

Here's a few:

- K-means

- PCA

- HMM (learned with expectation-maximization or Viterbi algorithm)

- Autoencoders

- Outlier detection, e.g. Chauvenet's criterion

Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.