Agglomerative Clustering

[secsy]

Been working on a model for currencies and wanted to know if anyone has tried something like this before. Given a history of prices, I create scaled instances of a certain length (around 60 points per instance), then cluster them by cosine similarity. The idea of clustering is simply to reduce the computation time required to compare arrays to find similar instances. The strategy relies on the assumption that "similar" price instances will reliably produce similar outcomes. The clustering, which I am doing with agglomerative clustering and then fitting with a classifier, speeds up the process significantly (~5x compared to array comparison for all instances). Would be interested to know if anyone has thoughts on this/experience doing this.

[contango_and_cash]

I have not done exactly what you are saying, but if the clustering itself is somewhat consistent, you could just pickle your fitted model with some lag t (1h, 1day), load the latest model at t+1 and transform on the fly. this alone will greatly reduce computation time regardless.



agglomerative clustering works with the added benefit of dendrograms to visualize (often a struggle with ML stuff) but i don't know enough about your problem.

[ctd]

It probably depends on what your ultimate objective is as to whether this approach will be fruitful.



I've used clustering methods, but to a certain extent with the aim of dimensionality reduction (a la PCA), where it can shine given appropriate context. It stands to reason it could work for you if you trusted the underlying assumption you mention and the clusters themselves (ie, it will take less time to classify 5 clusters than to classify 40 separate currency crosses).