Crypto Portfolio Clustering with and without optimization techniques (elbow method, PCA).

This is an anlysis of crypto market data, specifically the price-change data, for several crypto tokens. Two different KMeans unsupervised learning approaches were used to analyze this data, and come up with two different cluster plots. The first approach uses the original data, with standard scaling, and the second approach uses Principal Component Analysis to first reduce the features and then generate a model. In both cases, the Elbow Method was used to choos a k value for the number of clusters.

See full analysis details in the notebook crypto_investments.ipynb

A single dataset was used to anlyze this space:

• crypto_market_data.csv for price-change data over various time-periods
  • contains data: coin_id | price_change_percentage_24h | price_change_percentage_7d | price_change_percentage_14d | price_change_percentage_30d | price_change_percentage_60d | price_change_percentage_200d | price_change_percentage_1y

• Two KMeans algorithm analyses will occur, one with original feature data, one with PCA optimized data.
• Original: will use price_change_percentage_24h and price_change_percentage_7d for the cluster plot.
• PCA analysis will reduce to 3 components n_components = 3
• Each analysis will use the Elbow Method to arrive at the k-value for the KMeans algorithm

Original Feature Data
Using the elbow method with the original feature data (no PCA), suggested a k value of 4, as the inertia change really flattened after k=4. The resulting clustering plot with standard scalar (no PCA), was not very tight, and showed spread in each of the two main clusters (0 and 3). Specifically, there were two semi overlapping clusters for most of the data points, but also two clusters with only single outliers.

PCA Data
Using the elbow method with the PCA approach also yielded a k of 4, although the elbow plot was more stark with PCA than with original feature data. The resulting cluster plot with PCA was more distinct than that using the original feature data as well. The PCA scatter plot main two clusters (1,2) were better clustered although also had a little overlap. There were also two clusters which only had single outlier points, which is similar to the non-pca approach.

Overall, using the PCA approach, reduces the number of features down to a smaller set, which yielded results that were better as well as easier to work with.

Visual Comparison Summary

Below are screen caps showing the comparison for the Elbow Methods and resulting cluster plots when using the original standard-scaled vs PCA optimizations.

See full analysis implementation, interactive charts, and maps in the notebook crypto_investments.ipynb

This challenge uses python 3.7 and the following built-in modules:

• os

Additionally, it requires:

• matplotlib
• pandas
• JupyterLab
• hvplot
• scikit-learn

See installation below for specifics.

You will need Python 3.7, that supports for this application to run. An easy way to install python 3.7 is to download and install Anaconda. After installing anaconda, open a terminal/command-prompt, and setup a python 3.7 environment, and then activate it like so:

# create an anaconda python 3.7 environment
# name can be any friendly name to refer to your environment, eg 'dev'
conda create --name dev python=3.7 anaconda

# activating the environment
conda activate dev

# use pip to install the above modules, eg:
pip install python-dotenv
...etc...

The analysis is presented within a JupyterLab notebook. To launch JupyterLab, from the root of this repo dirctory:

# within repo root 
$ jupyter lab

You can now open the notebook crypto_investments.ipynb

David Lopez

MIT