Correlation Clusters [LuxAlgo]

15 ago

The Correlation Clusters is a machine learning tool that allows traders to group sets of tickers with a similar correlation coefficient to a user-set reference ticker.

The tool calculates the correlation coefficients between 10 user-set tickers and a user-set reference ticker, with the possibility of forming up to 10 clusters.

🔶 USAGE



Applying clustering methods to correlation analysis allows traders to quickly identify which set of tickers are correlated with a reference ticker, rather than having to look at them one by one or using a more tedious approach such as correlation matrices.

Tickers belonging to a cluster may also be more likely to have a higher mutual correlation. The image above shows the detailed parts of the Correlation Clusters tool.

The correlation coefficient between two assets allows traders to see how these assets behave in relation to each other. It can take values between +1.0 and -1.0 with the following meaning

• Value near +1.0: Both assets behave in a similar way, moving up or down at the same time
  
• Value close to 0.0: No correlation, both assets behave independently
  
• Value near -1.0: Both assets have opposite behavior when one moves up the other moves down, and vice versa
  

There is a wide range of trading strategies that make use of correlation coefficients between assets, some examples are:

• Pair Trading: Traders may wish to take advantage of divergences in the price movements of highly positively correlated assets; even highly positively correlated assets do not always move in the same direction; when assets with a correlation close to +1.0 diverge in their behavior, traders may see this as an opportunity to buy one and sell the other in the expectation that the assets will return to the likely same price behavior.
  
• Sector rotation: Traders may want to favor some sectors that are expected to perform in the next cycle, tracking the correlation between different sectors and between the sector and the overall market.
  
• Diversification: Traders can aim to have a diversified portfolio of uncorrelated assets. From a risk management perspective, it is useful to know the correlation between the assets in your portfolio, if you hold equal positions in positively correlated assets, your risk is tilted in the same direction, so if the assets move against you, your risk is doubled. You can avoid this increased risk by choosing uncorrelated assets so that they move independently.
  
• Hedging: Traders may want to hedge positions with correlated assets, from a hedging perspective, if you are long an asset, you can hedge going long a negatively correlated asset or going short a positively correlated asset.
  

Grouping different assets with similar behavior can be very helpful to traders to avoid over-exposure to those assets, traders may have multiple long positions on different assets as a way of minimizing overall risk when in reality if those assets are part of the same cluster traders are maximizing their risk by taking positions on assets with the same behavior.

As a rule of thumb, a trader can minimize risk via diversification by taking positions on assets with no correlations, the proposed tool can effectively show a set of uncorrelated candidates from the reference ticker if one or more clusters centroids are located near 0.

🔶 DETAILS

K-means clustering is a popular machine-learning algorithm that finds observations in a data set that are similar to each other and places them in a group.

The process starts by randomly assigning each data point to an initial group and calculating the centroid for each. A centroid is the center of the group. K-means clustering forms the groups in such a way that the variances between the data points and the centroid of the cluster are minimized.

It's an unsupervised method because it starts without labels and then forms and labels groups itself.

🔹 Execution Window



In the image above we can see how different execution windows provide different correlation coefficients, informing traders of the different behavior of the same assets over different time periods.

Users can filter the data used to calculate correlations by number of bars, by time, or not at all, using all available data. For example, if the chart timeframe is 15m, traders may want to know how different assets behave over the last 7 days (one week), or for an hourly chart set an execution window of one month, or one year for a daily chart. The default setting is to use data from the last 50 bars.

🔹 Clusters



On this graph, we can see different clusters for the same data. The clusters are identified by different colors and the dotted lines show the centroids of each cluster.

Traders can select up to 10 clusters, however, do note that selecting 10 clusters can lead to only 4 or 5 returned clusters, this is caused by the machine learning algorithm not detecting any more data points deviating from already detected clusters.

Traders can fine-tune the algorithm by changing the 'Cluster Threshold' and 'Max Iterations' settings, but if you are not familiar with them we advise you not to change these settings, the defaults can work fine for the application of this tool.

🔹 Correlations



Different correlations mean different behaviors respecting the same asset, as we can see in the chart above.

All correlations are found against the same asset, traders can use the chart ticker or manually set one of their choices from the settings panel. Then they can select the 10 tickers to be used to find the correlation coefficients, which can be useful to analyze how different types of assets behave against the same asset.

🔶 SETTINGS

• Execution Window Mode: Choose how the tool collects data, filter data by number of bars, time, or no filtering at all, using all available data.
  
• Execute on Last X Bars: Number of bars for data collection when the 'Bars' execution window mode is active.
  
• Execute on Last: Time window for data collection when the `Time` execution window mode is active. These are full periods, so `Day` means the last 24 hours, `Week` means the last 7 days, and so on.
  

🔹 Clusters

• Number of Clusters: Number of clusters to detect up to 10. Only clusters with data points are displayed.
  
• Cluster Threshold: Number used to compare a new centroid within the same cluster. The lower the number, the more accurate the centroid will be.
  
• Max Iterations: Maximum number of calculations to detect a cluster. A high value may lead to a timeout runtime error (loop takes too long).
  

🔹 Ticker of Reference

• Use Chart Ticker as Reference: Enable/disable the use of the current chart ticker to get the correlation against all other tickers selected by the user.
  
• Custom Ticker: Custom ticker to get the correlation against all the other tickers selected by the user.
  

🔹 Correlation Tickers

Select the 10 tickers for which you wish to obtain the correlation against the reference ticker.

🔹 Style

• Text Size: Select the size of the text to be displayed.
  
• Display Size: Select the size of the correlation chart to be displayed, up to 500 bars.
  
• Box Height: Select the height of the boxes to be displayed. A high height will cause overlapping if the boxes are close together.
  
• Clusters Colors: Choose a custom colour for each cluster.