Clustering Eurodollar Futures' Trading Days by Volatility

Clustering Trading Days by Volume and Range in the Eurodollar Futures Market from 2010-2015

• Use unsupervised learning methods to create a categorical feature based on cluster assignment. Cluster assignment is determined by the relationship between daily volume and range features for the front 6 quarterly contracts (closest to expiration) in the Eurodollar Futures market.
• After completion, a further study will be conducted using supervised learning methods to predict the cluster assignment for the upcoming trading day.

Read and Clean Data

• After downloading historical Eurodollar Futures market data from Quandl, I assembled 6 databases of historical end of day data for the “front” 6 contracts. These are the closest 6 quarterly contracts from expiration. For example if the current expiration is in December 2015, the contracts will be as followed:
  • 1 - December 2015
  • 2 - March 2016
  • 3 - June 2016
  • 4 - September 2016
  • 5 - December 2016
  • 6 - March 2017
• Contracts roll after expiration, ie, March 2016 will become Contract 1 in the above example and June 2017 will enter as Contract 6.
• After assembling the databases for the front 6 contracts I added each contract’s daily range and volume as features to be used for the clustering algorithm.

Use Unsupervised Methods to Create Trading Day Clusters

Examine Cluster Strength

• Unsupervised Learning shows the ideal cluster amount to be 2. However, we would like a little bit more separation, so we manually forced clusters to be 5.

## [1] 2

Cluster Stability

## [1] 0.5203835 0.5725811 0.5270334 0.5060634 0.7558619

## [1] 39 41 46 49  1

• With 5 clusters, the clusters are faily unstable; as clusters 1-4 are dissolved between 40-60x out of 100 runs. Only cluster 5 is considered stable.

Clusters

Number of Days per Cluster

## High.Vol-High.Range  High.Vol-Med.Range   Low.Vol-Low.Range 
##                  52                 164                 229 
##   Med.Vol-Low.Range   Med.Vol-Med.Range 
##                 512                 503

Visualize Clusters

Range and Volume by Year

• We see that 2012-2013 was dense with Low Volume - Low Range days. 2014 began the transition out of this stagnant period, and 2015 continues the trend toward increased volatility.

Cluster Breakdown per Year

• Once again we see that the majority of low volume - low range days occured in 2012-2013, with 2015 resembling the 2010-2011 period.

Year Breakdown per Cluster

• The increase in volatility is further demonstrated as over half the high volume - high range days occur in 2014-2015.

Visualize Clusters by Economic Events

• As expected, Nonfarm days are much more likely to be high volume - high range and high volume - medium range days. It is interesting to see that GDP and ADP days are relatively slow.

Summary

It is apparent that volatility is reemerging in the eurodollar futures market after a calm period in 2012-2013. This fits with the belief in an increased chance of an interest rate hike from the Fed. The current 2015 market resembles 2010 more than any other market in the past 6 years. The cluster analysis provides us with a classification variable that we can predict for upcoming trading days to help traders anticipate volatility.