Abstract
The main objective of the project was to cluster trading indicators using machine learning. After achieving this initial objective it expanded to analyze the shift of trading indicators within clusters with changes in the data in terms of time period, frequency to generate clusters. Daily prices of 63 stocks in Financial Select Sector SPDR Fund from 2006 to 2017 and minute data from 20017-4-06 to 2016-4-12 were considered for the analysis. 18 Technical Indicators were calculated for each of the 63 stocks and used to generate 3 clusters containing 6, 8 and 4 indicators. Similarly 3 clusters were generated using the minute data and they contained 7, 7, 4 indicators each with only 1 indicator shifting clusters. Analysis shows that clusters generated using both the daily data and minute data are weak and not substantial. However, the clusters remain the same over time and don’t change in spite of changes in frequency of the data. Thus it can be deduced that the clusters, however weak, remain the same for a group of underlying stocks which in this case belong to Financial Select Sector SPDR Fund. Both in the case of daily data and minute data, the maximum percentage returns are generated by cluster 3 which contains the indicators - KAMA, TWM, EMA and WMA.
Research Topics:
K-medoids, clustering, technical indicators, High Frequency
Researchers:
Research Group (2016 Fall):
Hani Suleiman, Master in Financial Engineering, Graduated in December 2017
Advisor
Dr. Khaldoun Khashanah
Main Results:
The returns calculated per cluster as the average percentage returns per buy sell cycle. Both in the case of daily data and minute data, the maximum percentage returns are generated by cluster 3 which contains the indicators - KAMA, TWM, EMA and WMA
Conclusions:
Open, High, Low, Close prices were downloaded for the 63 stocks from Financial Select Sector SPDR Fund and clusters were generated using 10 years daily data 4 days minute data. The clusters generated in both cases are fairly similar with a difference of just one indicator switching clusters. The clusters generated with daily data are - Cluster 1 (6): DM, AROON, MFI, RSI,CMO, Bollinger Bands; Cluster 2 (8): CCI, SMA, DEMA, TEMA, ROC, MACD, TRIX, PPO; Cluster 3 (4): KAMA, TWM, EMA, WMA; The clusters generated with minute data are - Cluster 1 (7): DM, AROON, MFI, RSI,CMO,PPO, Bollinger Bands; Cluster 2 (7): CCI, SMA, DEMA, TEMA, ROC, MACD, TRIX; Cluster 3 (4): KAMA, TWM, EMA, WMA. This leads to the conclusion that clusters remain same for the same set of underlying stocks irrespective of changes in time period or even changes in frequency from daily data to minute data. This idea can be taken further and tested with stocks in other industries, and also other type of assets, derivatives to confirm that clusters generated are native to a particular set of underlying assets. And both in the case of daily data and minute data, the maximum percentage returns are generated by cluster 3 which contains the indicators - KAMA, TWM, EMA and WMA