This project aims at helping Portfolio Managers to optimize their quantitative strategies.
Recently, particularly with the advent of new computer tools, software optimization and saving time are very important factors in a highly competitive sector like Finance.
In this context, the improvement of technical indicators generating a buy or sell signal is a major challenge and many tools have been created to make them more effective. This concern for efficiency led us to explore the best (and most innovative way) that would allow us a further improvement of these indicators. The technique of frequent patterns seemed to be the most appropriate method to allow Portfolio Managers to optimize their quantitative strategies.
Our method is to assign a signature to frequent market configurations. PLC "trading" is finding the most accurate signatures using a "back-testing" procedure for use with technical indicators to improve their performances.
Indeed it must be determined that the signatures combined with the indicator outperform indicator alone. To find and extract the frequent patterns we use the algorithm "FP-Tree" which we found to be the most efficient algorithm to perform this task.
Indeed the FP-Tree algorithm allows reducing the number of scanning of the database. The tree structure will allow the algorithm to have all the information it needs, what will have a positive effect on both the efficiency of mining and the processing time.
The only inconvenient of this algorithm related is that the construction of the FP-tree structure can be long and may consume a lot of system resources. We also introduced some epsilon value that will play the role of gauge, the larger epsilon is, the better opportunity we have to find frequent patterns. However, the smaller epsilon is, the more accurate is the result.
A graphical interface will be available to the Portfolio Manager so that it can return to its simple excel file containing the data he wants to be treated and the percentage of the minimum support. Our software will give him j + 1, j +5 and j+10 yields and standard deviation.
After several back-testing over different periods, we can conclude two things, the first is that our tests have failed to create effective signatures as major economic alterations between 1929 and 2008 can not be anticipable and the second thing is that our results also depend on the value of epsilon. We did many back-testing to be able to find the right value of epsilon. Indeed, we selected a range of values ??of epsilon to do the back-test: from 0.01 to 1. For example, 0.01 means that we only take a piece of information out of 100.
To enhance our results, we thought to integrate an exponential law to our algorithm, what would give more weight to recent information. Years of improvement of technical analysis have made recent information more precise and more complete than old information, so with this exponential law, we could give more weight to the latest results.