Thursday, September 15, 2016

Stock Trend Analysis and Trading Strategy


He et. al. (2006) developed a highly empirical trend analysis methodology to assess and anticipate future stock trends and trading strategies.  To do this, they formulated a three step approach which combines data partitioning, linear regression, and prediction.  Once combined they recommend using the newly minted trend prediction methodology; Trading based on Trend Prediction (TTP).  The authors additionally devised this methodology as they observed previously that, “patterns in long time series data repeat themselves due to seasonality or other unknown underlying reasons… [and] this information will be able to help decision-making on the trading strategy in stock market trading practice.”


At the outset the scholars developed sophisticated calculus for data preparation as they needed an ability to train the data for the analysis. The empirical analysis of this step can viewed as a “window” in the training series to observe test data. See Figure 1 below for visual representation:

S1 : p1, p2, …, pwtr
S2 : p2, p3, …, pwtr +1
Sn : pN, pn+1, …, pwtr+N-1

During the “data mining” phase it takes place over three distinct steps; Initialisation, Data Mining, and Test models on test data.  During the Initialisation phase, a training time period is selected to observe the stock data (i.e. – 1999-2000) whereby the data is sampled from the first day of the year to the last day of the year. During the data mining phase, the now trained data is “partitioned” into clusters (k-clusters) to be repurposed later for linear regression modeling. Finally, in the last step, the data can be modeled and calculations can be made for the returns and trading strategies can be identified.

The authors additionally delineate two distinct trading strategies: Na├»ve Trading (NT) and Trading based on Trend Prediction (TTP). NT is the basic buy low, sell high proposition, while TTP is the same as NT, but incorporates forward looking trend analysis on future stock seasonality for better decision-making.  To that end, they only “sell the share if the trend prediction is downward.”

In testing the method, the scholars ran the data for the 1999-2000 test period against the corresponding training period 1989-1998. What they found was, “TTP’s performance exceed[ed] NT’s performance in most countries… [and] clearly indicates that the trend prediction is able to find the correct trend in some cases.” However, they also found that TTP was not generalizable in predicting trends in all cases.

Finally, they conclude that the results show the proposed methodology improves the trading performance over some existing strategies in some cases. Further, they found that the methodology can correctly predict future trends in stock price, but not predict them well in other situations. They also leave this study open for further research and suggest future ways to further improve the method.


Despite this studies overall findings in support of correctly predicting, in some instances, future stock prices -- this study lacks much needed qualitative explanations and support (i.e. – 4 pages in length). The study, although heavily empirical, could make clear changes to further improve the usability of the study by appealing to a wider audience who wishes to utilize sophisticated trend analysis as mentioned herein, but do not have the mathematics background to do so.



  1. Did the study explain which situations that the method is accurate in predicting future stock prices?

  2. Aubrey, good catch! The referenced and utilized stock price indexes were from five different countries in a Buy and Hold (B&H) strategy; USA, UK, Canada, Taiwan, Singapore. That said, TTP was found to be most effective at predicting future stock price for the US and Singapore. For the UK, NT was slightly better than TTP. Canada was an outlier as only Genetic Programming* (GP) was able to effectively predict B&H strategies. Also, TTP was able to exceed B&H over most of the practical strategies.

    *Genetic Programming - Is a technique whereby a computer program is encoded to a set of genes that are then modified (evolved) using an evolutionary algorithm.

  3. Tom, this was a well written summary and I am curious as to what future ways the authors thought that this method could be improved.

  4. Instead of paraphrasing how they want to improve this study, here are the four ways they wish to take this study forward from the text:

    1.) A simple decision on classification of clusters is made using the linear regression model in the present work. We can further improve the accuracy of the trend prediction by using fuzzy or probabilistic decision systems in the future.

    2.) Improve the computation efficiency by using sophisticated and scalable clustering techniques... [references in-work citations]

    3.) Introducing scale change to pattern matching can discover similar patterns with different time scales.

    4.) Combine our method with other techniques, such as GP, for better and more sophisticated trading strategies.

    Thanks for your comment on the quality of the narrative!