Mining for Proitiable Patterns in the Stock Market

Mining for Rules

 

 The rules we mine for are similar to these by Liu (1998), Siberschatz & Tuzhilin (1996) and Zaki, parthasatathy, Ogihare, & Li (1997). They have the following format:

Rule type (1): a 3-day K-Line pattern          the stock’s price rises 10% in 10days.

Rule types (2) : a 3-day K-Line pattern         the stock’s price falls 10% in 10 days

 

   The search algorithm for the finding 3-day K-Line pat-terns that lead to stock price rise or fall is as following:

 

1.     For every 3-day K-Line pattern in the database

2.     Encode it by using the RPM method to get every day’s bit representation, cl, c2, c3;

3.     Increase pattern_occurrence[c1][c2][c3] by 1;

4.     Base_price = the 3^rd day’s closing price;

5.       If the stock’s price rises 10% or more, as compared to the base_price, in 10 days after the occurrence this pattern

 

Increase p  up[c1][c2][c3] by 1;

6.     If the stock’s price falls 10% cr more, as com-pared to the base_price, in 10 days after the occurrence of this pattern

 

Increase P  down[c1][c2][c3]  by 1;

 

   We used daily trading data form January 1, 1994, through December 31, 1998, of the 82 stocks, as shown in Table  as the base data set to mine for the price up and down patterns. After applying the above search algorithm on the base data set, the P up and P down arrays contained the counts of all the patterns that led price to rise or fall by 10% in 10days. In total, the up=patterns occurred 1,377 times, among which there were 870 different types of up-pat-terns; and the down-patterns occurred 1,001 times, among which there were 698 different types of down-patterns.

 

  A heuristic, stated below, was applied to all found

 

Table 1. 82 selected stocks

  

ADBE	BA	CDN	F	KO	MWY	S	WAG
ADSK	BAANF	CEA	FON	LGTO	NETG	SAPE	WCOM
ADVS	BEAS	CHKP	GATE	LU	NKE	SCOC	WMT
AGE	BEL	CLGY	GE	MACR	NOVL	SNPS	XOM
AIT	BTY	CNET	GM	MERQ	ORCL	SUNPS	YHOO
AMZN	BVEW	CSCO	HYSL	MO	PRGN	SUMW
AOL	CA	DD	IBM	MOB	PSDI	SYBS
ARDT	CAL	DELL	IDXC	MOT	PSFT	SYBN
AVNT	CBS	DIS	IFMX	MRK	RATL	T
AVTC	CBTSY	EIDSY	INTU	MSFT	RMDY	TSFW
AWRE	CCRD	ERTS	ITWO	MUSE	RNWK	TWX

 

Patterns to reduce the ambiguity of the patterns. Using the price-up pattern as an example, for a pattern to be labelled as a prise-up pattern, we think the time it appeared in P up should be at least twice as many as the time it appeared in the Pdown. Within all the patterns labelled as price-up patterns, they were sorted based on the ratio of the squared root of its told occurrence plus its occurrence as a price-up pattern over the occurrence as a price-down pattern.

       For price-up pattern:

Preference =    * Pup, if

* Pup,

 

For price-down pattern:

Preference =    * Pup, if

* Pup,

 

The find winning patterns with positive preference score are listed in Table 2.

Performance Evaluation

 

To evaluate the performance of the found winning pat-terns listed in Table 2, we applied them to the prices of the same 58 stocks for the period from January 1,199,

 

Table 2. Find winning patterns sorted by preference

Pattern Code	PO	Pup	Pdown	Preference
Up[00][20][91]	46	15	4	81.68
Up[01][28][68]	17	7	1	77.86
Up[07][08][88]	11	7	1	72.22
Up[00][24][88]	10	7	1	71.14
Up[00][30][8E]	9	7	1	70.00
Up[01][19][50]	28	12	3	69.17
Up[00][30][90]	39	21	9	63.57
Up[00][31][81]	26	8	2	52.40
Up[00][20][51]	18	8	2	48.97
Up[01][19][60]	24	9	3	41.70
Up[01][1D][71]	10	0	6	66.00
Up[00][11][71]	17	1	6	60.74.
Up[01][19][79]	35	3	10	53.05
Up[00][20][67]	18	2	5	23.11