Algorithmic Identification of
Chart Patterns
step-by-step identification procedures
which can be coded in a computer
program
Algorithmic Identification of
Chart Patterns
Key Benefits
• Instant scan innumerous charts
• Objective statistical analyses
• Powerful research means on combinations of
chart patterns with other analysis methods
Algorithmic Identification of
Chart Patterns
Major Obstacles
• Pattern Variation
endless ways for a textbook pattern to manifest itself
in a real chart.
• Scale
Almost all chart patterns of classic technical analysis
are scale-free
Outline of the Talk
• Brief review of methods used in academic
literature (finance and computer science)
• Ideas I have used in my articles
• A detailed example
Academic Approaches (Finance)
Three major identification methods in
academic financial literature.
• Smoothing price data
• Zigzag-ing
• Matrix template
Academic Approaches (Finance)
1st Method (Smoothing Price)
• smooth price data in a Max2
time span and identify
local extrema
Max1 Max3
• Define the pattern via
conditions for these
extrema
• Repeat the procedure
using different time Min2
spans. Min1
Academic Approaches (Finance)
2nd Method (“Zigzag-ing”)
• Apply a “Zigzag” swing Peak2
indicator to filter noise
(disregard changes below x% cutoff
and
threshold) identify
Peak3
peaks and troughs Peak1
• Define the pattern via
conditions for the peaks
and troughs
• Repeat using different Trough2
Trough1
cutoff threshold x%
Academic Approaches (Finance)
3rd Method (Matrix Templates)
-1 -1 -1 -1 -1 +2 +2 -1 -1 -1 -1 -1
-1 -1 -1 -1 -1 +2 +2 -1 -1 -1 -1 -1
-1 -1 +1 0 -1 +2 +2 -1 0 +1 -1 -1
A matrix template -1 -1 +2 +1 -1 +2 +2 -1 +1 +2 -1 -1
models the desired -1 +1 +2 +2 0 +2 +2 0 +2 +2 +1 -1
0 +2 +2 +2 +2 +2 +2 +2 +2 +2 +2 0
pattern using weights +1 +2 +2 +2 +2 +2 +2 +2 +2 +2 +2 +1
+2 +2 +2 +2 +2 0 0 +2 +2 +2 +2 +2
+2 +2 +1 +2 +1 -1 -1 +1 +2 +1 +2 +2
+2 +2 0 +1 -1 -1 -1 -1 +1 0 +2 +2
+2 0 -1 -1 -1 -1 -1 -1 -1 -1 0 +2
+2 0 -1 -1 -1 -1 -1 -1 -1 -1 0 +2
An iconic 12x12 matrix template for H&S
Academic Approaches (Finance)
3rd Method (Matrix Templates)
Superimpose a grid
(having the same dimensions as
the matrix template) in the
actual price chart
Academic Approaches (Finance)
3rd Method (Matrix Templates)
x x
x x
x x
The price action is x x x
x x x x x x x x
then converted into
x x x x x x x x x
matrix form x x x x x x x x x
x x x x x x x x x x x
x x x x x x x x x
x x x x x x
x x x
x
Academic Approaches (Finance)
3rd Method (Matrix Templates)
+2
x +2
x
+2
x +2
x
Sum up all weights for +2
x +2
x
the price action +2
x +2
x +2
x
-1
x +1
x +2
x +2
x +2
x 0x +2
x +1
x
x +2
0 x +2
x +2
x +2
x +2
x +2
x +2
x 0x
The higher the sum, the +1
x +2
x +2
x +2
x +2
x +2
x +2
x +2
x +1
x
+2
x +2
x +2
x +2
x +2
x 0x +2
x +2
x +2
x +2
x +2
x
better the price action fits
+2
x +1
x +2
x +1
x +1
x +2
x +1
x +2
x +2
x
the pattern model +2
x x +1
0 x -1
x +1
x +2
x
+2
x -1
x -1
x
+2
x
Academic Approaches
(Computer Science)
• Perceptually Important Points (PIP)
A smart modification of the zigzag-ing method: no use of % thresholds;
you a-priori define the number of swings you want instead.
• Symbolic Aggregate Approximation (SAX)
Uses alphabetic symbolic representation of segmented data series and
applies methods from bioinformatics and text data mining
• Support Vector Machines (SVM)
Sophisticated classification method which makes use of learning
algorithms in high-dimensional feature spaces
My Published Algorithms on Chart
Pattern Identification
Focus on:
Perception of Technical Analysts in mind
Elimination of Variation and Scale obstacles
Implementation in almost any technical analysis
software (Simplicity)
Execution Speed
To accomplish the above:
I created different algorithms per case (pattern)
Simple ideas I have Used (I)
• Chart patterns usually consist of
branches
• These branches have some dimensional
restrictions with respect to one another
• If we identify even just one of these
branches, then we can usually identify the
whole pattern
Example:
Identifying the top of the left shoulder in H&S
Area of
possible tops
for the left
shoulder
The dimensions of the green area are related
to the dimensions of the orange one
Simple ideas I have Used (II)
When a branch of a chart pattern is
more or less curved, you can
usually identify its important points
(no matter the scale)
Example:
Identifying the top of the head in H&S
The highest Check all prices
between P and Q.
of them is
the top of
the head
P
Q
Move backwards in time until
you find a price (Q) which is
lower than the last one (P)
A Detailed Example
Identifying the cup pattern
A modification of a simple algorithm presented in my February 2006
article “Identifying The Cup” in the Technical Analysis of Stocks and
Commodities magazine
What is a Cup pattern?
Rounding Bottom Rounding "Bottom" Rounding "Bottom"
after a downtrend after an uptrend after a sideways market
All these different types of rounding
"bottoms" are called cups and they
are generally considered bullish
patterns
Identifying the Cup
High of the last Bar
C
Move backwards in time until Consider the bars
you find a bar having a higher below the red line
Let C be the lowest
high than A.
low of all red bars
Identifying the Cup
“Imprison” the red bars behind a 5x5 grid of equal
rectangular cells using A and C as milestones
Identifying the Cup
A
1st Condition:
No Closing
price inside the
yellow cells
AND
No High price
inside the
C
orange cells
Identifying the Cup
A
2nd Condition:
Bar lows exist
inside both
green boxes
OR
bar lows exist C
inside both
blue boxes Height of blue and green boxes = 2/3 of the height of cells
Identifying the Cup
A
3rd Condition:
There are at
least 20 red
bars
C
Application
Application of the algorithm in the daily
charts of all S&P500 stocks from 1982 to
2014 reveals a total of 3,991 distinctive*
cups of various durations.
Scan took less than 20 seconds
distinctive*
No time overlap of more than 70% between any two
cups in the same chart
Algorithm in Action [Example 1]
Apple (Daily)
49 bars
49 bars 143 bars
Algorithm in Action [Example 2]
Big Lots Inc (Daily)
30 bars
95 bars
Algorithm in Action [Example 3]
C.H. Robinson WW (Daily)
22 bars
71 bars
Statistics for the S&P500 Stocks
(Daily Charts 1982-2014)
• Total number of cups: 3,991 • Median duration: 32 bars
• Minimum duration: 20 bars
• 80% of all cups were less than
• Maximum duration: 1,162 bars 3-months long (75 trading days)
Evaluation of Chart Patterns
Algorithmic identification gives us the means
to evaluate the effectiveness and efficiency of
a chart pattern
Because: it gives us detailed info about innumerous
manifestations of the pattern in actual charts
Example: Evaluating of Cup
Questions:
Does the cup pattern actually produce uptrends?
Are these trends exploitable?
What is the historical edge of the cup?
How confident we are that this edge is statistically
significant?
Example: Evaluating the Cup
A simple buy-only system was applied in the previous data
• Go long at the high of the identification bar when a cup is
identified
• Exit when the price falls below a % trailing stop-loss of 0.7
times the % distance from A to C
Example A
Go long
-21% here with a
-30%
trailing stop
loss of 21%
C
Example: Evaluating the Cup
A Profit Factor (PF) was calculated for every
hypothetical trade :
So, all cups are treated under equal terms (no matter
their height) allowing us to determine the "edge" of cup
pattern from this system's point of view
Example: Evaluating the Cup
Results for S&P500 stocks from 1982 to 2014:
Statistical Values Translation
• Mean PF = 0.166 Edge of cup (1982-2014) = 16.6%
• Profitable trades (%) = 39% Most trades produced loses
• CV = 1,105% PF’s vary too much in relation to the
edge of the system
• 99% confidence levels for We can be 99% sure that the
Mean PF : 9.1% - 24.0% [9% , 24%] interval captures the all-
time edge of the cup pattern
Epilogue
Algorithmic identification of classic TA
patterns is feasible and provides important
benefits for the chartists
Thank you
