StockSpotter.

com
SPECTRAL DILATION John Ehlers
[email protected]

Codifying the Fractal Nature

of Market Data

Impact on Technical Indicators

StockSpotter.com
OUTLINE John Ehlers
[email protected]

• Theoretical Basis of Market Data Structure

• Measured Market Data Structure
• Measuring Market Data Spectrums
• The Need to Think in Terms of Frequency
– Frequency is the Dual of Conventional Time Waveforms
• Filter Basics
• Indicator Dynamics
– The Impact of Spectral Dilation and What to Do About it
• An Introduction to www.StockSpotter.com
StockSpotter.com
Drunkards Walk John Ehlers
[email protected]

• Described in “MESA and Trading Market Cycles”

– Drunk steps right or left with each step forward
• Random Variable is position
• Results in the famous Diffusion Equation
• Describes the shape of a plume of smoke (or a trend)
– Drunk steps in the same or opposite direction as the
last step with each step forward
• Random Variable is momentum
• Results in the famous Wave Equation
• Describes a meandering river (or a cycle)
• The 2nd Order Partial Differential Equations are
nearly identical
• Results are that cycles and trends can coexist in
a complex mixture
StockSpotter.com
Swerling Model John Ehlers
[email protected]

• Peter Swerling statistically described radar echoes

– Pulses were noisy over time – due to complex airplane shapes
and changes in aspect from the fixed radar site.
– Model described as pure noise with memory
• I have synthesized market data as noise with an EMA
– Not bad for a simple model
StockSpotter.com
Hurst Exponent John Ehlers
[email protected]

• The Hurst Exponent describes the randomness

of a data series

“Hurst Exponent and Financial Market Predictability”

By Bo Qian and Khaled Rasheed
University of Georgia
 Measured & Modeled John Ehlers
[email protected]
StockSpotter.com
Market Spectrum

• Market spectrum amplitude

models as 1/F
• 2*(1 – Hurst Exponent)=  Quantization
• Spectral Dilation increases Noise

approximately 6 dB/Octave
• 1/F Noise is apparently
universal Spectral
Dilation
• Model shows two mandates
for Technical Analysis
1) We must stay several octaves
away from the Nyquist
Frequency due to Quantization
Noise
2) Indicators must compensate for
spectral dilation to get an
accurate frequency response
“Modelling Share Volume Traded in Financial Markets”
By V. Gontis
Lithuanian Journal of Physics, 2001, 41, No. 4-6, 551-555
StockSpotter.com
Quantization Noise John Ehlers
[email protected]

• Highest possible frequency has two samples per cycle (Nyquist Frequency)
– 2 day period on daily bars
StockSpotter.com
OverSampling John Ehlers
[email protected]

• Why not reduce quantization noise by

sampling more often?
– For example – hourly data to trade daily bars
• What is a day? 6 hours? 24 hours?
• Gap openings are a data issue
• Spectral dilation becomes an even larger
issue because several more octaves
range is included in the data
StockSpotter.com
Autocorrelation John Ehlers
[email protected]

• Correlates a waveform with itself lagged in time

• SwamiCharts Autocorrelation of a theoretical 20 Bar Sine Wave
• Vertical Scale also shows periodicity
 Autocorrelation John Ehlers
[email protected]
StockSpotter.com
Periodogram

• Autocorrelation Periodogram produces spectrum amplitudes that

are range bound by the correlation coefficient
• All other spectrum measurements must compensate for Spectral
Dilation for a true picture of the Measured Spectrum
• This is really how I “discovered” spectral dilation
StockSpotter.com
Filter Basics John Ehlers
[email protected]

I’m sorry! I just have to do this

Let Z-1 represent one bar of delay
4 Bar Simple Moving Average: Output = (1/4 + Z-1/4 + Z-2/4 + Z-3/4)*(Input Data)
Transfer Response = H(z) = Output / (Input Data)
More Generally: H(z) = b0 + b1*Z-1 + b2*Z-2 + b3*Z-3 + b4*Z-4 + ……….+ bN*Z-N

An EMA uses a previously calculated value, so with still more generality:

b0  b1 * Z 1  b2 * Z 2  b3 * Z 3  b4 * Z 4  .........  bN * Z  N
H ( z) 
a 0  a1 * Z 1  a 2 * Z  2  a3 * Z 3  a 4 * Z  4  ........  a N * Z  N

Therefore, filter transfer response is just a ratio of polynomials

The polynomials can be factored into their zeros
Zeros in the denominator are called poles

The rate of filter rolloff is 6 dB / Octave per Pole

Since we must use simple filters in trading we have only a few poles
in the transfer response - BUT – the data are increasing at the rate
of 6 dB / Octave. The result is there is no real filtering.
StockSpotter.com
SMOOTHING FILTERS John Ehlers
[email protected]

• The real reason to use averages or smoothing

filters is to remove quantization noise
Take Your Pick
10 Bar SMA and EMA 10 Bar SuperSmoother

EMA (1 pole)
SMA
2 double zeros
StockSpotter.com
SuperSmoother Filter Code [email protected]
John Ehlers

SuperSmoother Filter
© 2013 John F. Ehlers

a1 = expvalue(-1.414*3.14159 / 10);
b1 = 2*a1*Cosine(1.414*180 / 10);
c2 = b1;
c3 = -a1*a1;
c1 = 1 - c2 - c3;
Filt = c1*(Close + Close[1]) / 2 + c2*Filt[1] + c3*Filt[2];

Code Conversion Notes:

1) Filter is tuned to a 10 Bar Cycle (attenuates shorter cycle periods)
2) Arguments of Trig functions are in degrees
3) [N] means value of the variable “N” bars ago
StockSpotter.com
HighPass Filter John Ehlers
[email protected]

• HighPass Filters are “detrenders” because they attenuate low

frequency components

One pole HighPass and SuperSmoother does not produce a zero mean

Because low frequency spectral dilation components are “leaking” through

The one pole HighPass Filter response
StockSpotter.com
Roofing Filter John Ehlers
[email protected]

• Comprised of a two pole HighPass Filter and a SuperSmoother

The Roofing Filter guarantees only the desired

frequency components will be passed for analysis
StockSpotter.com
Roofing Filter Code John Ehlers
[email protected]

Roofing Filter
© 2013 John F. Ehlers

//Two Pole Highpass filter passes cyclic components whose periods are shorter than 48 bars
alpha1 = (Cosine(.707*360 / HPPeriod) + Sine (.707*360 / 48) - 1) / Cosine(.707*360 / 48);
HP = (1 - alpha1 / 2)*(1 - alpha1 / 2)*(Close - 2*Close[1] + Close[2]) + 2*(1 - alpha1)*HP[1] - (1 - alpha1)*(1 - alpha1)*HP[2];
//Smooth with a Super Smoother Filter
a1 = expvalue(-1.414*3.14159 / 10);
b1 = 2*a1*Cosine(1.414*180 / 10);
c2 = b1;
c3 = -a1*a1;
c1 = 1 - c2 - c3;
Filt = c1*(HP + HP[1]) / 2 + c2*Filt[1] + c3*Filt[2];

Code Modification Notes:

1) HP Filter is tuned to a 48 Bar Cycle (attenuates longer cycle periods)
2) SuperSmoother is tuned to a 10 Bar Cycle (attenuates shorter cycle periods)
3) Arguments of Trig functions are in degrees
4) [N] means value of the variable “N” bars ago
 Impact of Spectral Dilation [email protected]
John Ehlers
StockSpotter.com
On Traditional Indicators

• Spectral Dilation has impacted (distorted?)

the interpretation of virtually all indicators

Conventional Stochastic

Stochastic preceded by a Roofing Filter

 Roofing Filter Can Be John Ehlers
[email protected]
StockSpotter.com
An Indicator Itself

• Cycle Period is about twice the desired

trade duration

2 week nominal trade duration

3 month nominal trade duration

 John Ehlers
[email protected]
StockSpotter.com

Even better DSP indicators DO exist

StockSpotter.com
StockSpotter.com John Ehlers
[email protected]

• Analyzes over 5000 Stocks & ETFs each day

• Free indicator analysis
– Includes Advanced SwamiCharts
• Free and Premium Screeners
• Watchlists
• Swing Trading signals called – IN ADVANCE
– Performance is transparently tracked
• Monte Carlo Analysis of Performance
• Technical Presentations (this webinar is there)

This QR code will take

you to this presentation
at StockSpotter.com