** QuanTek development:** Here is the latest news
regarding the ongoing development of the

Internally the * QuanTek* program has been
extensively modernized, but the user interface (UI) looks much the
same as the original version from the late '90s. This needs to be
modernized from the old

I would like to build upon the basic **Adaptive Wavelet Linear Prediction**
filter design by making use of a wider variety of regressors or
**technical indicators**, and do further testing and
tweaking of these filters. This should also set the stage for
a version of the filter to work with **real-time data**.

Also the **volatility** will be taken into account in the **
Adaptive Wavelet LP **filter by making use of a **GARCH** model (**Generalized
AutoRegressive Conditional Heteroskedasticity**). This should improve
the predictive power as well as setting the stage for an **Options**
version of ** QuanTek**.

More **Statistical Tests **are planned, such as **Simulated Trading**
scenarios using various **Adaptive LP filters** and **Trading Rules**.

Also the versions for ** MetaStock** and

A new **Adaptive Wavelet Linear Prediction**
filter has been developed that seems to provide good results. This
filter solves some issues with the previous filter design. We intend
to refine and develop this filter further in future updates. This
new **Wavelet Linear Prediction** filter is designed
for **non-stationary time series** such as financial returns series.
It should give a better estimate of the N-day returns, up to 128 days,
without "fitting to the stochastic noise" as in the standard
filters. This filter is also set up so that it can be generalized to
more complex types of filter. The next improvement will be an **adaptive**
version of this filter, and also the filter will be generalized to a
**multivariate** version which does the whole portfolio calculation
at once.

Also the **Optimal Portfolio** calculation was updated and
refined, going beyond the modified **Markowitz **method
that was previously used. I had to develop a new modified version of
the **Markowitz **method that was suitable for individual
portfolios (see the article in the **Help** file). This
corrects a fundamental issue with the standard **Markowitz **
method.

I have also introduced a new set of **Help**
dialogs. Eight of these are shown from the **Welcome to
QuanTek! **dialog, which is available from the Main Window,
and another eight are shown from the **Data analysis - Graphs
& Trading Rules**, which is available from the Graph Window.
These Help dialogs present a summary of key features of the **
QuanTek** program, and also provide links to the
main

This version downloads **free Yahoo data** as well as parsing
**ASCII** files. The historical data is parsed from **
Yahoo** *.csv files from their website, but the daily update
may be done automatically. The time for the daily update may be set,
and then the download is done automatically even while the computer
is asleep.

This has been a series of beta versions, in which I have
essentially rewritten and modernized the whole program. Many of the
numerical routines and data download routines were obsolete and
needed to be updated. The ASCII parse routines were also updated and
improved, in particular to implement the parsing of **free
Yahoo data**.

In addition, a set of **Wavelet** routines was
implemented, incorporating **Wavelet smoothing** in
various displays, as well as a new **Adaptive Wavelet Linear
Prediction** filter design. This design is more appropriate
for financial data than conventional designs, such as the **
Standard LP ** filter that we still use, because it
minimizes "fitting to the noise" and helps to separate the "signal"
from the "noise" in financial data.

A number of **Statistical Tests** have been devised and included
in a **Statistical Tests Dialog**. These were devised to test the**
Linear Prediction** filter as well as the** ** **Wavelet
routines** to make sure they were operating correctly.

I developed an **encrypted Authorization Key** for the program
to enable licensing for fixed subscription periods, which uses a personal
**Pass Code** for each subscriber.

** QuanTek** enables you to design your own custom

This version of ** QuanTek** underwent much testing and development of the

Also a **Correlation Test - Filters** dialog was developed in
order to test directly the predictive properties of the various **Linear
Prediction** filters. In this version, six different **LP**
filters were included, and any one of these filters can be selected
to use for computing the **Price Projection** and **Trading Rules**.
However, it was found necessary to be able to set the filter parameters
manually, for each security individually, in particular the **Fractal
Dimension**. This has been done in the subsequent version of
** QuanTek** by bringing back the

The **Time Horizon** adjustment could now be made separately for
each security, rather than one setting for the whole program.
This is important because each security has its own optimum **time
horizon** for best performance of the **Price Projection**.
Now this may be set either in the **Correlation Test - Filters **
dialog when testing the filters or in the **Correlation Test - Indicators**
dialog when testing the indicators.

A new dialog box, called the **Trading Rules Filter & Momentum
Weights** dialog, was added, where the **Trading Rules** can be
selected and displayed. The new selections for the **Trading
Rules** include three new controls, called the **Increment**, the
**Threshold**, and the **Compression**. In addition the
three **Momentum Weights**, which are the weights of the three **
Momentum** indicators in the **Trading Rules**, are set from this
dialog. This new dialog is also called from the **Diagnostic
Test**, so a set of **Trading Rules** based on the three saved
**Momentum** indicators can be set and back-tested independently
from the data file itself.

** QuanTek** now works with either

Here were some of the new features of ** QuanTek 3.0**:

**Price Projection:**A**Price Projection**is computed using one of a choice of**Linear Prediction**filters, and displayed on the Main Graph. This**Price Projection**is also used as part of the computation of the**Momentum**indicators and**Trading Rules**. One of the**LP**filters uses the**Wavelet spectrum**, one uses the**Periodogram spectrum**, and one computes the**autocorrelation**directly. At present, all of these filters assume a**stationary stochastic process**over the past**1024 days**, which seems a reasonable approximation to the true**non-stationary stochastic process**underlying the financial markets. You can experiment with the results of using the different**LP**filters. Linear Prediction: The**Hybrid LP Filter**dialog is used to set the**Linear Prediction**filter for each security separately. The**type**of filter may be chosen, as just described, and two parameters which are called the**Order of Approximation**and the**Fractal Dimension**can be set. The**Order of Approximation**selects the degree of smoothing of the filter spectrum, by setting the length of the series of**LP coefficients**. The**Fractal Dimension**sets the low-frequency response of the filter, corresponding to modeling the time series as a**fractionally differenced (FD)**orprocess. The*long-memory***Hybrid LP Filter**dialog also has attached to it a dialog box which displays the**correlation**between the filter output**future returns**. Using this**Correlation Test - Filters**dialog, the**predictive power**of each filter can be tested directly for any settings of the filter**type**and**parameters**.**Momentum Indicators:**The main feature ofis that you can design and test yourself a wide variety of*QuanTek***technical indicators**based on the**Linear Prediction**filter and the**Savitzky-Golay smoothing**filter. These consist of practically every conceivable oscillator-type indicator that can be made out of the past price data. These indicators can then be tested for**correlation**with**future returns**using the**statistical tests**in. After a set of three of these indicators is designed, the*QuanTek***Trading Rules**indicator is formed by a weighted sum of these three indicators, with various**filter rules**applied. The effectiveness of the final**Trading Rules**indicator can be tested in a variety of realistic trading scenarios by means of the**Diagnostic Test**.**Statistical Tests:**There are two**correlation**tests which measure**correlation**with**future returns**of the**technical indicators**and the output of the**Linear Prediction**filter. These give a direct indication of the effectiveness of these two methods of estimating**future returns**. There is also a**Diagnostic Test**, which is a back-testing routine to test the**Trading Rules**in a variety of realistic trading scenarios, for a given**time horizon**. Two of the**statistical tests**are the standard**Periodogram spectrum**and the**Wavelet spectrum**, which are two different methods for computing the power spectrum of the stock (log) price returns. These give another indication of the possible presence of**correlation**in the returns data.**Main Graph:**The**scrolling graph**has been improved, with**four different scales**that cover the whole data set. Now there is no limit to the time span of historical data that can be displayed. This also makes it possible to do a better statistical analysis, using deep historical data. The graphs are now available in either**black**or**white**background. The black background looks great! These four graph scales display the**Price Projection**, various**buy/sell signals**,**Bollinger Bands**, and on the highest scale is a**Candlestick Graph.****Portfolio Optimization:**also incorporates a*QuanTek***portfolio optimization**routine, making use of the**Markowitz method**. The**correlation**between the**returns**of all the securities in the portfolio (data files in a folder) is computed and used, along with the**expected return**for each security derived from the**Price Projection**, to compute the**optimal portfolio**which should**maximize returns**and**minimize risk**. The parameters for**risk aversion**and**margin leverage**can be set by the user, to achieve portfolios with different degrees of**risk vs. return**.

* StockEval* was a preliminary version of our stock trading
program, designed to incorporate up-to-date methods of computation and
a

* StockEval* used only data from

These were the main features of ** StockEval**:

**Price Projection:**The main**technical indicator**in theprogram was a**StockEval****Price Projection**, in which future prices are estimated up to 100 days in the future. Two of the main approaches we tried were a**regression**of the future returns on a set of exponential moving averages of the past prices, and the**Linear Prediction**routine. The first method used a set of**exponentially weighted moving averages**on different time scales, and the differences between**EWMA**s on adjacent time scales were taken as the basis functions (essentially**MACD**indicators). The future one-day returns were then regressed on these basis functions, and the whole procedure was extrapolated into the future one day at a time. This method apparently gave reasonable results. However, a simpler method is to utilize the standard**Linear Prediction**method, which is somewhat equivalent, although it leads to a "noisier" outcome due to a greater number of parameters being fitted. Both of these methods gave good results on the**Diagnostic Test**(a back-testing routine), but the excess noise in the trading rules on short time scales was a difficult problem. However, after applying the**Savitzky-Golay**digital smoothing filter to get rid of the short-term**stochastic noise**, the results were definitely positive, and led to rather spectacular short-term trading gains according to the**Diagnostic Test**, in the strongly trending market of the late '90s.**Linear Prediction:**The**Linear Prediction**filter used inwas a standard, publicly available**StockEval****LP**filter which models the time series as an**auto-regressive (AR)**process. This filter was used with the maximum number of parameters (1024) in order to try to get the best possible long-term prediction. (The short-term**stochastic noise**was then filtered out using the**Savitzky-Golay smoothing**filter.) However, after analyzing this particular filter, it turns out that it is actually*adaptive*in a certain sense (when used with the full set of parameters), in that it gives greater weight to the most recent data. This explains why we were getting pretty good results on the short-term trading rules as well, as measured in the**Diagnostic Test**.**Harmonic Oscillator:**Theprogram featured a set of three novel**StockEval****technical indicators**based on the**Savitzky-Golay**digital smoothing filter, which we call the**Harmonic Oscillator**indicators. The**Savitzky-Golay**filter can compute derivatives (rates of change) of the smoothed prices, and these smoothed prices and their first and second derivatives were used to construct the set of**Harmonic Oscillator**indicators with*zero lag*and a more straightforward interpretation than the normal ones. From these indicators, a set of**trading rules**was derived. These trading rules gave good results on the**Diagnostic Test**, although they only applied to a trading scenario in which the position is varied smoothly by making a small adjustment every day (which would only be practical for large portfolios).**Main Graph:**The Main Graph ofwas rather novel. It was a scrolling graph of the entire data set, presented in one long "panorama" which you could scroll across. There were four different scales on the graph, separated by a factor of two, and the horizontal and vertical axes both scaled the same way, so that the slope of the graph was preserved when rescaled. The price axis of the graph was logarithmic, so a constant distance along the vertical direction represented a constant percentage change. This made it possible to judge the relative performance of different stocks directly "by eye", just by comparing directly the slopes of the graphs (a feature not seen in the usual stock price graph). Underneath the log prices, the log volumes were plotted, relative to a reference line that represented the average log volume. This made it easier to see when the volume is larger or smaller than its average value.**StockEval****Portfolio Optimization:**also incorporated a**StockEval****portfolio optimization**routine, making use of the**Markowitz method**. The**correlation**between all the stocks in the portfolio (stock data files in a folder) was computed and used, along with the**expected return**derived from the**Price Projection**, to compute the**optimal portfolio**which would**maximize returns**and**minimize risk**. The parameters for**risk aversion**and**margin leverage**could be set by the user, to achieve portfolios with different degrees of**risk vs. return**.