Questions and Answers
|
|
What are your Neural Networks abilities to predict stocks?
Our tools maximize using Neural Networks (NN) up to the limit what we can get from NN. But the problem of predicting shares prices is not in NN abilities. Any statistical method works well if the future is predetermined by historical data. Even the best NN with a huge input of historical prices cannot predict all the time very well if the historical factor becomes weaker than other ones. That is why NN as a perfect tool itself has limited abilities.
Have you done any forward testing?
When people say "forward testing" in trading systems, as a rule, it means that the system with indicator buy/sell signal was optimized on past periods and then was tested on future periods. In NN forecasting, a program is trained/optimized on past periods and then tested by comparing predicted and actual prices (these actual prices are excluded from prior optimization). However, it is called "back testing" to emphasize that it was tested on a past period relatively to users (but it is a future period relatively to the program). That past period, normally, the latest known one, is not used to train NN (out-of-sample).
We encourage our users to perform back/forward (out-of-sample) testing on their own for unbiased results.
I ran S&P-500 prediction using NNSTP-2 and it did not show properly the upward trend in March 2009.
You could try to narrow the input data period to make "Training cases number " around 100. By doing that, you increase the importance of more recent historical data. Also if cases number is more than a few hundreds, the forecast can be too noisy (please see the recommendation in Users Manual).
Please note, statistical predictions work the best for technically conditioned retraces. Since a big move and reversal happen in the stock market mostly because of fundamental reason or news, it is hard to predict such events using statistical methods. The testing of NNSTP-2 has showed - the stock market moves that were induced by news were not always synchronized with technical pre-disposition. So that it is important not to rely only on statistical forecast results, do not put too much trust in any method alone, and make your own conclusion.
Why SMAP-3 "Fitting Deviation" and the "Total Deviation" are always less than the lowest "Deviation" in the small spreadsheet at the bottom left hand side?
It is normal that a single cycle cannot fit perfectly. Most historical price curves consist of many overlapping cycles. So that total deviation between actual curve and composed solution (consists of many cycles) must be less than only one fitted cycle deviation.
|
Do your tools use out-of-sample data to test predictive capabilities?
Out-of-sample testing is used in SMAP-2, NNSTP-2 (plus sample testing), and FTA-2 (indicators and waves forecasts). More explanations presented in the User's Manuals.
When I click on the download link, my system shows a security warning. Is it safe to download and install your software?
The safety and quality of our products are our major priorities. If you are downloading from our website, it should be no problem. We use a secure web hosting. Also our products are well tested (including several industry-leading security solutions); we fix any reported bugs fast. If you have any trouble with our software, please report us at support@addaptron.com or use our on-line form to send an instant message.
Which SMAP-3 forecast is better, on "Advanced" page or the regular forecast?
A regular (automated) forecast is created by using three methods and one adjustment. In advanced mode, only a simple cycle analysis is used. However, it has more flexibility. As example, it enables working with arbitrary period cycles (some users believe in particular cycles).
Why in some cases the SMAP-3 forecast in "Advanced" mode can differ from a regular one?
A regular (automated) forecast is created by using three methods composed with auto-weight and one adjustment. In advanced mode, only a simple cycle analysis is used but it has more flexibility (and more "human factor"). That is why these two forecasts may differ even for the same time frame. Advanced option can be useful for the users who believe in particular cycles and prefer to set own periods, for example, two-week cycle (period 10 trading days).
Help section Advanced in User's Manual has some explanation (if you have outdated version, please update your SMAP-3).
I got the error message "The application failed to initialize properly..." after installation. Any advice?
If you are receiving this error message, you need to install the MS .NET Framework on your PC. Please make sure that you follow properly the pre-installation procedure described here: http://www.addaptron.com/download.htm
|
|
What data do I need to use the software packages?
TFW-1 uses the data provided by Addaptron Software (by clicking on the button "Get data from cloud").
Concerning SMFT-1, users should provide the software with input data. SMFT-1 works with simple input files. They are the most commonly used EOD CSV-files. The file format is the following:
Date,Open,High,Low,Close,Volume,Adj Close
2011-09-21,2.90,2.96,2.77,2.79,137000,2.79
2011-09-20,2.91,2.96,2.89,2.89,102300,2.89
. . .
The files should be copied under C:/SMFT1/INPUT/ subdirectory. Also there are data for fundamental analysis in FTA-2. They are entered manually by users. More details can be found in the user's manuals after installing the software.
I used SMAP-3 and noticed that sometimes back-testing and forecast failed. Any advice?
SMAP-3 uses optimized parameters and automatically re-adjust some of them depending on a situation. The only thing that can change forecast even in an opposite direction - is a chosen time frame (historical period for input). Users should use a common sense to choose a good one or better analyze a few ones. It is important to use back testing curve deviation from actual curve to estimate the accuracy for a selected historical period. The current version of SMAP-3 combines result (with different weights depending on back-testing success) from two methods. One of them decomposes a curve into trend and periodic components. More details about used models presented in the Manual (Help menu).
Forecast built on statistical methods may fail because of redistribution of driving factors - news, investors' sentiments, regulatory impact, etc. You could play with time frame and find the one which gives a very good back-testing; however, the forecast may still fail. That is why it is important do not put too much trust in any method alone. There is an Expert Method. As its illustration, an experimentalist shows a pen and asks about 40 people to write down their estimate of the length. Then he collects notes and calculates the average number - normally it is almost 100% accurate. Why it works? Everyone makes error in different directions so that averaging gives a precise result. In the same way - it is wise to get forecasts using different methods (tools) and make your own conclusion.
I noticed that other stock cycle software packages compute a Bartel's test statistic to assess the statistical significance of each cycle. Does SMAP-3 do that?
The Bartels Test is used to measure the stability of the amplitude and phase of each cycle. It is a good approach to study a stable system with a certain set of cycles. The stability of cycles can be important. However, please be aware, the stock market is more like as a evolving semi-cyclic system.
A few latest versions of SMAP-3 fit curve by assigning higher weights for the most recent historical period that allow increasing sensitivity to the freshest stable cycles. SMAP-3 starts fitting process by searching for the most stable cycles (constant amplitude, period, and phase) with the biggest amplitude. So that one can analyze the cycles significance by their position on the list (use Output → Select Output File → View TXT-file). Upgraded SMAP-3 displays the deviation between each solution (extracted cycle) and actual curve. The deviation is in inverse proportion to statistical stability.
|