BigT1
Sergeant Major
- Messages
- 951
Pharaoh, Admnistrators, Phantoms, FPA Members, Suckers, and Scammer(s),. . . So, any updates? Think Dustin will offer to join the FPA's Performance Test Program?
.
Today I launch a full-blown 90-day forward test of TradeVantage AUE V1.0 from Herr Dustin Pass and Konstantin.
I have spent two weeks on and off, studying and implementing this complex Neural Network System.
I will attempt to explain what this is really about, what was omitted, and what was mis-stated during marketing; and together we shall see if this method is useful or not.
.
To start, I wish to thank Konstantin for providing me with the opportunity to perform this test. I wish it understood, as I have already stated several times, I am uncertain as to the degree to which Konstantin is a knowing participant in this appauling marketing fiasco. All I can say is that he has provided me with access to it and offered his support. I can also add that as I originally expected, Konstantin appears to be a knowledgeable programmer and computer scientist of high order, with weak English skills, which may have contributed to his being abused in business. However, I have no insight into his motivations, or his knowledge of securities law or business ethics.
.
The NN system Konstantin developed and Dustin Pass marketed obtains simple, raw daily OHLCV data from their dedicated, private server. It is accumulated and continuously updated for local user updates. The local user's client is to refrain from performing these updates except at three specific times each trading day, geared toward the convenience of forex traders around the globe. The client system expects and requires this once-a-day "Updating" to be done manually, at a very specific time. Failure to observe this restriction results in a partial current day bar that unduly influences the signaling. I consider this to be a design weakness. This updating should be automated in any future version upgrade releases. In addition, the server should not provide partial daily bars to the client, to avoid the likelihood of inaccurate calculations.
.
Users have a choice of executing daily trades on daily bars that close at the top of these (GMT-4) hours:
8am ET, the NY Market Open
3am ET, the London Market Open, or
7pm ET, the Asian Market Open.
.
Once a Time Zone is selected, the local client price bars will be generated based only on that chosen market's Close/Open time. Therefore, trading the system on three different time zones with differing results is entirely possible. This is a normal condition of trade for any Daily Charting system, including this one.
.
For this test I have chosen to use the Asian Open at 19:00 (7pm ET) for several reasons:
1- The trading week begins on Sunday at 5pm ET and ends at 5pm on Friday. This should produce 5 nearly equally lengthy time bars per week; Monday - Friday, plus a 2-hour bar for Sunday. This bar will emphasize weekend gaps of importance with high volume for analysis purposes, and be small and unremarkable in the calculations at all other times. Monday will be a 22 hour bar, as will be Friday. In my opinion, any other time zone choice would produce more compromised structures.
.
2- The majority of the world's daily forex trading volume on any given day will be fully contained within these daily bars. In addition, trade volume and price movement is generally minimal between rollover time (5pm ET) and the Asian Open, which really doesn't get underway until 8pm ET when Japan opens.
.
3- News events are very often scheduled for precisely 3am and 8am. This means added variability and uncertainty of obtaining consistent, non-chaotic updates of the daily bar charts close/open at the very moment all pairs are to undergo updating, analysis, and order entry. Ed: I consider this a serious design flaw in this program. The 7-8pm ET hour avoids almost all news, and provides the end-user a sufficient time window to accomplish the manual batch update and calculation process, plus time to consider and enter the daily trades with minimal slippage.
.
4- I have too much else going on at those other market opens to worry about this daily system.
.
Once the data is obtained, the newly minted system needs to receive its initial training. The system knows nothing about prediction without self-training. NNs work by breaking the problem of prediction down into a series of simpler questions, assigning a process to analyze each item, then combines the results from all the subordinate processes to find the best result. To the best of my understanding this system utilizes 3 Moving Averages, a Zig-Zag reversal pattern, plus preference inputs: expected trade frequency, trading style, and desired noise filtration. The number of layers and the number of neurons is adjustable, with the known trade-off being the computing time necessary to work all the elements is proportional to the required granularity.
.
Konstantin provides one basic "Type 1 Default" condition set, plus 4 additional pre-packaged Types for a total of 5 stock models, all with reasonably quick learning times of about 10 minutes per pair for 1-yr historical data sets. Together these 5 models yield better or worse results once learned on a given data range, due to the differing nature of the particular pairs being tested. The idea is to train each pair on all models ranging from "Inert" through "Chaotic", then select the model that can be said to have performed best over the input data period, for each pair.
.
Now here in lies the LIE!
Once calculated, the models reveal how they would have signaled during the learning period had each model indeed known this precise future data in advance! That is all a NN can do; find the best curve fit! The presumption is, that the past is prologue, and should this pair behave in the future as it did in the past, the curve fit strategy should be profitable. Therefore, it signals you to trade in the anticipated direction; but, possibly right into a brick wall (Roadrunner: Beep-Beep).
.
Ed: Now in physical systems like vibration analysis, fluid dynamics, aero-dynamics, such curve-fit methodologies have a strong correlation with reality, simply because the laws of physics are presumably immutable, and well studied molecules can be expected to behave in a predicatable manner. However, when it comes to chaotic, irrational, human behavior, these methods often have no known corollary in the physical world, behavior approaches random, and thus the basic theorem brakes down.
Therefore, those signals generated and displayed that occurred during the training period have no real relevance at all, either past or future. Only in the limit, as volatility approaches zero (calmest waters) does future reality approach the models' predictive result. Any market movement at all forces the model to later take the new information into account, and therefore, it is always certain to be wrong to some degree. That error is inversly proportional to the the complexity of the model, increases with the time that has elapsed since the last learning cycle, and with the quantity, accuracy, and length of the historical data included in the learning cycle, and with the current price volatility relative to the typical volatiltiy experienced during the learning cycle.
.
So it was on those retroactively optimized, curve-fitted, psuedo-results, that Herr Dustin and Company marketed this hope-filled product to you; touting phantom "learned curves" as "Past Performance"; as though these signals had actually occured in real time. Nothing could be further from the truth.
What a sin.
.
One might ask: "How certain can we be that the future will resemble the past"? My answer: "Not very certain at all". However, in order to permit a fair and proper test on each model's predictive results, NN program designers by design, provide a mechanism to Learn on a "Range of Data" from the past; but most importantly, also allow for the intentional exclusion of the most recent data from the "Learning Process". These excluded bars are intentionally held in reserve, so the finished, educated model can later be asked to make predictions on the "Real and most Recent" future data, without having had any prior exposure to it.
So certain recent daily bars can, and should be held "Out-of-Sample" during the learning process; then later used to confirm how well the model performs, had it been trained in the recent past and then allowed to signal its way forward into the future.
Later re-learning brings into the model all that data that was previously excluded, yet still leaves an unincorporated "Tail" for again projecting forward; and on you go, over and over again.
.
Ed: Konstantin and Dustin Pass certainly know all this in detail, for Konstantin indeed wrote it into the program! Furthermore, K openly suggests retraining the model every two to three weeks, a normalcy for this type of analysis. Yet, none of this was even hinted at in the sales presentations; and worse, was openly denied by their intentionally "kept ignorant" support team.
.
This Test.
In order to obtain the best results possible, I selected only the 12 most liquid and poorly correlated pairs from the 23 available on their server (~50%). I then built 60 learned models (5 for each pair), training for "The best combination of winning trades and profitability", rather than for "Maximum Profits" as was used in the public marketing. See the illustrated attachments. Also, the public reveal was done using the Default Model (Type 1), and with just 1-year of historical data.
.
In order to obtain even better results, I utilized 18 months of backdata; and initially excluded the last 30 days from the learning process. I then reviewed the 60 learned models for performance and drawdown. More importantly, I then pushed the models to predict over the latest 30 days. The results in every case were nowhere near as good as the models' learned outputs suggested they might be. I compared results from each group of 5 tests, and for 10 of the 12 pairs, I selected the most desireable model. Two pairs, had results so poor, that I chose to invoke the NN program's Advanced Function, enabling me to build deeper models, with more layers and neurons, and incorporating the better testing characteristics gleaned from the 5 basic runs. You will see these marked as type "Custom B" in the screenshots I have attached. Once I finished building the 12 most reasonable models, I then built 12 identical new models using these results and trained them on the full 18 months of data, right up thru the current date (June 2).
.
At 7pm tonight I "Updated" the models with the Sunday, June 3 Bar of 2 Hours, precisely at 19:00; and within the next 30 minutes I entered positions on 4 trades, signaled as new opens for tonight, or where I could enter those trades already in progress as signaled earlier, where I could get price advantage over the original entry price, and which were not too deep in the hole. Some existing trades in progress, where they were in extended gains or in deep drawdown, I ignored. Soon I will be synced up, as trades open and close day-over-day.
.
The demo test account is running at ATC Brokers, an ECN/STP provider clearing FXCM, and which charges commission. See my review of ATC on the Brokers Review page.
This daily trading model should be very insensitive to broker spreads. Stop Loss is placed at 2 times the max drawdown seen during the learning period, which can be quite large, so I am limiting position size to 0.10 standard lot per position on a $40K account. With 12 pairs, our max position size will be 1.2 lot at any one time, with an average size likely below 1 lot.
As of tonight, 4 trades are open. Again, see attached screenshots.
.
I hope everyone involved will benefit from this precision test.
Attachments
Last edited:
