Envision...
Current Update: February 28, 2011, #2
The final result of this signals set, trading a total of 1,628 positions, initially sized at 0.05 standard lots, with compounding to a final lot size of 20.30 standard lots, over two years, is as follows:
Chart 11: 9 Trading Models, 1,628 positions, profit baseline $2,228,064, PF=2.51
The compounded interest performance of the system is 44,561% ROI in two years. This calculation assumes an initial account size of $5,000 and entails a MAXIMUM potential risk of 19.39% sustained over the entire 24 month period. (That is to say, this backtest return is achieved by never subjecting the account to as much as a 20% total risk at any one time over two years.)
The *average* order size of all orders placed over the two years, is 2.83 standard lots, due to sizing increases caused by compounding.
The following list recaps the improvements obtained by the modeling procedure from February 6/11 forward. This list is of the non-compounded results only:
Chart 1: 6 Trading Models, 637 trading signals, profit baseline $8,895, PF=1.53
Chart 2: 6 Trading Models, 595 trading signals, profit baseline $10,525, PF=1.71
Chart 3: 8 Trading Models, 696 trading signals, profit baseline $11,667, PF=1.67
Chart 4: 9 Trading Models, 737 trading signals, profit baseline $12,581, PF=1.68
Chart 5: 9 Trading Models, 669 trading signals, profit baseline $12,639, PF=1.72
Chart 7: 9 Trading Models, 665 trading signals, profit baseline $12,977, PF=1.75
Chart 8: 9 Trading Models, 703 trading signals, profit baseline $13,497, PF=1.76
Chart 9: 9 Trading Models, 769 trading signals, profit baseline $14,546, PF=1.79
Chart 10: 9 Trading Models, 1,628 positions, profit baseline $41,159, PF=2.30
Previous Update: February 28, 2011
Finally.
After 516 lines of detailed methodology task logging, started on February 6/11, the newest signal set for "cable" is complete! Not counting the first six days of modeling during Feb 6 - Feb 11, a total of approximately 416 CPU hours of modeling work has been performed, multiplied by approximate CPU core usage of 11.1 cores at work simultaneously, each at 3.84 Ghz. This works out to approximately:
63,833,702,400,000,000 clock cycles -- about 64 quadrillion internal clock tics.
In order to complete this signal set yet another engine enhancement was coded to facilitate arbitrary clone position placement within signal drawdown zones. Version 4.3, hopefully the last enhancement to the 4th generation trading engine, brought the current 4th gen trading core to 14,470 lines of code.
The end result with the clones placement is a signal set unlike anything we've seen before: Due primarily to the new fine-tuned stop control features in the latest 4th generation engine we have witnessed the first signal set that actually runs the engine with *on average NEGATIVE risk*! What this means is that during backtested operation over 24 months, with the engine placing and controlling 1,628 trades, on average, at new position opening time, if all positions were immediately liquidated the system would actually increase the current bank balance, due to uncommitted positive equity represented in positive trades. As mentioned, this is something that has never been seen before with the engine, and is the result of continuous and aggressive dynamic stop control management on all open positions.
The final result of this signals set, trading a total of 1,628 positions sized at 0.05 standard lots over two years, with no compounding is as follows:
Chart 10: 9 Trading Models, 1,628 positions, profit baseline $41,159, PF=2.30
The simple interest performance of the system is 823% ROI in two years. This calculation assumes an initial account size of $5,000 -- which entails a MAXIMUM potential risk of 9.44% sustained over the entire 24 month period. (That is to say, this backtest return is achieved by never subjecting the account to even as much as a 10% total risk at any one time over two years.)
The next chart that will be published shortly is with the exact same trades, but with the sizing modified on the orders to maintain a 1:1 ratio between the lot sizing and the amount of USD in the account. (This will cause compounding, as for example, when the system increases from the original testing balance of $5,000 to $6,000, the sizing will increase accordingly from 0.05 to 0.06 standard lots.)
Previous Update: February 27, 2011
We have finally completed the fine-tuning of our new set of 9 signals. The total time to model the signals, not counting the final step that we are currently working on of placing the clone positions, has been 400 total hours x 11.15 CPUs, or 4,460 CPU hours. This equates to 6.2 x 10^16 CPU clock cycles, that is to say, 64 quadrillion ticks of the clock within the CPU cores. The end result is nothing short of astonishing, with an incredibly smooth and consistent set of 769 individual signals to trade over two years, at a Profit Factor of 1.79.
Chart 9: 9 Trading Models, 769 trading signals, profit baseline $14,546, PF=1.79
The final stage of this signal development project should be published shortly, giving the results *WITH* clone position placements both with and without compounding.
Previous Update: February 23, 2011
The current fine-tuning of 4 of the 7 "signal termination" parameters is proceeding nicely, with 6 of the 9 signals already showing some improvement. The current improvements, as of 327 total hours x 11.15 CPUs, or 3,646 CPU hours, are shown in Chart # 8 and are summarized as follows:
Chart 8: 9 Trading Models, 703 trading signals, profit baseline $13,497, PF=1.76
Previous Update: February 22, 2011
Chart #6 is the last chart for the current phase of modeling. Two of the nine models improved their PFs and profits, so a new consolidated backtest was executed at 309 hours total CPU time x 11.15 CPU cores. The final result for this phase is:
Chart 7: 9 Trading Models, 665 trading signals, profit baseline $12,977, PF=1.75
The idea of re-modeling the trend signals has been changed, with the focus shifting now instead to a final fine-tuning of the some of the dynamic stop control parameters. This is commencing immediately. (We are also working on a potential methodology improvement related to re-modeling the trend dependencies by dedicating a single CPU core to testing this change.)
Previous Update: February 21, 2011
Nearing the end of the current phase the system has achieved Profit Factor of 1.72 with the same (but further tuned) signals used in the Chart 4 update on Feb 19/11. The total time is now at 285 CPU hours x 11.2 CPUs x 3.84 Ghz which equates to 4.4 x 10^16 CPU cycles into the current modeling. The Chart 5 summary is as follows:
Chart 5: 9 Trading Models, 669 trading signals, profit baseline $12,639, PF=1.72
Regarding risk and return calculations that should be applied against this modeling: The testing is all performed on a large value account with minimum size orders in order to ensure that automated compounding features of the engine do not produce ever larger trades in relation to the account growth in order to prevent compounding effect skewing. With the testing order size of 0.05 standard lots, under the current testing the volume of orders and maximum potential drawdown represent a maximum possible loss at any one time during the 2 years of back-testing of only $520 and an average potential loss at the time of opening each new order of only $98. What this means is that an account size of only $5,200 with this order sizing and no compounding would achieve growth to $17,839 in two years under this current model, which represents a simple return of 234% GROWTH in the two years, without using any compounding.
To illustrate the skewing power of compounding, Chart 6 is provided that shows the use of the exact same 9 trading models, operating in the exact same manner, on an account of initial size of only $5,200, while allowing the size of orders to grow proportionally with the account. Interestingly, the engine keeps the maximum risk quite low, hitting only 12.14% max risk during the 24 month backtest with an average risk at the time of each order opening of only 2.09%. At the same time, the compounding effect raises the return rate to slightly over 645% GROWTH in the two years, hitting a final balance of $38,762 from an initial investment of $5,200 in 24 months. That said, due to the more violent drops at higher sizing, the Profit Factor lowers to 1.66 with compounding turned-on.
We have decided to run the current non-bidirectional trading models (that is, 7 of the 9 models in the current modeling group) through the final version 4.22 engine for another round of trend-protection-signal modeling. This is being done as that final version incorporates a new "constrained time delta" feature that may serve to provide more accurate market trend protection features to these trading signals. This will extend the development period of this set of trading models by something in the order of 4 to 6 days.
Previous Update: February 19, 2011
The fourth chart is from the combined signal test done on February 19/11 at 232 hours x 11.2 (approx) CPU cores = 2,599 CPU hours over-clocked to 3.84 Ghz per CPU core = 3.6 x 10^16 processing cycles, that is to say, 36,000,000,000,000,000 (36 quadrillion) cycles, give or take a few trillion.
Before going into the details of this current backtest, a quick summary of all four of the 2-year backtesting charts is provided:
Chart 1: 6 Trading Models, 637 trading signals, profit baseline $8,895, PF=1.53
Chart 2: 6 Trading Models, 595 trading signals, profit baseline $10,525, PF=1.71
Chart 3: 8 Trading Models, 696 trading signals, profit baseline $11,667, PF=1.67
Chart 4: 9 Trading Models, 737 trading signals, profit baseline $12,581, PF=1.68
The most recent chart, #4, represents the start of the "modeling within trend constraints" phase of the abbreviated methodology being employed at this time. It is likely that we will invest around 10 quadrillion or so CPU cycles (around 3 days with 11 over-clocked cores) improving the A.I. during this present phase.
Previous Update: February 18, 2011
The third chart is from the combined signal test done early morning of February 17/11. This signal generation result (at 162 hours x 11.25 (approx.) CPU cores = 1,823 CPU hours at 3.84 Ghz), obtained with avoidance of the curve-fitted trend signals that contributed to the second result at 89 hours, results in a total of 696 trades over two years at a Proft Factor of 1.67 (down marginally as expected.) To date, the stages used in this methodology have not yet implemented fine-tuned trading decisions within the market trends indicated by the trend indicating signals -- that is to say, trading decisions are still somewhat coarse, not yet fine-tuned to the market conditions that the signals will now tend to trade in due to completion of the trend support signals.
The next stage of this methodology currently being used has already commenced -- that is the fine-tuning of the trading signals within the trend constraints. The final stage will involve placing DDCC positions within the trading signal decisions. As usual, these two final stages will increase both Profit Factors and ultimately the number of trades as well.
Previous Update: February 16, 2011
Due to some out-of-range parameters two of the modeling CPUs crashed over the past few days, however we were able to salvage some of the modeling work from those two CPU cores. In the meantime we discovered a curve-fitting anomaly related to "trend protection signals" that is exhibited under certain circumstances of modeling. Due to this minor problem another release of the 4th generation engine had to be designed and programmed to support avoidance of that specific potential curve-fitting issue. This *hopefully final* version of the 4th generation engine is already in production use, reworking some of the signals for non-curve-fitted trend recognition. As this issue required revisiting the previous in-the-works signals, we accordingly expect the Profit Factor to drop off slightly on the next aggregate signal test, which we expect to have ready in the next 12 - 48 hours.
Previous Update: February 13, 2011
After only 89 hours of training (multiply by approximately 11.5 for number of active CPU cores, yielding approximately 1,024 CPU hours of A.I. modeling) we are seeing 595 signals/market positions being taken in 2 calendar years of backtesting with these new signals at an aggregate Profit Factor that has risen to 1.71. More fine-tuning of the trend detection parent signals will be performed over the next few days, followed by trading signal tuning, followed by DDCC placement work -- all of which will continue to improve the Profit Factors associated with these new signals.
Previous Update: February 12, 2011
Currently we are working on the final set of signals to be produced for the GBPUSD currency pair with the 4th generation engine. The early backtest results after only approximately 800 CPU hours at 3.84 Ghz (achieved in less than 3 calendar days with 12 CPU cores) are 637 trading signals, one position taken per signal, in 2 years from February 1/09 to January 31/11 at a modest Profit Factor of 1.53.
Watch for improvements in the number of trades and the Profit Factors over this backtesting time period over the next few days as we further develop these signals.