I have made my four Collective2 systems more accessible by extending the Free Trial period to 30 days, and by dropping the subscription rate from $99/mo to $50/mo. This price drop is temporary because it is now below my standard subscription rate (http://www.schulenberg.com).

Hypothetical trading performance, whether calculated by Collective2 or TimerTrac (http://www.timertrac.com), can be rather close to actual (real-world) performance provided that: (1) trades are relatively infrequent, and (2) traded equities have high liquidity. In the case of the IWM/TWM/UWM system, only the equity IWM (Russell 2000) is traded -- and this is actually accomplished via the 2x ETFs (UWM/TWM) to provide a 'reasonably safe' amount of leverage. The leverage might be considered 'reasonably safe' because predictive methods are used to try to generate accurate signals for every trading day. Without this never-ending battle to predict market direction every day, drawdowns would be too great to permit trading such a volatile equity with 2x leverage. As it is, some unpleasant drawdown occurred at the end of June and the beginning of July when my predictive algorithms just didn't 'see' the market strength and persisted in staying Short. This was subsequently rectified by enhancing the Quarterly Logic to handle that particular kind of momentum-driven trend, but such after-the-fact improvements cannot change the past. My standard procedure is to do a post-mortem analysis every day that my prediction is grossly wrong, and try to add new algorithms (or enhance current ones) to deal with 'similar' market conditions in the future. The downside of such continual changes is that back-testing and optimization takes an increasing amount of cpu time (5-6 hours) every evening.

An important feature of this particular predictive system (IWM/TWM/UWM) is that it is only averages about 5 trades per month. This low rate is due to two factors: (1) the use of a predictive signal that holds positions for about 4-5 trading days on average, and (2) limiting the portfolio to a single ETF like IWM.

Historical (actual/hypothetical ... take your pick) performance of the IWM signal has been about 30% per year over the past 3 years (as tracked by TimerTrac), This is rather good, but it is interesting to compare this figure with the 'back-tested' performance of the IWM model. During much of the past 3 years the back-tested model has produced more like a 300% annualized gain, and this supports my empirical rule (which at least tends to be reasonably accurate for my kinds of models) that actual trading performance will generally be about 1/10th as good as the back-tested performance. There have been many modeling enhancements during the course of 2011, however, and back-tested performance is now closer to 400% ... suggesting that actual/hypothetical trading performance may reach about 40% per year (for IWM itself) ... or 80% per year (hopefully) by using the 2x ETFs (UWM/TWM). This is the target figure for the IWM/TWM/UWM system, which is currently annualizing at about 55% after 10 months in operation.

The development of the current equity models extends back about 10 years, and has generated over a million lines of code (although only about 300,000 lines have survived into the current models). There have been four major milestones along the way, and at each of them I was 'sure' that I had constructed the ultimate system, and that there would be little more to do...

The first milestone was reached in 2002-2003 with the development of the Mechanical Trading System (MTS). This program processed about 9000 equities, constructed complete portfolios of different sizes (stocks and ETFs), and was being set up for fully automatic trading. It was soon apparent, however, that trading accuracies were not as good in actual trades as I had expected. This led to converting this large piece of software into a Preprocessor whose sole objective is the calculation of an effective 'temperature' for the stock market. This was done by changing the buy/sell rules for the various portfolios in rather peculiar ways that permitted striking an analogy with an ideal gas (whose molecules could be compartmentalized into momentum bands), and using a financial equivalent of the 'Maxwell-Boltzmann equation'. From this point onwards the Market Timing Indicator (MTI) has remained the sole output of the Preprocessor, and this is a major part of the 'predictive' machinery used today. Another major part is comprised of the Generalized Candlestick algorithms, and the final major component -- a host of neural networks -- completes the predictive machinery.

The next milestone was the creation of the Grail System of 1000 customized stock and ETF timing models. As is reflected in the name, I really thought that this was going to be 'it' -- and in fact the trading accuracies of the Grail System models were better than those in the portfolios generated by the Preprocessor. Still, the annualized gains and drawdowns proved to be less earth-shattering than I had hoped. Nonetheless, the Grail System models (all 1000 of them) survive within my software system and are used as inputs to the next stage of refinement -- the Class "A" Voted Signals.

The Class "A" Voted Signals focused on 140 stocks and ETFs, and the idea was to utilize ALL available information (Preprocessor information, Grail System information, standard technical indicators, proprietary technical indicators, fundamental data, 'Fed' calendar data, VIX data, etc.) to produce the most accurate possible signals. Since much of this data is inherently contradictory, the 'voting' process tries to ensure that an optimal decision is derived from this large mass of data. The Class "A" Voted Signals are undeniably much more accurate than anything that had been developed previously, and these were used during the first year of TimerTrac tracking of the IWM signal (as well as QQQ, SPY, DIA, MDY, and SMH). But this was not to be the end...

The next step was to take all of the models and data (Preprocessor, Grail System, Class "A" Voted Signals, etc.) and try to increase performance (trading accuracy and drawdown) one more notch. The Class "B" Voted Signals are cross-voted signals, that is, each signal peeks at the other 139 stock and ETF models to try to deduce the relative strengths of different sectors and industries, and also to incorporate additional minor algorithms that try to detect short-term momentum-driven movements that would otherwise not be picked up by the predictive algorithms. These Class "B" Voted Signals are currently used to populate all of my various systems, but although there are 140 different stock and ETF signals, the ETF signals remain the most accurate, and of these the IWM signal remains the top choice (for the best combination of annualized gain, drawdown, and liquidity).

In addition to the major focus on stock and ETF timing models, another large collection of software machinery has been developed to deal with mutual funds. This has led to the creation of four long-term signals that are ideal for managing funds within retirement accounts. This is because they trade frequently enough to avoid many of the more significant market dips, but they do not trade too frequently to run afoul of the restrictions imposed by mutual fund and variable annuity managers. There are over 1000 mutual fund signals that are generated daily, and the mutual fund machinery includes some interesting logic that uses Monte Carlo techniques to try to 'deduce' the current holdings of every fund. For stock and ETF traders, however, the long-term signals are a useful guide because they help categorize the overall market trend. For example, Longs are substantially safer during periods when the long-term signals remain in the BUY state.
Schulenberg 2X-Hedged IWM (^IWM2)