TimingCube - The ins and outs of backtesting - financial model backtesting

Current Signal Performance as of

Signal Type	Trade Date	Index	Return since issued
		Nasdaq 100
		Russell 2000
		S&P 500

Market Update

Stocks posted strong gains this week as several events lined up almost perfectly to lift the market. As a result, the major averages all rose each day of the week. First, investors cheered the cease-fire between Israel and Hezbollah in Lebanon. Then Tuesday, the PPI (Producer Price Index) for July came in at 0.1%, well below the 0.4% reading that was expected. The core PPI, which excludes food and energy, even fell by 0.3%. These numbers imply that inflation is under control and give the Fed more room to hold interest rates steady. Stocks rose all day after the release of the report. On Wednesday, the core CPI (consumer Price Index) also came in below expectations at 0.2%, reinforcing the view that inflation is in check. Just as they did the day before, stocks surged on heavy volume and the positive market action caused our Model to issue a new Buy signal after the close Wednesday. Supported by lower oil prices, stocks kept going up Thursday. They did so again Friday despite disappointing earnings news from Dell, as the market cheered a boosted buyback plan from Microsoft.

For the week, the Nasdaq 100 and Russell 2000 respectively gained 6.03% and 4.81%. As for the S&P 500 , it closed 2.81% higher. Both the Russell 2000 and S&P 500 are now back above their respective 50-day and 200-day exponential moving averages (EMAs). As for the Nasdaq 100, it has reclaimed its 50-day EMA and rests just below its 200-day EMA. With the improved market tone, we now have a Buy signal in effect.

Trend Timing School

The ins and outs of backtesting

Before going too far we would like to thank you all for the many supportive comments you sent us after the publication of the current Model's backtest and the statistics comparison. It goes without saying that there were also many questions about the process of backtesting and how it is done, which we will attempt to answer here. Yet, as many of your questions highlighted, last week's published results were a snapshot of one single time period (1989 to July 14, 2006) and one single index (Nasdaq 100 ). You want to know what the returns would have been for your particular index/ETF/mutual fund of interest. You also want to see results over specific, more recent time periods. Well, your requests have been answered on the "Results" page where all the tools now reflect the current Model which went into effect on July 14, 2006 and its backtesting to 1989. Returns for all "live" TimingCube Signals since June 18, 2001 can be found on the "Trades History" page.

For those of you not familiar with the features of the "Results" page, it is worth pointing out that it has proven a very handy tool to assess how well a particular index or investment vehicle would have performed with the four TimingCube strategies, over any period of time since 1989 through today, including 2006 Year-to-Date. Many subscribers have used this tool in conjunction with our recommendations to steer their portfolios towards the stronger indices (e.g. international markets). The data allows gauging both the relative strength of various indices and their correlation with the broad markets and our Signal. Results are available by market index for the three we track (Nasdaq 100, Russell 2000 , S&P 500 ), but also for any ticker symbol of your choice and for pairs of bull/bear mutual funds from the ProFunds and Rydex families. The tools range from Annualized, Cumulative, TimingCube Wealth calculator and TimingCube Chart. Most of the tools have a pull-down menu to select the starting date.

Coming back to the topic of backtesting, we need to stress that it is the basis of all investment strategies, because only observation of past markets can produce viable strategies. Before a broad stock market investment strategy can be applied with any confidence, it needs to be tested, improved and tested again over wide data samples, meaning long periods of time and many broad market indices. Suspicious systems are over-optimized for selected time periods. Having Rules A for period 1, and Rules B for Period 2 is called curve fitting, and cannot be trusted as valid backtesting (because going forward, one never knows which of the rules should be applied). Our mandatory conditions are that any rule enhancement must be applied to AND must improve performance over the entire period.

All investment strategies are somehow based on the principle that what happened in the past will happen in the future. The catch is that all models will encounter new market conditions as the markets adjust to counter profitable strategies by the masses. It is the nature of markets. In turn, any successful long-term investment model must include a continuous improvement process to adapt to such new market conditions.

Our initial testing period dated back to 1989 and after our first signal on June 18, 2001 our system produced outstanding results for the 5-year period. When taking a closer look, in early 2004 the U.S. markets entered a different "trendless" phase which had not been seen to this degree in our entire backtesting period. The first 2½ years after going live were incredible, but as the markets changed, the returns disappeared and in 2006 became losses.

Anyone who analyzed our backtested results has noticed that the earlier years have not changed significantly between our original Model and the revised one. This is because that testing period (the so-called "in-sample" data) had been well researched and tested before, and few enhancements remained. When it comes to the newer market phase, there were several enhancements, such as the detection of concurrent up and down trends, which became apparent. These enhancements were applied successfully to the entire backtest period (1989 - July 14, 2006) as mandated for any improvement to our Model.

So, how does the revised Model make such a difference in a mostly trendless stretch since 2004? Even though the market has been pretty trendless since 2004, there have been some up and down moves that our original model did not catch very well. The new Model does, and that is why the returns are much better.
We thought it was interesting to contrast the original and current Models by listing the improvements seen over the last 3-years in Table 1 below.

Table 1: Original and revised Models comparison (Cumulative % return for Long and Short strategy)

		Since 2004		Since 2005		2006 Year-to-date
Market	ETF	Original	Current	Original	Current	Original	Current
U.S. - Nasdaq 100	QQQQ	-7.66	67.25	-11.44	35.30	-11.21	11.00
U.S. -Russell 2000	IWM	16.40	97.15	5.49	46.39	-4.37	22.00
U.S. - S&P 500	SPY	9.88	39.12	1.77	19.96	-3.58	9.47
India	IFN	59.05	477.06	41.04	274.85	-17.07	92.21
Brazil	EWZ	16.40	135.50	57.15	103.29	-5.38	45.66
Emerging Markets	EEM	44.49	116.88	43.65	78.36	-3.08	38.74
Austria	EWO	78.62	215.22	50.31	82.03	-6.93	34.93
Germany	EWG	33.44	114.58	31.99	65.19	2.78	34.41
Europe, Australia and Far East	EFA	43.63	94.51	34.12	49.78	0.93	25.24
All values are as of market close on August 17, 2006

What is most telling is that improvements come evenly over the three most recent years (while also improving the long term since 1989), and that the improvements are broad based. We had to include India for laughs (it almost hurts to think about it), but frankly, we could not really find a broad market which did not do well in the backtesting.

After the backtesting, the real-time test is what really counts. Our Model had already tested well during bullish, bearish, and changing markets since 1989, but now, with 5 additional years to our "in-sample" data we expand the variety of markets our Model is proven against.

FAQ of the Week

Question: What are the current Model's average trade durations?

Since we published the backtesting of the current Model last week, a number of option traders have asked us to breakout the average trade duration statistic into Buy and Sell signals, as well as between winning and losing ones. This data can be useful in selecting appropriate expiration dates for option contracts. Still, remember that these are averages and that in fact signals have varied in length between 2 and 531 days.

Average duration of winning trades
Buy	243
Sell	61

Average duration of losing trades
Buy	45
Sell	25

For more information about option trading strategies, read the following Trend Timing School articles: Options basics and terminology, Simple option trades and strategies, and Implementing the TimingCube strategies with options.

Warm wishes and until next week.

The TimingCube Staff