System building from the ground up
One question I asked all the interviewees for my book, "Market Beaters," was whether or not they tested only in support of pre-existing ideas. I expected some disparity: while some would only be comfortable confirming their theories, others would accept exceptional results for their own sake. I was wrong. Virtually every trader wanted some idea of what was behind the findings. A trading theory pretty much always precedes number crunching.
System developers are obsessive about markets and their historic performance. Their constant chart scrutiny all but ensures a bottomless well of inspiration. That’s how I recently found myself re-visiting a frequent obsession: if and how the September–October stock market telegraphs its traditional downside trajectory ahead of time.
Let’s first make sure that the myth of market crashes does not significantly outstrip reality (good mechanical system developers accept nothing on faith). We know there have been historic periods of extreme down moves. By visually scanning those time frames, hopefully we can spot recurring behavior that we can incorporate into a system. "Bad days" (below) shows daily S&P futures charts around three notorious events; Oct. 19, 1987 (Black Monday), Sept. 11, 2001 and the financial melt of September 2008.
The next step is to establish a control: without filters or other qualifiers, how does the market fare during the three targeted time periods? "Trial and error" (below) reveals the outcome of shorting and then buying back the full-sized S&P futures between Sept. 1 and Sept. 30, Oct. 1 and Oct. 31, and Sept. 1 and Oct. 31.
We get a rather qualified confirmation of the autumn downward bias. It’s clearly not a good idea to hold shorts from the beginning to the end of October. Overall, September is a far more negative month, despite October’s legendary crashes. Those melts should make further investigation worth the effort, despite the non-dramatic initial findings.
There are probably many things the three charts have in common leading up to the price declines. Some of these aspects are non-consequential and will have no correlation with future profitability; some will have immediate validity. A good percentage will suggest something that may be valuable but should be supported by additional testing. Here’s what I noticed: It is historically profitable to short the opening of the fifth trading day of the month in September or October if the fourth day’s low was lower than the third day’s, and the third day’s was lower than the second day’s.
"Bad days" displays this price action. The setup occurred in both September and October of 2008. You would have made $27,025 in the September short and $26,175 in October. Your total would have increased to $77,450 had you let your original short run through the end of October (see "Optimize," below). However, when we extend the methodology to the entire S&P futures history, the results are less impressive.
The good news is we’ve improved over the original sell-at-the-onset, cover-at-the- close of the two months. The bad news is profits over the 28-year period fall short of the single anomalous crash year of 2008. Twenty-seven years, in other words, rack up a net loss. As is too often the case, we have to discard a seemingly good idea.
Our two choices from here are to give up or to rethink our approach. What if we started eyeballing the years which were counter to the expected down years, rather than those that confirmed the trend? The year 2009 comes immediately to mind. That was when the market bottomed in April and then steadily rallied through year’s end. Is there anything there that might jump-start another study?
"Plan B" (below) illustrates a cup bottom formation. A cup is a three day formation in which the second, or middle day, has the lowest low. The formation is most noteworthy when it is reversing a multi-day trend, as we see in "Plan B." It visually depicts a convex bottom, the same way the bearish "cap" counterpart suggests topping (middle day is the highest high). For years, my analysis has shown a wide application of cup/cap biases. Let’s continue to go short on the fifth day as before except when there is a completed cup formation within the first four days.
Each step has improved the key numbers relative to the previous incarnation. The best tool for comparing apples-to-apples is arguably the ROA, or return-on account. That figure gives you the percentage that your account would have increased had your initial capital been equal to the worst drawdown — $9,100 in the top row of the table in "Plan B." That row’s $30,800 divided by the $9,100 equals 3.3846; $30,800 is 3.3846 times as big as $9,100. Multiplying the figure by 100 and rounding down produces the displayed 338% increase.
We still don’t have a trustworthy system. While our process has eliminated negative outcomes, it also has dropped our total number of trials to statistically suspect levels. Besides, our overall net still falls short of 2008 in two out of the three tests. Our "better" numbers are deceptive.