Why Use Indicators?
There are obvious functional reasons to use indicators–not the least of which is the ability to have a means to create buy and sell signals instantaneously. In a market full of unpredictable events, having something to guide a buy or sell decision amidst all of the noise and confusion can be invaluable. Ultimately, indicators can even create automated buy and sell signals which gets you that much closer to sitting on the beach sipping margaritas so to speak—though I always caution against having these visions!
But taking a step back from the pragmatic, those who do a lot of testing and research will find that indicators are like “Lego” building blocks that can be put together to generate a far superior result. Doing so in a common sense fashion can make the sum far more valuable and robust than the individual parts (think AggM or AggZ etc). This is the hidden gem, and perhaps the most overlooked area in backtesting. The general idea here is to combined two indicators into one composite indicator. There is nothing better for research than preserving our sample size and number of observations . Every time we use a stock screen or a sequence of trading rules, we narrow the universe of either candidates or observations. The more conditions we use, the fewer the degrees of freedom we have. This gives us the potential for unstable results when tested out of sample. Note that the degrees of freedom is the most important part of the t-stat and other statistics that traders typically use. However, the actual calculation performed is always an approximation based on sample size yet typically ignores the number and type of variables used. Ultimately true measures of degrees of freedom would consider this factor–and many traders would be a lot richer for it. (On a side note expect me to produce a new and improved version of the t-statistic in the future. This will probably be more useful than any trading system or indicator).
Another feature I have noticed in testing dozens of variables (including fundamental data and economic indicators) is that many specific market effects–such as rules or conditions–have a great deal of noise and the potential for “false discovery” even after applying statistical tests. After all, no matter what the statistics we use, at a 95% confidence interval a full 5% of observations will be purely due to chance. The smart researcher will look at the stability by quartile or quantile of an effect–ie is there a consistent separation across buckets. Using DV Indicators which are almost always scaled by percentiles, it is very easy to do this kind of research. It also makes them a lot easier to combine in a way that will reduce noise and improve an effect. From this vantage point it is truly irrelevant which is the “best” indicator based on selective testing methods performed. The advantage of percentiles is that they can be easily combined and tested—try doing that with other indicators and you will quickly run into problems due to different measures of scale. If you can tell me how to combine MACD, RSI, TRIX and a Z-score into one efficient continuous signal–than I am all ears! Another advantage is that percentile indicators always give you a near constant number of trades so you can keep your exposure consistent. A final comment is that you always want whatever is working best now and it matters little to me what that happens to be at the time. If you have a complete set of indicators–trend ,countertrend, and composite versions, you will always find something that is working well.
The best base for adaptation that I have tested, is to use a strong effect such as daily follow through, rsi2, dv2, intermarket indicators or any other simple and frequently occuring condition. They are obviously not perfect due to their inherent simplicity, but they create a meaningful foundation for improvement. Of course the trick is you have to ensure that these effects are still relevant in today’s market. Ideally you can then use unrelated variables or indicators that may or may not be comlementary and then create a sort of Modern Portfolio Theory type portfolio of signals that will produce a superior final result to any of the individual components. There are many dimensions to consider for this type of approach-not the least of which is time. I will save deeper thoughts for another time, but I would encourage you all to think about indicators in a brand new way—as building blocks for research and signal or even numerical aggregation.