This is the time of year when we typically start hearing seasonal predictions for the stock market. There are a few problems with seasonality, the biggest of which is that market data is highly random, and analytical tools will tease seasonal tendencies out of randomness. Until you develop some intuition about how “seasonal” random data can look, and how volatility can distort seasonality, it is hard to truly understand the value (or lack thereof) in seasonality.
I created a short video that uses the technique of “feeding” seasonal analysis random data with a known seasonal bias. In this video, you will see that even moderate volatility completely swamps the seasonal tendency, and also creates apparent patterns where none actually exist. Consider, also, that most of the random data did not have anywhere near the degree of volatility that actual market data typically has.
I have found this technique–creating random data with known but random characteristics and then analyzing it with the tools of technical analysis–to be invaluable. It is one of the best tools I have found to build a sense of how chance can impact our trading results, and to give deep insight into how tools and indicators actually work. (See the appendix of my book for other illustrations of this same concept applied to MACD.)
Pingback: Tuesday links: a time to garden | Abnormal Returns
A very nice and profound talk. Thank you for sharing this.
Why don’t you use 50-60+ years worth of data? Looking at 3 year segments is statistically insignificant and meaningless. No wonder why you can’t find a seasonal link.
There are plenty of good arguments for not using 50+ years of data. I have, of course, looked at both long and short term seasonality. Don’t assume that a 10 minute video contains the total content of all seasonal work I’ve done. Also, the point is not that “I couldn’t find a seasonal link”… the point is how stable and reliable those seasonal tendencies might or might not be.
Do you successfully invest using seasonality based on 50-60+ years of data?
Another excellent lesson in trading the statistics not the heuristics. I love it. Thanks Adam.
Thank you!
Hi Adam,
What is your recommended way to generate random price data? Something like the Geometric Brownian Motion model? I’ve tried a few different things when defining a baseline and, sometimes I realise that some “randomnesses” are more random than others; they all look like reasonable price series but, outcomes are not always consistent (small edges do not appear consistently with all methods).
Thanks and regards.
There are many ways to do it, depending on what you are trying to accomplish. One is to use some type of distribution (mixture of normals can work reasonably well). Another is to get actual price data, convert it to returns relative to the previous bar’s close (meaning high, low, and close are all returns relative to that previous close), and then “scramble” the bars into random order. A quick and dirty way to dodge the distribution question since you are using actual price data as your source.