Does the RSI Indicator Give a Trading Edge?
The Relative Strength Index (RSI) is an oscillator that outputs a value between 0 and 100 which is intended to show how "oversold" or "overbought" a security is. If the RSI value is high, it's alleged that a down move is coming and vice versa. Does it hold up?
The video below is me showing my work, but you can skip to the results below if you'd prefer.
Tested Allegation
If the RSI value is overbought (greater or equal to 70) then the probability of a down day tomorrow is higher than 50%.
Parameters
14 day period (industry standard)
RSI value of 70 is overbought
Data
Data used is the Dow Jones Industrial Average daily close from 1928-09-30 to 2020-07-30.
Method
Conditional probability using Bayes theorem. This is a simple test using high school statistics. Ordinarily people would use a t-statistic for this sort of thing, but I think that over-complicates it.
The question this approach will answer is: what is the probability of a down day tomorrow, given that we are overbought today?
There are 23,046 samples in this particular analysis which is decent, but we can still fool ourselves with randomness. To avoid that, I'm going to run the same identical analysis against white noise data. I'll compute the RSI for that white noise data, check for overbought conditions, and check the same conditional probability.
If RSI looks like it works on random noise as well as it works on the real data, then it's not particularly likely that RSI is any more useful than random noise.
Results
The conditional probability of a down day following an overbought day for this series is 48.2%. In other words, you're more likely to have an up day than a down day following an overbought day.
I generated five random series to do the same analysis on for comparison, and these we the conditional probabilities for those particular random results:
49.2%, 49.5%, 47.6%, 53.5%.
In other words, RSI appears to "work" better on some of these random series than it does for the real data.
RSI is noise, not signal. (for this set of parameters and data)