A recent article in Technical Analysis of Stocks and Commodities helped me understand what Wilder's Relative Strength Index is all about. I'm going into detail about this because it will help me understand it better and perhaps will help you as well.
This classic indicator of 'overbought' and 'oversold' price levels, first published in 1978, was intended for application to NON-TRENDING assets. The levels of 30 and 70 as the boundaries of normal, and the 14-day lookback period, were empirically derived in the context of certain commodity markets, and are not appropriate for all markets at all times.
Wilder's method can be expressed as RSI = 100 * SUMUP/(SUMUP + SUMDOWN) over a particular lookback period, where SUMUP is the sum of price changes from close to close on up days and SUMDOWN is the absolute value of the sum of price changes from close to close on down days. For some examples:
Price goes up $1 on Monday and $2 on Tuesday -- SUMUP is $3, SUMDOWN is $0 so RSI = 100 over two days
Price goes up $1 on Monday, down $2 on Tuesday, and up $1 on Wednesday -- SUMUP is $2, SUMDOWN is $2, and RSI is 50 over three days
Price goes down $1 on Monday and $2 on Tuesday -- SUMUP is $0 so RSI = 0 over two days
To put it simply, RSI will be greater than 50 in bullish circumstances and less than 50 when the bears rule.
The really interesting questions are:
How many days of data are enough?
Where should the 'overbought' and 'oversold' levels be set?
What about trending markets?
To take the first question first, consider a commodity with a perfectly regular price cycle 28 (trading) days long: the price goes up from $25 to $75 over the first 14 days and then drops back to $25 over the next 14 days and so on forever. Obviously the best lookback period would be 14 days and useful levels for 'overbought' and 'oversold' would be $70 and $30 respectively. As long as this commodity behaves like that, you can make a lot of money by buying when the price hits $30 and selling at $70. The implication is that the best lookback period for a non-trending commodity would be half its cycle length.
Note, however, that the appropriate levels for the 'overbought' and 'oversold' triggers may differ according to the lookback period. Assuming you would be satisfied with trades that work 95% of the time, AND assuming that RSI values are normally distributed around a mean of 50, you could calculate appropriate trigger points for any given lookback period. The 90% confidence limits are defined as 0.5 +/- 1.645 * SQRT(0.25/n) where n = number of periods. Results must be multiplied by 100 when applied to RSI. It turns out, interestingly, that 70 and 30 are not mathematically related to Wilder's original 14-day lookback. The article in TASC V. 32:5 (28-32) shows these calculations in detail and includes a table with some sample results. But take note of that second assumption....
Especially note that many assets have irregular cycles and a tendency for prices to trend up or down over many weeks or months. This is typical of equities among others. In that case, it is typical for the RSI to enter 'overbought' territory repeatedly and then make new highs, and the inverse can happen with 'oversold' levels on the way down. The RSI will therefore mislead you into selling winners and buying losers when applied to trending assets. That is because, in a trending market, the mean RSI is not 50. That's really important to know!
What the RSI can tell you about trending assets is the direction of the trend. In an important uptrend, RSI values will tend to hang above 50, while major downtrends will show RSI values chronically below 50. Therefore it's important to look at the RSI over time before you start trading any 'overbought' or 'oversold' signals. A good candidate for RSI trading will show regular cycling between 'overbought' and 'oversold' readings.
You may find RSI data helpful for trending assets when used in conjunction with another indicator that shows major trend violations. If, for example, RSI shows a burst of 'overbought' readings and THEN there's evidence that the uptrend has been violated (topping patterns, moving average crossovers, or trendline penetrations), you can point to this accumulation of sell signals with more confidence than you could to an unsupported trend violation. In that case, the extreme RSI readings could indicate a 'blowoff top' or 'selling climax'. As always, accurate forecasting depends a great deal on experience and judgment.
Bottom line: NOTHING about RSI is magical, especially not the classic lookback period and 'overbought - oversold' levels.