RSI Indicator

The formula

RSI looks at the last N price changes, usually N = 14. It adds up the gains and the losses separately and averages each. The ratio of the two averages, the relative strength, is then squeezed into a 0 to 100 scale.

The first average is a plain average of the first 14 changes. After that, the common method (Wilder smoothing) updates each average: new average = (previous average x 13 + this bar's value) / 14. Some tools use a plain moving average instead, so values differ a little between platforms.

A worked example

Take the made-up closes in the picture, starting 100, 101, 102, 101, 103. Over the first 14 changes the total gain is 12.0 and the total loss is 5.0. The average gain is 0.857 and the average loss is 0.357. RS = 0.857 / 0.357 = 2.400, so RSI = 100 - 100 / (1 + 2.400) = 70.6.

The price then falls from 108 to 88 over several bars. As losses enter the averages and gains drop out, RSI falls to a low of 28.3 at bar 24. That is below 30, the usual oversold line. At that bar the close was 88 and one bar later it is 89, so in this short made-up path the low reading happened to mark the low. In the straight fall in the next picture it does not: the reading stays at the floor while the price keeps dropping.

How a bot misuses it

• Buying oversold in a downtrend. The reading can stay low for as long as the price keeps falling. A bot that adds at each oversold signal is averaging down, so it grows into a loss. See average down and the DCA bot.
• The same 30 and 70 everywhere. The lines are conventions, not laws. The right levels, if any, depend on the market and the time frame, and finding them by trial is overfitting.
• Costs. A bot that trades every time RSI crosses a line trades often. Each trade pays the spread.
• Settings that change the number. A shorter period makes RSI jumpier, and a longer one smoother. Tuning the period to fit the past is a classic way to build a good-looking backtest that fails later.

How to check a rule that uses RSI

Test it across more than one kind of market: a range, an uptrend and a downtrend. Count the trades, and look at the average result per trade after costs and the worst losing run. Check the max drawdown. Then keep a part of the data back and test once, as in backtesting. The result in a downtrend is the one that tells you the most about the rule.

Reading it together with the trend

A common refinement is to use RSI only in the direction of a longer trend: take low readings as pauses inside an uptrend, and ignore them in a downtrend. That is a second rule, with its own settings, such as how the trend is measured and over what length. Each extra setting is another chance to fit the past, so the same tests apply to the pair of rules as to one. Compare the pair with the plain buy-and-hold result and with doing nothing, after costs, on data it has not seen.

What it does not tell you

It does not say where the price goes next. It summarises the last 14 bars and nothing more. No setting of it removes risk, and this page is education, not a signal.