# Outlier Detector with N-Sigma Confidence Intervals

7244 views
A detrended series that oscilates around zero is obtained after first differencing a time series (i.e. subtracting the closing price for a candle from the one immediately before, for example). Hypothetically, assuming that every detrended closing price is independent of each other (what might not be true!), these values will follow a normal distribution with mean zero and unknown variance sigma squared (assuming equal variance, what is also probably not true as volatility changes over time for different pairs). After studentizing, they follow a Student's t-distribution, but as the sample size increases (back periods > 30, at least), they follow a standard normal distribution.

This script was developed for personal use and the idea is spotting candles that are at least 99% bigger than average (using N = 3) as they will cross the upper and lower confidence interval limits. N = 2 would roughly provide a 95% confidence interval.
Release Notes: Added extra confidence bands for 95%.
I recommend the standard values N = 2 and N = 3 that provide, respectively, approximately 95% and 99% confidence intervals.
Note: I suggest using smaller sample sizes (between the 30 and 100 last candles) for sigma estimation as they tend to represent better the recent volatility. I also suggest to use sample size=400 for long-term average volatility.
Release Notes: Corrected bias in sigma estimation.

Remark: the original interpretation is a bit misleading. When the series crosses over the interval limits, one can say that the current candle length is 95% or 99% as extreme as the expected length.
Also in hypothesis testing, one could say that the hypothesis of the candle length being within the expected range is rejected at 5% or 1% significance level.

As closing prices can be seen as a random walk on chart, this series is basically modelling its error.
Release Notes: Updated graphic layout for better visualization.

An analogous approach for candle length is just thinking of it as changes in the closing price ( I would rename ir as Price Change Outlier Detector if TV allowed it!).
A green column means a positive change in price and a red column means a negative change in price. These changes are always relative to the last price.

If a column crosses from from the inner band to the outer band, the change in price is considered to be approximately 95% as extreme as expected.
If it crosses both bands, the change in price it considered to be approximately 99% as extreme as expected.
Release Notes: Minor bug fix.
Release Notes: Fixed sigma estimator so it doesn't depend on the present value.
Release Notes: Fixing last fix.
Open-source script

In true TradingView spirit, the author of this script has published it open-source, so traders can understand and verify it. Cheers to the author! You may use it for free, but reuse of this code in a publication is governed by House Rules. You can favorite it to use it on a chart.

Want to use this script on a chart?

I kinda see what's going on but not really xD
So how do you basically use it?
kiero
@kiero, can be used along with other indicators (RSI and Volume, for example) to detect dips as there's always a fast change in price. Candles that are 99% longer than expected can be good indicators. Of course everybody knows when a candle is locally longer than average, so this oscillator is just one formal way to measure if they lie within the "expected" range.
tvjvzl
@tvjvzl, thanks! I've been thinking it could be used to detect liquidity grabs: when a candle has a long wick, but is actually not an outlier. But maybe it's only good in hindsight, needs checking.