# Solved – Median absolute deviation only can be used for anomaly detection for time series without a trend

I think MAD only can be used to detect anomalies for time series without a trend because it relies only a stable median to detect anomalies. It should be OK for time series with seasonality. Just seek confirmation here. This question is limited to when MAD is used as a standalone algorithm without coupling with other methods for anomaly detection.

Suppose we have a set of observations: $$(x_1,…,x_n)$$
$$Median = median(x_1,…,x_n)$$
$$MAD = median(|x_1- Median|, |x_2- Median|,…,|x_n- Median|)$$

Then we can use $$Median$$ +/- 3*$$MAD$$ as thresholds to detect anomalies.

Contents

As said in the comments, using MAD as you proposed assumes that you are dealing with i.i.d. variables. For time series this is obviously not the case, as the time-series changes over time, so the method would not be appropriate. What you could do instead, and is commonly done, is to adapt the approach to the fact that the distribution of the time-series changes over time.

• If you can assume, that the only thing that changes over time is the mean, than you could detrend the data first and then use the method like MAD. To do this, you would first need to estimate the trend of the time-series and subtract it from the data. For doing this, you could use something like rolling average, exponential smoothing, LOESS, or a number of other methods, depending on the mature of your data. One such example is given by Rob Hyndman in the answer to the Simple algorithm for online outlier detection of a generic time series question.

• However, it also can be the case that not only trend, but also variability of the time-series changes over time (see example below, taken from the Forecasting: Principles and Practice book by Rob Hyndman and George Athanasopoulos). In such case the whole distribution changes over time, so you need a method to account for that as well. Simple solution is to do windowed estimates (e.g. split your data to daily, weekly, monthly etc. periods and do local anomaly detection within the windows).

• Another case is if there is seasonality in your data. Then you need to do seasonal anomaly detection (how does this example differ from the data we usually observe on Mondays, or in August, or on 8 AM, etc).

Of course, in real-life data it might be the case that you would need a mix of those approaches, or a tailor-made approach for your data. The key take-away message is that you need to consider how does the distribution of your data changes over time. It is rather not i.i.d., otherwise you would not consider it as a time-series, so using MAD directly is a bad idea.

Rate this post