Anomaly Detection

Flag irregularities and exceptions

The Problem

Anomaly detection usually isn't as simple as checking if a metric is greater or smaller than a given value, because the time series fluctuates due to seasonality and trend, creating long-term peaks and troughs.

How can you implement an anomaly detection system that is able to flag true inconsistencies in your data, and minimise the false positives?

The Solution

Our machine learning algorithms are trained to predict whether a given value is outside the normal bounds, given all features that determine what should be considered 'normal' for a given time slice.

The output of the model is an alert than triggers every time there is a spiked anomaly or a pattern in the data that is atypical.

We also build machine learning algorithms to identify the strong emerging trends in social media hashtags and can distinguish these from seasonal fluctuations and weaker trends that will plateau.


Stream data (usually time series)


1. A model that predicts whether a given time point represents an anomaly, outputting a predicted likelihood

2. Classification of anomalies by type, (e.g. system failure, demand spike, abnormal seasonal behaviour)

3. Implementation of the model into your existing tracking, or creation of a suite of visualisations that enable you to monitor the live system


Your anomaly detection system will implement state of the art technologies used by the likes of Twitter, alerting you to potential problem before it's too late.

More information