CrunchMetrics detects anomalies in any time series data based on cyclicity, trend and seasonality. In case when the data is cross-sectional and a particular KPI/metric can be defined as function of other associated KPIs/metrics, the regressor functionality fits the best curve by automatically selecting all optimum associated metrics and best model (linear, polynomial, lognormal etc.) based on distribution of data. A statistically significant deviation of actual value from the predicted value is called an outlier. Furthermore, alerts can be configured for the following scenarios:

• Scenario 1 – Anomaly in time-series (univariate) signal and an outlier from the regressor (multivariate) module
• Scenario 2 – Anomaly in time-series signal but not an outlier based on regression
• Scenario 3- Not anomalous but an outlier

The alerting and actioning modules can be configured to trigger custom alarms and actions for each of the aforementioned scenarios.

Example:

In the CPG world, sales of a brand and product is dependent on several factors like price, inventory level, competition, demand, expected consumers’ income, advertising cost, CPI and other external and macro-economic factors. A univariate forecast on sales might not present an entirely true picture. Leveraging historical data, CMVariate will automatically select all the predictors which can accurately predict Sales as a function of other variables (factors). CMVariate automatically selects the best fit model and automatically tunes the hyperparameter to get the best accuracy.

An example output of CMVariate can be:

Sales = α0 + β1*P + β2*CA + β3*IL+ β4*Sp + β5*CH

…………………..

Where: P – Price of the product
CA – Advertising Cost
IL – Inventory Level
SP – Sales of competing product
CH – Holding Cost

Anomaly detection points out to an unexpected behaviour in one variable based on its history and just provides a univariate view. An observed spike or dip in sales could be because of various other factors like change in price, low inventory, lower ad effectiveness etc. CMVariate evaluates the effect of other variables on Sales (variable under observation) and indicates the order and magnitude (percentage relative contribution) of effect that each of the other variables is having on Sales at any instance (anomalous point or series of anomalous points). Additionally, if there is a dip observed in Sales and Sales as function of other predictors is also expected to go down, the business can quickly investigate the root cause and take the corrective action.

In addition to root cause detection, CMVariate is also integrated with the alert module of CrunchMetrics to configure alarms for reducing business specific false alarms. For example – a particular regional head can choose to raise alarm if and only if there is an anomaly observed in Sales and Sales as function of other variables does not show the same directionality.

Summon the power of Augmented Analytics to help you identify risks and business incidents in real-time.

Sign up for a pilot run on your data

Shashank Shekhar

Shashank Shekhar is Data Sciences leader with diverse experience across verticals including CPG, Retail, Hitech and E-commerce domains. He joined Subex from VMware where he was heading Data Sciences practice for transformational projects. In the past, he has worked in Amazon, Flipkart and Target and has been involved in solving various complex business problems using Machine Learning and Data Sciences.