Most frequent questions and answers
Yes, the algorithm is capable of detecting structural change by fitting locally stationary autoregressive models and can effectively handle structural trend breaks by doing model selection using Minimum Description Length. Effects of day of the week, week of the month, day of the month, month of year, special event, holidays etc. can be determined and controlled.
CrunchMetrics’ proprietary algorithm very efficiently handles business specific cyclical variations,
incorporates real time feedback from user to models and self-evaluates and iterates over true
targets. Also, it is more scalable and has significantly lesser latency than competing solutions.
CrucnhMetrics anomaly detection algorithm outperformed other competing algorithms by ~9%-
22% on Kaggle public datasets.
Below is an illustration comparing CrunchMetrics with other competing solutions and algorithms:
The algorithm splits train, test and validation data to measure the accuracy. We tune our models’ parameter based on business specific random and cyclical variations. Our proprietary model iterates over true targets and accuracy is evaluated using metrics like MASE, Scaled MAPE and quantile loss.
The algorithm works efficiently on any time series data and identify patterns in any time series
data. The algorithm has been designed to distinguish between systematic pattern and random
noise and identify any change in pattern in a fast and accurate fashion.
CrunchMetrics provide the ability to exclude data points from retraining based on events. An Event could be set to suppress the anomalies during a particular time frame, or it could be set to exclude few consecutive data points (based on time window) from retraining. User will have the ability to specify affected KPIs and dimensions (or cardinalities) for an event. Retraining will be happening periodically choosing a training window based on the rollup level. Within that window, if there was an event set to exclude the data points from retraining, system will exclude any anomalies occurred in that time window and will proceed with the retraining.
CrunchMetrics shows association between different KPIs and metrices using a tag-based
methodology. CrunchMetrics provide the ability to correlate multiple KPIs along with the ability
to establish a parent child relationship within the scope of the Tags. For example, if user finds an
anomaly pattern in a site, the same could be drilled down to the individual cells within the site,
which potentially causes the anomalies. Similarly, if user finds an anomaly in “Traffic” (Measure) in South Wales (Tag), the same could be correlated with the anomaly in “Connected Users” (Measure) in South Wales.
CrunchMetrics algorithm has the intelligence to automatically set the retraining window in a way that it covers all possible seasonality. With every retraining, the system will learn the seasonality in the data and will point out the anomalies accordingly. With each anomaly, system is intelligent enough to identify them and give less importance in the next retraining. In case of prolonged anomalies, CrunchMetrics provide ability to the user to set up events, based on which data points could be ignored from the next training.
CrunchMetrics indicates severity of an algorithm using a measure termed as Anomaly Score. The
anomaly score is a statistical measure to represent the deviation of the anomalous point from
the predicted / expected value. Higher the score, higher the probabilities of true anomalies.