Simplify time series modeling without loss of information from high granularity data sets

Improve time series anomaly detection and classification

Develop software library for real-time anomaly detection

Background

Time series modeling is one of the most widely applicable topics in statistics and probability theory, from finance to manufacturing to business to engineering. Complicated time series can cause difficulty in anomaly detection, which can in turn hurt decision-making, monitoring, and forecasting. Many solutions leverage "machine learning", but succeed mostly in becoming uninterpretable, unreliable, and unstable.

Formal ARIMA modeling is a time-tested mathematical solution, and anomalies of various types can be formally defined rather than relatively defined. However, implementation of time series anomaly detection and classification is difficult and can fail for certain types of datasets.

We propose a new type of time series modeling called fuzzy ARIMA, in which we aggregate high granularity data sets into fuzzy numbers, thus avoiding the typical loss of information when aggregation is performed. This results in far simpler modeling, as shown in [insert citation], though these authors did not extend the paper to full ARIMA or anomaly detection.

USES

Business

Is a particular sales number anomalous? Is it indicative of a coming trend? Time series anomaly detection and classification facilitates better business decisions

Monitoring

Anything that needs to be monitored in real time--IoT, network traffic, metadata, and many other applications require fast decision-making, without a loss of valuable information. Troubleshooting requires knowledge of issues, not just flagging them.