MIT researchers have developed a system that directly integrates forecasting functions on top of an existing time series database.

Its simplified interface, which they call tspDB (Time Series Forecast Database), allows the inexperienced user to generate a weather forecast, future stock prices, disease risk, and more in just a few seconds.

The new system is more accurate and more efficient than current deep learning methods in performing two tasks: predicting future values ​​and filling in missing data points. The success of tspDB lies in its new algorithm for predicting multivariate time series data, i.e. data containing more than one variable, which depends on time.

In the weather database, for example, temperature, dew point, and cloud cover depend on their previous values. The algorithm also evaluates the volatility of the multivariate time series to give the user a level of confidence in their forecasts.

Even though time series data is getting more and more complex, this algorithm can efficiently capture any time series structure. We seem to have found an appropriate lens to look at the complexity of time series data models,” says senior author Devavrat Shah, Andrew and Erne Viterbi Professor at EECS and Fellow of the Data, Systems and Society Institute of Information Systems and Decision Systems Laboratory.

Abdullah Alomar (Electrical Engineering and Computer Science PhD student) and Devarat Shah were joined by lead author Anish Agrawal, a former EECS PhD student and now a postdoc at the UC Berkeley Simons Institute, on the paper. The study will be presented at the ACM SIGMETRICS conference.