Collaborating Authors

Detect the Changes in Timeseries Data


In this post, we will provide an example of how you can detect changes in the distribution across time. For example, let's say that we monitor the heart rate of a person with the following states: We can work with two different packages, the changepoint and the bcp. We will try to test the changes in mean. As we can see, it detected 4 distributions instead of 3. As we can see, it returns the posterior Mean as well as the probability of a change at that particular step. We can set a threshold like 30%.

The Time-Series Ecosystem


Time-series analysis has been studied for more than a hundred years, however, the extraordinary growth of data available from numerous sources and more frequent growth of data alongside the growth of computer power (GPU & Multicore) makes the analysis of large-scale time-series data possible today in a way that was not previously practical. The use of time-series data has been traditionally linked to sectors where time is not just a metric but a primary axis, such as in finance, Industrial IoT, and energy. However, in the last 10 years, it is starting to be generally used in other sectors such as marketing, gambling, or any other sector where performance monitoring and time-series analysis is needed. There are three main solutions in the ecosystem to treat, analyze, and visualize time-series data. These are Time-series Databases, Time-Series Data Analytics Solutions, and Machine Learning Platforms.

How do I forecast a timeseries of data using GARCH(1,1)?


HI, I'm new to GARCH, but I've got daily data of TV Ratings. I've been trying to forecast this for future, and a quick background - the data is non-stationary, has high seasonality (weekly, monthly & yearly). I've tried UCM, but forecasts for weekly data using UCM are easier to handle, and daily level of forecasts aren't making the cut. Which is when I turned to GARCH to see if I can quickly get some high level estimates into the future. I'm stuck with trying to get the forecasts for both the "conditional mean" and the "conditional variance" for t periods in the future.

Time-series Generative Adversarial Networks

Neural Information Processing Systems

A good generative model for time-series data should preserve temporal dynamics, in the sense that new sequences respect the original relationships between variables across time. Existing methods that bring generative adversarial networks (GANs) into the sequential setting do not adequately attend to the temporal correlations unique to time-series data. At the same time, supervised models for sequence prediction - which allow finer control over network dynamics - are inherently deterministic. We propose a novel framework for generating realistic time-series data that combines the flexibility of the unsupervised paradigm with the control afforded by supervised training. Through a learned embedding space jointly optimized with both supervised and adversarial objectives, we encourage the network to adhere to the dynamics of the training data during sampling.

Clinical Event Time-Series Modeling with Periodic Events

AAAI Conferences

The focus of this work is on developing models that can accurately predict events in complex multivariate event-time series derived from electronic health records (EHRs). One common characteristic of many EHR-based event time series is that they are periodic and events are repeated at regular time intervals. Examples of such events are the ordering of laboratory tests in intensive care unit (ICU). Since periodic events are quite frequent in EHRs in order to define a high accuracy event prediction process, the periodicity of the event occurrence needs to be properly modeled. In this work, instead of trying to combine all periodic information for all event time series into a common space (e.g. hidden space) we propose multiple simple periodic mechanisms that help us to drive the expression of individual events in time. We show that these simple periodic mechanisms can be effectively combined with more complex neural architectures capable of modeling the dependencies among different types of events. We test our new model on the clinical event prediction problem that consists of hundreds of clinical events in EHRs derived from MIMIC-III database. We show that our model that relies on simple periodic mechanisms is able to outperform competing baseline models in the multivariate event prediction task.