The world is about to be flooded with data, with the incubation of the Internet of Things and the overall trend of digitization generating data in places that we have never seen before. According to research from Cisco, more than half a trillion devices will be Internet-connected by 2030. This new era of connectivity is resulting in a strong correlation between data signal creation and time-series data, leading to a whole new wave of data creation – and a whole new set of challenges. At Analytics Ventures, our expertise is in time series AI, and so I thought it would be helpful to provide a quick overview of time series data, time series analysis, and some everyday examples of it in practice.
In simple terms, time-series data refers to a consistent stream of data sets over the course of a period of time. Analyzing this type of data has become a recent area of focus in artificial intelligence, as accurate forecasting is becoming increasingly vital across all kinds of industries in order to make more informed decisions.
In more technical language, a time series is a stream of data within a time domain of a particular signal, for example, your heart rate. Your heartbeat is a single signal, with a stream of data measuring the activity of your heart rate over a period of time.
I was first introduced to the concept of applying AI to time series analysis by observing NASA’s work surrounding solar winds. I witnessed the disturbances to satellites by cosmic winds. Fascinated by this, and the correlations between the multiple frequencies and irregularities, I realized that this was an entirely new world to explore in data and analytics.
It is clear that virtually every industry stands to benefit immensely from the automation of precision forecasting, from healthcare to finance, manufacturing, and distributed energy.
Machine Learning has been proven powerful in imaging, natural language, and speech because of huge annotated datasets available. On the other hand, time series problems usually do not have big annotated datasets. Also, the data from different domains exhibit considerable variations in important properties and features, temporal scales, and dimensionality. Further, time-series analysis requires the algorithm to learn time-dependent patterns within and across multiple modalities, unlike images or speech. Time series analysis mostly includes clustering, classification, anomaly detection, and forecasting – each of which is uniquely useful to the business.
Going back to your heart rate, an application of classification a sudden change in your heart rhythm (cardiac arrhythmia) can be detected by the algorithm to identify this fatal problem. Just like the doctor, the algorithm, by seeing enough cases, would learn to identify the pattern underlying this problem.
Essentially, applying AI to time series analysis allows us to better uncover the meaning that is hidden in patterns. In today’s market, this is the type of intelligence that business leaders are really after.
Another example of time series data is the everchanging price of a stock on the stock market. The stock itself progresses through the course of the day, with its price fluctuating in tandem. Finally, the stock closes out at the end of the day. And, like clockwork, the journey starts up again the next morning – just as the sun rises each day.
Let’s say a grocery chain has a historical time series of the demand for granola, and how often the granola has been sold in a particular store over the last three years. The chain wants to do a better job of predicting the demand for its granola on a day-by-day basis. To do so, it would use data sets such as granola sales on a single day over the last three years, overlaid with a seasonal dataset, or a dataset around a certain holiday, day of the week or price. By overlaying different times series elements, a certain sequence of patterns emerges, creating confidence around the ability to make accurate predictions around how much granola to stock on the shelf.
It is important to point out here that no two days are the same, regardless of what we spend our days doing – there is always a different set of circumstances to react to. That is where Machine Learning comes in, rather than just your standard data analysis, because it is impossible to adjust to what you have never seen before with standard data recognition.
We live in a world of time series. Historical data is simply the starting point of the deep learning process, which is why applying AI to time series analysis is one of the more exciting recent innovations in the space. With 5G and the IoT, a wealth of data is about to be unlocked across the globe, and with time-series AI significant benefits can be realized across all kinds of verticals for all kinds of populations.