Rethink Time and Data in Your Organization

Unlike the wishes Cher sings in her famous song, you cannot turn back time. But with the tools that have become available, you have a better chance of predicting time or, more accurately, predicting if occurrences in a time series sample will continue a decision-influencing trend.

Facebook Prophet and TensorFlow, issued by Google, are two machine learning protocols aimed at enticing developers to create exciting data science applications. Technology and analytics managers should view these tools as ways to expand their DataOps capabilities and expand their initial steps into machine learning.

Created by the Facebook core data science team, Facebook Prophet provides a reliable time series forecast where processing capacity is an issue. Prophet is based on an additive model to address how non-linear trends fit with yearly, weekly, and daily seasonality. The framework aids businesses when data contains periodic trends, such as retail holidays or the discovery that a sudden event impacted a trend. R programming and Python versions were launched a year ago, so businesses can leverage open source resources to create models. The source code and examples are available on GitHub.

I have previously reported on TensorFlow -- you can read about it here. The neural network framework also offers an additional suite of probability models; in R the models are called as a separate library. This allows for more advanced statistical models to be built into the model easily. In the case of time series, users can apply a Bayesian structural time series. A Bayesian structural time series is a set of probability models that includes and generalizes many standard time-series modeling concepts. Its purpose is to highlight statistical details for more accurate comparisons between time series data of current and previous periods. The TensorFlow probability library allows a model to incorporate the Bayesian Structural Time Series.

Why so much interest in time series reporting? If you stop and think about it, time series reports are as common as an Excel spreadsheet, since many tools display time series data. Just take one look at a web analytics solutions or social media analytics report.

But the visualizations of time series data in those solutions are not really designed with statistical analysis in mind.

For example, a web analytics solution like Google Analytics can provide time series results for referral traffic, and the results can permit decisions on which sources are consistently sending traffic to a website. But suppose you needed to predict how sustainable a trend for a given referral source can be? The slope of a trendline may not be immediately discernable from a flatline if the length of time is long enough. I speak from experience: Years ago I needed two and half days to determine the top conversion sources of my first client’s search traffic, because the visitor volume grew with a slow-developing growth pattern.

With today’s data sources, it is also likely that the frequency pattern of a given time series is not linear. This means observations would display successive increases and decreases in a logarithmic or curvilinear pattern. A tool with statistical capability would detect these nuanced trends far better than a standard solution would. Finance professionals who conduct stock market predictions know the value of better statistical capability well. They use advanced tools to create accurate time series predictions because noise and volatility in the data obscures the trend.

The latest tools make that needed statistical capability possible, speeding up analysis that creates meaningful decisions. Random noise in the data can be filtered out as well. The ability to separate data into components is why tools like Prophet and TensorFlow are so valuable. But advanced analysis can also be done in other dashboards such as Tableau, or they can be created as a visualization model in Python or R programming, as Prophet has provided.

Time series is a simple analysis that sometimes contains complex statistical nuances. Examining those nuances can reveal the right details quickly, helping a team make data-influenced decisions faster and better.