Time series modeling is a challenging area for data science. It is a difficult problem: it has its own history of specialized techniques and many modern machine learning approaches do not work without modifications. For example, decison tree based approached struggle with extrapolation and many approaches struggle with the dependencies induced between training examples. Traditional time series approaches, however, are not often able to take full advantage of all the data available in hetergeneous datasets. When solved, the business opportunities are plentiful. These include challenges that many companies face in operations and decision making. Example applications include sales forecasting, staff planning, and marketing.
This talk will cover how to use multiple open source tools to build a time series solution. This will cover several different modeling approaches including traditional time series approaches as well as how to adapt machine learning approaches to time series problems. Examples of approaches covered include the prophet library from Facebook, ARIMA models using statsmodels as well adapting well-known machine learning approaches. In addition, it will cover how to combine different time series approaches into a single model. It will also cover how to include business considerations into the solution such as building models when information is not yet available or it will take some time until the model can actually be used. As this talk will move quickly into a few different approaches and how to use them, familiarity with Python and machine learning approaches is assumed. No previous experience with time series is required.
As Data Science Engineering Architect at DataRobot, Mark designs and builds key components of automated machine learning infrastructure. He contributes both by leading large cross-functional project teams and tackling challenging data science problems. Before working at DataRobot and data science he was a physicist where he did data analysis and detector work for the Olympus experiment at MIT and DESY.
Event Timeslots (1)
Track A (Upper Floor)