Time-Series Forecasting With Orbit
User guide to deploying Uber’s new model for sales forecasting
In a world of growing data consumption, time-series analysis has become an increasingly common and essential technique for data scientists. This step-by-step user guide to leveraging Uber’s new time-series model ORBIT is a continuation from 5 Machine Learning Techniques for Sales Forecasting. Together, these two posts elaborate on a few common forecasting methodologies.
Getting To Know Orbit
ORBIT (Object-Oriented Bayesian Time-Series) is a time-series package that was designed to be easily implemented and generalized. Although built on probabilistic programming languages, the Python package allows for the model to be deployed similarly to a scikit-learn model. The interface therefore allows for the easy implementation of Bayesian exponential smoothing and autoregressive time-series modeling.
Orbit is an evolution of the exponential smoothing model where predictions are generated using weighted averages of past observations and the weight of an observation decreases exponentially as time from current date increases. Exponential smoothing models have a level (smoothing) component and a trend component. Orbit currently contains two models, Local Global Trend (LGT) and Damped Local Trend (DLT). LGT is a combination of a fully additive ETS, ETS(A,A,A) and autoregressive model. LGT uses a log-transformation to account for multiplicative trend which means all input…
