LinkedIn recently opened-sourced Greykite, a Python library originally built for LinkedIn’s forecasting needs. Greykite’s main algorithm is Silverkite, which delivers automated forecasting, which LinkedIn uses for resource planning, performance management, optimization, and ecosystem insight.
While using predictive models to estimate consumer behavior, data drift has proven to be a great challenge during the pandemic in 2020. In such a situation, predicting future expectations is challenging as well as necessarily helpful to any business. Automation, which allows for repeatability, can increase accuracy and can be used by algorithms to make decisions further down the line. According to LinkedIn, Silverkite has improved revenue forecasts for ‘1-day ahead’ and ‘7-day ahead’ and Weekly Active User forecasts for 2-week ahead.
Greykite library provides time-series tools for trends, seasonality, holidays, etc., allowing users to fit their preferred AI models. The library includes experimental plots and tuning templates, which define regressors based on data characteristics and forecasting needs such as hourly short-term forecasting and daily long-term forecasting. The templates’ tuning knobs cut down on time it takes to find a good forecast. The Greykite library also allows users to customize a model prototype for algorithms, allowing them to mark known anomalies (and decide whether to ignore or change them).
Greykite, which can detect outliers, can also choose the optimal model from a pool of candidates based on past performance results. Users may describe a collection of candidate forecast configurations that capture various types of trends instead of tuning each forecast individually. Also, the library offers an overview that can be used to evaluate the impact of individual data points. Greykite can, for example, determine the magnitude of a holiday, check the effect of a changepoint on the pattern, or demonstrate how a particular feature can benefit a model.
With Greykite, a “next 7-day” forecast that has been trained on over eight years of daily data takes hardly a few seconds to produce forecasts; LinkedIn claims that its entire pipeline takes less than 45 seconds to complete, including automated changepoint identification, cross-validation, backtesting, and evaluation.
The LinkedIn research team writes that the Greykite library’s algorithm, Silverkite is a fast, accurate, and highly customizable algorithm for forecasting. Greykite also offers model analysis with intuitive tuning options and diagnostics. It can be used to benchmark different algorithms and has a single interface for doing so. LinkedIn has successfully used Greykite for a variety of market and technology metrics use cases.