In 2017, Meta open-sourced Prophet, a device for producing prime quality forecasts for time sequence knowledge that has a number of seasonality with linear or non-linear progress. It’s based mostly on an additive mannequin the place nonlinear tendencies are match with yearly, weekly, and day by day seasonality, plus vacation results. Prophet works finest with time sequence which have robust seasonal results and a number of other seasons of historic knowledge and is strong to lacking knowledge and shifts within the pattern, and handles outliers nicely.
Prophet’s star rose rapidly, and has been downloaded 32,247,931 occasions on the time of writing. Nevertheless, currently the device’s recognition has dipped. Dr Valery Manokhin, MBA, CQF, identified that current papers on time sequence haven’t used Fb Prophet as baselines because it doesn’t carry out nicely throughout any normal forecasting process.
“Each mannequin needs to be evaluated in opposition to easy benchmarks to find out if a mannequin provides enterprise worth. A easy benchmark might be utilizing present gross sales worth to foretell worth for tomorrow. Or it might be a price from final yr. Forecasting is a recreation of kin, not absolutes, and therefore to find out if the Prophet (or another mannequin) provides worth, one ought to set up a correct metric and naive forecast as a benchmark first,” he mentioned.
Criticism
Prophet was developed to make it simpler for consultants and non-experts to ship high-quality forecasts. In line with a examine printed final yr, the Prophet carried out poorly in comparison with AutoRegressive Built-in Transferring Common (ARIMA). Prophet’s forecasts have been subpar even when it used extra knowledge than different fashions. The examine concluded that the Prophet did not ship the products in a number of use instances. That mentioned, Prophet might be helpful in sure contexts the place fast and preliminary forecasts are wanted.
Prophet has been going through criticism for its underlying assumptions being weak or too simplistic. Additional, the device was designed to cope with holidays recognized prematurely and tends to overlook observations and outliers. Additionally, Prophet doesn’t straight think about the current knowledge factors which impacts the efficiency in instances the place prior assumptions don’t match.
Is Prophet nonetheless related?
In a current paper ‘ PROFHIT: Probabilistic Sturdy Forecasting for Hierarchical Time-series’, the researchers haven’t used Prophet as a baseline mannequin.
A baseline mannequin acts as a reference in a machine studying undertaking and helps contextualise the outcomes of skilled fashions.
“The explanation that Prophet doesn’t work nicely is that it’s a variant of inflexible Generalised Linear Mannequin. A curve-fitter that’s not aware of native patterns as a result of Prophet doesn’t embrace auto regressive phrases. There are papers and impartial research that demonstrated that even in conditions for which it was expressly designed for – particularly time sequence patterns with tendencies and volatility – Prophet is usually outperformed by different mannequin courses akin to ARIMA and plenty of extra.
“By now, it has been established by many tutorial research and papers that the Prophet isn’t an excellent mannequin. The truth is, after knowledge scientists like myself and others commented on this difficulty, Prophet improvement group eliminated a few of the unique efficiency claims from their web site,” Manokhin mentioned.
Regardless of all of the shortcomings, Prophet does have one benefit. Since it may possibly mechanically discover seasonal tendencies and provides easy-to-understand parameters, Prophet doesn’t require a lot prior information of forecasting time sequence knowledge.
