prophet forecasting tool illustrates the distinction between a traditional statistical approach compared to the newer machine learning/data science paradigm.
This distinction is cultural: it seems that the motivation behind
prophet was to quickly make accurate forecasts (predictions), instead of getting bogged down in building models satisfying certain theoretical properties, which may or may not yield useful results.
The “statistics” in
prophet may seem old-school, instead of new. The novelty is that the tool focuses on quickly getting useful results: you input your data and in 2 lines get your predictions. Then, you can fine-tune these by adjusting various “knobs”, or use the predicted curve in an experiment to see how it actually performs against another candidate model. Probably, this data-driven approach will yield better results much quicker than if you invested time into trying to satisfy model assumptions and worrying about other “academic” concerns.
The machine learning mindset is just that – a new and different mindset to solving problems, some of which may be old.
To illustrate this in another way, say you wanted to learn time-series forecasting. You could take a traditional statistics course on time-series methods, and you would probably learn a lot of theory but not spend a lot of time with real data to obtain useful results. For that, a better alternative would be to plug in some data into
prophet, and get a trend forecast that you could then fine tune by playing with it, and perhaps test in a (simulated) experiment. All on day one. How long would you have to wait to do this in a traditional course?
In essence, the newer approach is less focused on reaching promised “asymptoptia” and other abstract worlds, and more focused on quickly getting practical and useful results, which is especially important in industry.
Originally posted at pavopax.github.io