Building on ideas from Meta’s Prophet package to create powerful features for time series machine learning models
Meta’s Prophet package¹ is one of the most widely-used packages for time series. At least anecdotally, according to me, after looking through a list of time series articles that I’ve bookmarked for later reading.
Sarcasm aside, I have used the package before and I love it.
Another great resource for time series modelling is Vincent Warmerdam’s talk titled “Winning with Simple, even Linear, Models”² where he touches on modelling time series with linear models (with a bit of preparation).
Now, there are some elements of data science which blur the boundaries of art and science — think hyperparameter tuning, or defining the structure of a neural network.
We’re going to lean into the art and do what a lot of the great artists have done: borrow ideas from others. So, in this series of articles we’ll be borrowing feature engineering ideas from Prophet, and linear modelling ideas from Vincent to perform our very own time series regression with a real-world time series.
Let’s touch first on what the overall goal is, before we hone in on feature engineering.
The overarching goal is simple — to generate the most accurate forecast of future events across a specified time horizon.
We’ll start from scratch with a time series containing only a date variable and the quantity of interest. From this, we’re going to derive additional bits of information which will allow us to model future outcomes accurately. These extra features will be heavily “inspired” by Prophet.
We’ll then feed our engineered data into a lightweight model, and let it learn how to best forecast into the future. Later on, we’ll dive into the model’s internal workings — after all, we’ll need to understand what’s driving our forecasts.
Now that we’ve seen the forest let’s get a close up of the trees, starting with a look at our data.
Source link
