Borrowing concepts from Meta’s Prophet to construct a strong time sequence regression mannequin
On this observe up article, I proceed my mission to construct Frankenstein’s time sequence monster by combining concepts from the favored Prophet package¹ and the discuss “Profitable with Easy, even Linear, Fashions”².
After we’ve reminded ourselves of what we’re as much as we’ll contact on the regression mannequin — what it’s, and why it’s particular.
We’ll then transfer on to hyper-parameter tuning utilizing time sequence cross-validation to get an “optimum” mannequin parameterisation.
Lastly, we’ll validate the mannequin utilizing SHAP earlier than making the most of the mannequin kind to permit bespoke investigations and handbook changes.
That’s a whole lot of floor to cowl — let’s get cracking.
Apart: we coated the underlying knowledge preparation and have engineering in a earlier article, and so are leaping straight to modelling. Compensate for what went on there:
Let’s remind ourselves of what we’re doing.
The top objective is straightforward: to generate probably the most correct forecast of future occasions throughout a specified time horizon.
We began from scratch with a time sequence containing solely a date variable and the amount of curiosity. From this, we derived further options to assist us mannequin future outcomes precisely; these had been closely “impressed” by Prophet’s strategy.
That brings us to the place we are actually: about able to feed our engineered knowledge into a light-weight mannequin, coaching it to forecast into the long run. Afterward we’ll dive into the mannequin’s inside workings.
Let’s remind ourselves of what the info appears like earlier than we supply on.
We’re utilizing real-world knowledge from the UK — on this case, the STATS19 highway site visitors accidents knowledge set which…
