Forecasting with Prophet

Instruction: Detail the process of creating a forecast model using Facebook's Prophet library for a hypothetical time series dataset. Include considerations for incorporating holidays and handling outliers.

Context: This question evaluates the candidate's proficiency with the Prophet library, a tool developed by Facebook for forecasting with time series data. Candidates should discuss the steps involved in preparing the data, creating the model, tuning model parameters, incorporating holiday effects, and handling outliers. The answer should also include how they would validate and evaluate the model's performance.

Updated

Example Answer

I would first define the forecast horizon and prepare a table with dates in ds and the target in y. I would check duplicates, units, missing periods, and how much history supports each seasonal pattern. Then I would fit Prophet and compare it with a simple baseline on rolling forecast cutoffs.

For retail demand, I would include relevant holidays and planned events, then test whether those terms help on later periods. Adding a holiday is not a guaranteed improvement. For suspected outliers, I would distinguish recording errors from real demand spikes. A confirmed bad observation can be marked missing while keeping its date; a genuine promotion may need an event term. I would report horizon-specific errors and interval coverage, especially around those events.

Make it your own

Specify the target, event calendar, forecast horizon, and what qualifies as a data error.

Why this works

It connects configuration choices to evaluation rather than promised accuracy.

Follow-up to practice

What would you do with a genuine event that occurs only once?

Related Questions