Prophet Forecasting

Prophet was released by Facebook's Core Data Science team as an open-source forecasting procedure. Its model combines a trend, recurring seasonal effects and optional holidays or other specified events. Trend changes can be represented through changepoints; weekly or yearly cycles can be fit when the history supports them. The official documentation says Prophet works best for series with strong seasonal effects and several seasons of historical observations. A model that fits one historical curve well still needs a test on later data. A daily series may have weekday and annual patterns. Prophet can fit seasonal components and include a holiday table with named dates. Those dates must be supplied for relevant history and future periods; the model does not infer every local closure or one-time disruption. The additive default treats seasonal effects as an amount added to trend. A multiplicative mode can be more appropriate when seasonal amplitude grows with the level of the series. The choice should reflect the observed pattern and be checked out of sample. The data cadence matters. Prophet's documentation warns that forecasts in regular gaps, such as weekend hours absent from a weekday-only history, can behave poorly because the unobserved seasonal positions are not identified. With monthly data, asking for daily values can also produce unstable within-month components. Forecast at the supported cadence and inspect the generated dates. Sudden shocks, changing promotions, policy changes or data corrections can invalidate a pattern that looked stable in training. Evaluate with rolling or time-ordered cutoffs, never random shuffling that lets later observations influence earlier predictions. Compare error with a baseline such as repeating last week's value or last year's same season, as appropriate. Review errors by horizon and by important periods, not only one average. Prediction intervals from the model are conditional on its assumptions and can miss unmodeled shocks. Record the training window, holiday calendar, settings and cutoff so a later forecast can be reproduced and challenged.