Reading Valeriy's comments on the X post I still don't understand the hatred for Prophet. I know perfectly well how it works technically and I know its limitations perfectly, but at the same time I can state with certainty that for certain time series times it works very well

From my perspective the authors make bad claims. The title of the paper literally states it’s for “at scale” forecasting but takes like a second to fit for one time series and doesn’t have many speed ups available. It doesn’t really do well in most scenarios. It Leaves wayyy too much signal in the residuals. Hard Changepoints are also really scary to productionalize without knowing what you are doing but none of these points are ever brought up by the authors, instead they mostly say (at least used to) that it was a good automatic solution for time series. Now they have changed their tune but the damage is done. Prophet can win but so can forecasting the last value plus 10, doesn’t mean it’s actually good.

If you have time series with a clear trend and clear multiple seasonalities, it usually works quite well, often outperforming other approaches like ARIMA+MSTL in my experience. Additionally, you don't always find useful information in the residuals. Ignoring them can sometimes enhance robustness to noise, resulting in a simpler model that is easier to explain and interpret. However, you can also combine Prophet with another model to address residuals, or use NeuralProphet, which can also leverage any remaining residuals after modeling the trend, seasonalities, and exogenous variables.

I don’t think ARIMA or MSTL are super fantastic either. If you look at benchmarks like from monash or Peter Cotton’s micropredictions prophet lags miles behind. I probably don’t hate it as much as others but it’s not worth the time to fit it. If the trend is clear and you have multiple seasonality then it’s way easier to fit your own procedure or you can fit numerous other better optimized methods in the time it takes to fit prophet. From a professional note when we used to have prophet it would account for a very small percentage of “wins” in our stat engine.

Honestly, the benchmarks by Peter Cotton are more or less garbage. Prophet performs well with specific types of time series, like the one I mentioned earlier, which aren't included in Peter Cotton's benchmarks. If you have a time series without a clear trend or clear multiple seasonalities, and you can forecast it in an autoregressive way, it's obvious that Prophet won't work well for it, making such benchmarking entirely pointless. If you have mastered forecasting algorithms and understand when to use them based on their peculiarities, having a library like Prophet in your toolbox is still useful. P.S. I haven't used Prophet in a while, but according to the Git history, it seems that the latest versions have been significantly sped up. Additionally, Prophet can be easily parallelized using Spark or other distributed computing platforms.

last I tried pip installing it failed haha. But yeah Micropredictions is not super useful but is still a data point. We also have M4 and M5 both of which does not have a good showing for Prophet, I am not aware of a large benchmark that does show well for Prophet. Most of the time I see prophet showing up in like wind/solar predictions and other fields that are newer to time series. At the end of the day - there are just way better alternatives....like some of my work 😀 But ultimately that is why people don't like it - large promises with not much delivery on those promises and an overabundance of its use due to the parent company attached to it.