The sad reality is that many/most people trying to perform forecasting don’t really understand how the models work and the limitations, let alone how to test a time series model.
I tend to avoid inter disciplinary forecasting research because participants from other disciplines can’t understand or don’t want to know about the limitations and push ahead regardless.
Would you be willing to share some resources on testing time series models? I'm in the "many people" boat right now with gradient boosting, and that kind of thing sounds super interesting. Hard to know where to start as an outsider.
Not the OP. I find the clasic Time Series Analysis book of Box & Jenkins to be a good starting point for "basic" time series forecasting. And even that book is a difficult read: pleasant book but the theory and concepts are still hard. A deep understanding of stationary signals is a must for time series forecasting which is pretty hard to grok in my limited experience.
> “…we’re going to focus more on forecasting than, say, classification or anomaly detection—though a lot of the libraries below offer these features too.”
Doing some weekend warrior hacking in the accounting space. Reading this passage, I’m wondering if I could expect these packages to help with reconciliations--namely, multiple system reports for sales Vs. deposits and fees…??
Right? If you're going to forecast, then you could certainly apply the same to more-likely-than-not matching.
He has a wrapper for all of these: https://github.com/microprediction/timemachines
And he compares their performances against each other: https://microprediction.github.io/timeseries-elo-ratings/htm...