,
5 tweets,
2 min read
Read on Twitter
My take: in Macro, careful analysis of the data is FEASIBLE and PAYS OFF, so it should be done before blindly applying off-the-shelf methods. Many new papers want to have cool buzzwords in their title but ignore the lessons from the previous decades & make basic mistakes....
There is tons of decades-old research highlighting the importance of careful data treatment in time series (trend, seasonality, outliers), which is simply IGNORED, with embarrassing results!
Moreover, many of these papers compare their performance to very simple models (e.g. AR(1)) which are not relevant benchmarks for a nowcasting problem, where the timely addition of incoming information is key...
Machine Learning methods seem to work very well with large cross-sections of iid data, but macro data is small in both the T and N dimensions, it is highly correlated both in time and the cross section.
More research is needed in understanding the performance of these methods in such environment. Until then, #MachineLearning and #BigData in economics will be just overhyped hashtags for policymakers wanting to sound cool!
