Machine Learning Advances for Time Series Forecasting

Check out this fine survey by Masini, Medeiros and Mendes,

https://arxiv.org/pdf/2012.12802.pdf

For me the coolest thing is new insights into optimal regularization and subset averaging for density forecast mixtures. Amazingly, and very much related to the survey (but not widely recognized, including in the survey), optimally-regularized regression-based combinations and subset-average combinations are VERY closely connected. You can see the connection clearly in both of the papers below, in the first for point forecasts, and in the second for density forecasts. Effectively, the optimal regularization *IS​* subset averaging!

Diebold, F.X. and Shin, M. (2019), "Machine Learning for Regularized Survey Forecast Combination: Partially-Egalitarian Lasso and its Derivatives," International Journal of Forecasting, 35, 1679-1691.

Diebold, F.X., Shin, M. and Zhang, B. (2021), “On the Aggregation of Probability Assessments: Regularized Mixtures of Predictive Densities for Eurozone Inflation and Real Interest Rates,” arXiv:2012.11649.

Mixtures of Predictive Densities

Here's a new one, with Minchul Shin and Boyuan Zhang, "On the Aggregation of Probability Assessments: Regularized Mixtures of Predictive Densities for Eurozone Inflation and Real Interest Rates".

We propose methods for constructing regularized mixtures of density forecasts, exploring a variety of objectives and regularization penalties, and using them in a substantive exploration of Eurozone inflation and real interest rate density forecasts. The coolest thing is seeing what the regularization actually does. It has very different effects before and after the great recession. From the Great Recession onward, regularization moves density forecast probability mass from the centers to the tails, correcting for overconfidence. And it does so in real time, with no look-ahead cheating...

Welcome to 2021

 Welcome back!

I insist that 2021 will be better.  Here's a light one to start it off.  

I recently had occasion to visit the web site of Marcin Zamojski, a top young econometrician at the University of Gothenburg. He describes himself there as:

[I am a] frequentist at heart. I am also a strong believer in 'all models are wrong, but some are useful'. I am warming up to machine learning. If the data has a time dimension, count me in.

Wow!  Is that Marcin Zamoiski or Frank Diebold?