Meet AnaliTeG!

Dear Colleagues & Friends!

Today me and Петр Гринцер proudly unveil our new state-the-art Deep Learning model for forecasting inflation. Since it’s one of the first deep learning model designed for medium term economic forecasting we decided to break with long held tradition of deep learning community and avoided using “deep” or “net” in the model name.

AnaliTeG stands for Analitical inflationary Tensor Generator. The basis of our work are two papers: by Radford et al 2016. “Unsupervised Learning with Deep Convolutional Generative Adversarial Networks” and by Yeh et al 2017 “Semantic Image Inpainting with Deep Generative Models”

In essence we treated inflation data as “pictures” and trained DCGAN (Radford et al) on regional inflation data to generate random inflation samples from the same distribution as actual data.

On the second stage we used Yeh et al approach to generate not just random inflation samples but samples conditional on recent history and our assumptions about the future.

We used multiple averaging across all data axis (regions, price categories, time) to mitigate overfitting risk and reduce the impact of random shocks on a single forecast.

The model was trained on inflation data for 81 region 05.2005-12.2017. One region (Saints Petersburg) was left for testing. So the model test is in-sample on the axis of time and out-of-sample on the axes of regions and price categories.

On the intuitive level model acts as a vast consensus of experts who has been observing inflation data in various environments for many years and now are asked to guess how inflation may look like over next 12 months. Individual forecasts may be weak but as a consensus they perform well.

We spent nearly as much time creating framework for interpretation of the results as we spent building the model itself and now we hope that we managed to lift the lead of notorious “black box” ML models are considered to be. We can test various scenarios and measure sensitivities accounting for any second round effects and bidirectional causality between variables. In that sense our approach at some point may rival semi-structural (New Keynesian) and structural (DSGE) models but unlike them AnaliTeG doesn’t require priors containing our hypothesis about what economy actually is. The machine will figure it out itself.

We suffered from limited computational budget and lack of proper reference in economic literature, so our choices while building the model were arbitrary and in most cases simplest ones, designed just to test whether concept is viable rather to find best solution. So there is enormous scope for improvement. This is just the beginning of the beginning.

If any questions please contact me or Petr directly via messengers!

Sincerely yours,

Leave a Reply

Your email address will not be published. Required fields are marked *