Deep Portfolio Theory is a portfolio selection method published by J. B. Heaton, N. G. Polson, J. H. Witte from GreyMaths Inc.
Authors' codes are proprietary, so I (this github repo owner) can only try to code this notebook myself for experiment. I am not the author and is not related to the original authors. This code may not achieve satisfying results as the paper states. Maybe I misunderstand some parts from the paper, so I hope that someone can continue the research and contribute to the framework. (you are welcome to open issues.)
You may find relevant papers according to the lists:
The authors use "auto-encoding, calibration, validation and verification" as machine learning steps. In computer science, we are more comfortable to call them "auto-encoding, validation, testing and verification". But we will still follow the terms the authors use in this repo.
For the graph below in Page 13, for convenience, let's name upper left, upper right, lower left, lower right as A, B, C, D.
Python 3, Keras (Tensorflow Backend)
Downloaded from Bloomberg Terminal
Dates: from 2012/01/06 to 2016/04/29 (aligned with the paper)
As Section 2 of the paper states, stock data shall be treated as a matrix $X \in R^{T \times N}$, a market of $N$ stocks over $T$ time periods. You can consider it like: $T$ is number of data points (varied), $N$ is number of features (fixed).
IBB Index Data (ibb_uq.csv)