This paper is a practical review of training a Black-Litterman allocation algorithm with LSTM-RNN, the Long Short-Term Memory Recurrent Neural Network. The author describes and follows a well-establish training framework that provides feedback from the Black-Litterman framework back to the neural network. The novel approach can be interpreted as the machine learns from how Black-Litterman “thinks” and adjusts its views accordingly. Compared to most other research that disengages market-view generation from portfolio construction, we believe this paper shows more significant potential as market evolves.
This paper examines industry sector allocation for the US large-cap market. Unlike individual stocks, industry ETFs are with less noise, hence easier to learn. Historical price performances of the sector ETFs are plotted:
The paper uses 71 distinctive financial ratios for machine learning. The recurrent part of the network analyzes the time series of the prices and ratios. LSTM-RNN is a complex model. This paper incorporates a 3-layer model with about 31,000 parameters, which makes impactful training difficult.
The outcome highlights not just the performance improvement, but the smooth allocation that prevents large buy-sell swings. Statistically, the model outputs an enhanced, equally-weighted allocation. The table below shows the performance metrics comparing to value-weighted (VW), equally-weighted (EQ), and simple Markowitz optimized (MKW) benchmarks. MKW stands out when transaction costs are excluded; however, LSTM might be the winner if in actual investing.
Novel Asset Allocation: LSTM-Based Black-Litterman Model, Q. Wan, Master Thesis, Technische Universität München, 2020
Advisors: Prof. Dr. Christoph Kaserer, Prof. Peter Lin
Author: Q. Wan (linkedin.com/in/qianwan41)