Deep Learning for Global Tactical Asset Allocation

17 Pages Posted: 11 Nov 2018 Last revised: 3 Mar 2019

Date Written: October 19, 2018


We show how one can use deep neural networks with macro-economic data in conjunction with price-volume data in a walk-forward setting to do tactical asset allocation. Low cost publicly traded ETFs corresponding to major asset classes (equities, fixed income, real estate) and geographies (US, Ex-US Developed, Emerging) are used as proxies for asset classes and for back-testing performance. We take dropout as a Bayesian approximation to obtain prediction uncertainty and show it often deviates significantly from other measures of uncertainty such as volatility. We propose two very different ways of portfolio construction - one based on expected returns and uncertainty and the other which obtains allocations as part of the neural network and optimizes a custom utility function such as portfolio sharpe. We also find that adding a layer of error correction helps reduce drawdown significantly during the 2008 financial crisis. Finally, we compare results to risk parity and show that the above deep learning strategies trained in totally walk-forward manner have comparable performance.

Keywords: Tactical Asset Allocation, Neural Networks, Deep Learning, Regime Detection, Portfolio Construction

JEL Classification: C00, C10, C45, C50, G00, G11

Suggested Citation

Chakravorty, Gaurav and Awasthi, Ankit and Da Silva, Brandon, Deep Learning for Global Tactical Asset Allocation (October 19, 2018). Available at SSRN: or

Gaurav Chakravorty (Contact Author)

Qplum ( email )

Harborside 5, 185 Hudson St, Suite 1620
Jersey City, NJ 07311
United States
2013772302 (Phone)

HOME PAGE: http://

Ankit Awasthi

affiliation not provided to SSRN

Brandon Da Silva

OPTrust ( email )

1 Adelaide Street East
Toronto, Ontario M5C 3A7

Here is the Coronavirus
related research on SSRN

Paper statistics

Abstract Views
PlumX Metrics