Abstract

SAMANIEGO ALCANTAR, Ángel. Portfolio Optimization with Long-Short Term Memory Deep Learning (LSTM). Rev. mex. econ. finanz [online]. 2025, vol.20, n.2, e862.  Epub Oct 03, 2025. ISSN 2448-6795.  https://doi.org/10.21919/remef.v20i2.862.

The objective is a methodology for weighting financial assets in an investment portfolio. It is contrasted by the components of the Dow Jones Industrial Average (DJIA). For this purpose, portfolios with investment horizons between 1 and 2 years are studied using Long-Short Term Memory (LSTM) optimization. The best portfolio was with an investment horizon of 1.5 years. The neural network is trained with 1,000 observations and more than 2,777 portfolios are simulated. The model outperforms the DJIA by 73% to 85%, with a geometric mean annual return differential between 3.7% and 5%. The components of the DJIA in history are used to allocate assets to portfolios between 2008 and 2021. It is recommended that the methodology be contrasted in conjunction with another methodology for selecting financial assets. The conclusions are limited to assets that make up the DJIA. Mostly in the literature, neural networks are used for the short term; this paper contrasts the model to the long term, seeking to weigh assets and not future asset prices. The conclusion is that the LSTM model can be used for this purpose, for investment horizons of 1 to 2 years.

Keywords : G11; G17; C61; Artificial neural network; portfolio diversification; deep learning; LSTM.

        · abstract in Spanish     · text in English     · English ( pdf )