Forecasting the Future Value of a Colombian Investment Fund with LSTM Recurrent Neural Networks (LSTM)

Recurrent neural networks are a tool that is currently used in time series, a widespread use of these networks is the forecasting of future prices in financial time series. One widely used recurrent neural network model is the LSTM (Long Short-Term Memory) model, proposed by Sepp Hochreiter and Jürgen Schmidhuber in their paper called LONG SHORT-TERM MEMORY published in 1997. This model solves the long term memory problem of recurrent neural networks by adding a selective memory cell which acts as a "filter" to choose what kind of information is important to keep and what kind of information is irrelevant and can be discarded. This paper seeks to analyze whether the original LSTM model is relevant to forecast the price of a Colombian mutual fund without taking into account exogenous variables (economic news, macroeconomic and microeconomic variables, underlying assets such as stocks, options, bonds, etc.) that affect its behavior. In this study, three funds with a majority composition of shares in the Colombian market were taken, which have higher volatility levels than public or private bonds.