Estimating Value-at-Risk (VaR) on time series data with potentially heteroscedastic dynamics is a complex challenge. This task often involves dealing with limited data and a high degree of non-linearity, which poses difficulties for both traditional and machine learning estimation methods. In this paper, we introduce a new Value-at-Risk estimator that leverages a long short-term memory (LSTM) neural network provides more accurate predictions of the VaR of each of the Tunindex, ADI, MASI and TASI indices. A number of previous studies have confirmed the predictive ability of a classical feedforward neural network compared to traditional statistical models. To this end, our study uses an LSMT neural network to produce forecasts of future returns which are then used to calculate the VaR through the Bootstrap Historical Simulation (BHS). The model is evaluated using the MAE. The empirical results indicate that the VaR forecasting is reassuring in the case of ADI index for the 99% or 95% VaR, but not at all in others indexes cases.