Long-short term memory networks (LSTMs) offer a promising approach for analyzing shock propagation in multivariate economic and financial systems. To harness their potential, this working paper introduces the LSTM multiplier response function, analogous to the impulse response function of a standard linear VAR model. This response function presents several advantages. Firstly, it more effectively captures nonlinear dynamics inherent in complex systems compared to linear VAR models. Secondly, it accounts for the system’s current and historical states, reflecting the intuition that negative shocks have amplified effects under adverse conditions. Thirdly, the method involves applying shocks directly to variables of interest, obviating the need for establishing causality or orthogonalizing the system. To illustrate, the paper compares LSTM and VAR models by fitting them to a multivariate economic system. Leveraging the superior forecasting accuracy of the LSTM, it is demonstrated that the LSTM multiplier response function exhibits similar qualitative features to VAR impulse responses, highlighting its usefulness in economic and financial applications.