This paper applies a recurrent neural network, the LSTM, to forecast inflation. This is an appealing model for time series as it processes each time step sequentially and explicitly learns dynamic dependencies. The paper also explores the dimension reduction capability of the model to uncover economically-meaningful factors that can explain the inflation process. Results from an exercise with US data indicate that the estimated neural nets present competitive, but not outstanding, performance against common benchmarks (including other machine learning models). The LSTM in particular is found to perform well at long horizons and during periods of heightened macroeconomic uncertainty. Interestingly, LSTM-implied factors present high correlation with business cycle indicators, informing on the usefulness of such signals as inflation predictors. The paper also sheds light on the impact of network initialization and architecture on forecast performance.

翻译：本文应用循环神经网络LSTM进行通胀预测。作为时序建模的理想模型，LSTM能够顺序处理每个时间步并显式学习动态依赖关系。本文还探索了该模型的降维能力，以挖掘可解释通胀过程的经济学含义因子。基于美国数据的实证结果表明，相较于常见基准模型（包括其他机器学习模型），估计的神经网络虽具竞争力但未呈现显著优势。研究发现LSTM在长预测周期和宏观经济高度不确定时期表现尤为出色。值得注意的是，LSTM推导的因子与商业周期指标呈现高度相关性，验证了此类信号作为通胀预测因子的有效性。本文还揭示了网络初始化和架构对预测性能的影响机制。