基于LSTM的政策效应预测模型及其应用

LSTM-based Policy Effect Prediction Model and Its Application

文章提出了一种基于长短期记忆神经网络(LSTM)的政策评估方法,将深度学习技术融入反事实推断预测。首先,使用LSTM方法拟合政策干预前处理组变量间的复杂关系;然后,预测政策干预后处理组变量的反事实推断结果,在此基础上测算政策方案的平均处理效应;最后,剔除模型误差效应,评估政策方案的经济效应。以湖北省外贸进出口优惠贷款利率政策为例,测算该政策带来的宏观经济效益。结果表明,基于LSTM的方法在预测精度上优于线性模型SCM-LASSO和人工神经网络(ANN),且模型误差效应的纠正可以显著提高政策效应的评估精度。

This paper proposes a policy evaluation method based on Long Short-Term Memory(LSTM)neural network,which combines deep learning techniques and counterfactual inference prediction.First,the LSTM method is used to fit the complex relationship between the group variables before the policy intervention,and then the counterfactual inference results of the group variables after the policy intervention are predicted.On this basis,the average treatment effect of the policy program is measured.Finally,the model error effect is eliminated and the economic effect of the policy scheme is evaluated.Taking the preferential loan interest rate policy of import and export of Hubei Province as an example,the paper estimates the macroeconomic benefits brought by the preferential loan policy.The results show that the LSTM-based method is superior to SCM-LASSO and ANN artificial neural network in forecasting accuracy,and that the correction of model error effect can significantly improve the evaluation accuracy of policy effect.