[Paper Review] Economic Recession Prediction Using Deep Neural Network

Zihao Wang, Kun Li+2 more|arXiv (Cornell University)|Jul 21, 2021

Monetary Policy and Economic Impact41 references4 citations

TL;DR

This paper proposes a Bi-LSTM with Autoencoder and Attention Layer (BiLSTM-AA) model to predict U.S. economic recessions using macroeconomic and market indicators. The model achieves 94.8% accuracy and forecasts recessions up to 6 months in advance, outperforming traditional methods like SVM and logistic regression in out-of-sample tests for the 2008 crisis and COVID-19 recession.

ABSTRACT

We investigate the effectiveness of different machine learning methodologies in predicting economic cycles. We identify the deep learning methodology of Bi-LSTM with Autoencoder as the most accurate model to forecast the beginning and end of economic recessions in the U.S. We adopt commonly-available macro and market-condition features to compare the ability of different machine learning models to generate good predictions both in-sample and out-of-sample. The proposed model is flexible and dynamic when both predictive variables and model coefficients vary over time. It provided good out-of-sample predictions for the past two recessions and early warning about the COVID-19 recession.

Motivation & Objective

• To develop a deep learning model capable of accurately forecasting the onset and end of U.S. economic recessions using commonly available macro and market indicators.
• To address the challenges of temporal instability, data imbalance, and limited training samples in recession prediction.
• To improve early warning capabilities by enabling the model to predict recessions before official dates using only current-period data.
• To design a dynamic, flexible model that adapts to changing economic regimes and handles data latency through autoencoding.
• To evaluate the model’s robustness through rigorous out-of-sample testing on historical recessions, including the 2008 financial crisis and the COVID-19 recession.

Proposed method

• The model employs a Bidirectional Long Short-Term Memory (Bi-LSTM) network to capture temporal dependencies in time-series macroeconomic and financial features.
• An autoencoder is integrated to learn compressed, meaningful representations of the Bi-LSTM’s hidden states, enabling future-state prediction and mitigating data latency issues.
• An attention mechanism is applied to focus on the most relevant time steps and features during prediction, improving interpretability and performance.
• The model is trained to predict recession events at time $ t+1 $ using data from time $ t $, simulating real-time forecasting with a rolling window approach.
• Feature engineering includes standard macroeconomic indicators such as BAA yield spreads, industrial production (INDPRO), and term spreads.
• Model training uses a combination of supervised learning for classification and unsupervised pre-training on the autoencoder to enhance representation learning with limited data.

Experimental results

Research questions

• RQ1 Can a deep learning model with Bi-LSTM and autoencoder architecture outperform traditional machine learning models like SVM and logistic regression in predicting U.S. recessions?
• RQ2 To what extent can the BiLSTM-AA model provide early warnings for recessions using only current-period data?
• RQ3 How does the inclusion of an attention mechanism and autoencoder improve predictive accuracy and robustness in imbalanced, low-sample regimes?
• RQ4 What is the optimal time window for input features to balance economic cycle sensitivity and data latency?
• RQ5 How do key macroeconomic variables like BAA yields and industrial production influence the model’s recession predictions?

Key findings

• The BiLSTM-AA model achieved 94.8% accuracy on out-of-sample testing, significantly outperforming SVM (86.2%), logistic regression (86.2%), and DNN (89.6%) on the same test set.
• The model predicted the 2008 financial crisis and the COVID-19 recession with 6-month lead time, demonstrating strong early warning capability.
• With a 6-month input window, the model achieved the highest performance, with 81% F1-score for recession detection and 96% for non-recession periods.
• The attention mechanism contributed to a 6.9 percentage point improvement in F1-score over the model without attention, highlighting its role in feature prioritization.
• Feature sensitivity analysis showed that a 30% increase in BAA yields advanced the predicted recession onset, while a 30% increase in industrial production delayed it, confirming economic intuition.
• The model maintained high precision (97%) and F1-score (96%) for non-recession periods, indicating low false-positive rates despite class imbalance.