Recession Prediction Using Multiple Machine Learning Methods and Historical Economic Data

Abstract

Abstract Recent economic recessions have been triggered by a diverse range of factors. Consequently, traditional predictive methods have only been partially effective at identifying them. This paper investigates the use of machine learning methods to improve the accuracy of economic recession prediction in both the UK and USA. LR, LDA, KNN, DTC, GNB, SVC, NN, RTC, LSTM, CNN and XGB models were tested with economic data from 1900 onwards. The UK data included gross domestic product, unemployment rate, inflation, FTSE 100 index, yield curve and levels of debt. For the USA the 50-day simple moving average 10-year treasury rates minus 50-day simple moving average 3-month treasury rates was utilised. Averaged F1, recall and accuracy for 100 iterations of each model were used to assess overall performance. Confusion matrices displayed each model’s predictions against actual events. During training, accuracy was high for all models, ranging from 1.0 to 0.81. Recall, F1 and confusion matrix results varied widely. LSTM performed best with recall and F1 values 0.96 and 0.97 and correctly identifying 11 in 12 Positive USA events. KNN, DTC, RFC and XGB displayed good results with recall 0.99 to 0.75, F1 1.0 to 0.69 and correctly identified 2 in 3 Positive events. LR, GNB and NN returned less reliable predictions with recall 0.24 to 0.54, F1 0.32 to 0.49 and correctly identifying 1 in 3 Positive events. LDA, SVC and CNN were inadequate with recall 0.07 to 0 and F1 0.12 to 0, returning few true Positive predictions. Using recent data, most models predicted the USA avoiding recession during 2023-24. However, the probability increased from 0.01 to 0.5 by mid-2023, then reduce once more. LR, LDA and LSTM initially predicted no recession, but the probability rapid increased to between 0.83 and 0.97 by April 2024. Although recession may be avoided, modelling suggested an increasing risk. The results confirm the usefulness of machine learning in recession prediction. The importance of diverse training datasets is highlighted. Performance of the individual algorithms varied, with several poor at accurately identifying the rare recession events amongst abundant non-recession events. The neural network models, especially LSTM and XGB proved most accurate. To further improve performance, work to refine the training datasets and further utilise advanced models, such as LSTM is required.