Integration of Econometric Models and Machine Learning- Study on US Inflation and Unemployment

In this study we compare the in-sample-accuracy to evaluate the performance of Econometric models and Machine Learning models on the Time Series data. Enclosed to explore techniques which perform better for Time Series Classification to predict the state (High, Medium, or Low) of each quarter by studying macroeconomic variables in the United States: Inflation and Unemployment. In the direction of improving the models using machine learning techniques and investigating how they are incorporated in time series data to improve the efficiency of the predictions. We perform a comparative analysis of various models for this classification problem. In ML, Logistic regression, K-Nearest neighbors, Support vector machines, Gradient boosting and Random forest models were explored. In Econometrics, Autoregressive Moving Average and Autoregressive Conditional Heteroskedasticity models were explored. The results showed that Machine learning models are superior compared to the traditional Econometric models for time series data. The best model for Unemployment data was EGARCH in Econometrics and K- Nearest Neighbors to predict both 2 states and 3 states in ML. The best model for Inflation data was EGARCH in Econometrics and Linear SVM, Random forest to predict 2 states and 3 states respectively in ML. Even though the ML models lack the interpretability and clarity in the exact internal process, these models have resulted exceptional in terms of accuracy in predictions. Econometric modelling would be more suitable, if we focus to only understand the effect and interpret the casual effect of the data.