Improving empirical models and forecasts with saturation-based machine learning

This paper combines two threads of Harry Markowitz's research-uncertainty and data mining-to demonstrate a methodology for evaluating and improving the accuracy of empirical models and forecasts, focusing on forecasting. Machine learning with indicator saturation provides a generic framework that includes standard techniques for forecast evaluation, such as mean squared forecast errors, forecast encompassing, tests of predictive failure, and tests of bias and efficiency. Saturation techniques are applicable to both economic and non-economic models and forecasts. This paper illustrates the methodology with forecasts of the U.S. federal debt and of the U.S. labor market. Forecast evaluation is fundamental to assess the forecasts' usefulness and to specify ways in which the forecasts may be improved.