In developing and emerging economies, the accuracy of macroeconomic forecasts is often constrained by the limited availability of data both in time series and in cross-section. Given this constraint, this paper uses a suite of machine learning methods to explore if they can offer any improvements in forecast accuracy for headline CPI inflation (y-o-y) in 3 emerging market economies: India, China and South Africa. For each forecast horizon for each country, we use a host of machine learning models and compare the accuracy of each method to 2 benchmark models (namely, a moving average forecast and SARIMA). For India, we find that the deep neural networks out-perform the benchmark forecast for all horizons except the 1 month ahead forecast. The reduction in forecasting error ranges from 44% to 63%. For South Africa, the neural network model provides a reduction in forecasting error between 42% and 57% for the 1 year forecast. For China, the reduction in forecasting error is much more modest ranging from 5% to 33%. An average forecast using different neural net methods performs much better than any individual forecast.