Inflation forecasting using tensor-train recurrent neural networks
School of Business | Master's thesis
Unless otherwise stated, all rights belong to the author. You may download, display and print this publication for Your own personal use. Commercial use is prohibited.
AbstractThis thesis studies the effectiveness of Tensor-Train Recurrent Neural Networks (TT- RNN) in the field of inflation forecasting. The research question addressed is: "How effective is the TT-RNN model in forecasting inflation?" To explore this question, a novel machine learning-based (ML) approach was employed in which RNNs were augmented with the TT decomposition method. The research aimed to assess the TT-RNN model’s forecasting accuracy across various horizons for inflation in 12 different countries, spanning 1-month to 12-month predictions, in the year of 2022. Empirical work involved a comprehensive comparison of the TT-RNN model against a standard RNN model as well as benchmark models commonly used in macroeconomic forecasting, including Autoregressive Integrated Moving Average (ARIMA), Vector Autoregression (VAR) and Dynamic Factor Model (DFM). In comparing the models’ performances, root-mean-square-error (RMSE) metric was used. The empirical findings revealed that the TT-RNN model outperformed the benchmark models in short, medium, and long-term forecasts across diverse economies. Specifically, the TT-RNN model produced the most accurate forecasts in 38 out of 72 instances, and in 21 instances where it did not rank first, it had the second-most accurate predictions. This demonstrates the versatility and robustness of the TT-RNN model in a wide array of inflation forecasting scenarios. Notably, these findings challenge the conventional wisdom regarding the suitability of Big Data models and especially the DFM model in inflation forecasting. In a volatile economic environment characterised by complex data patterns as in 2022, the TT-RNN model’s inherent non-linearity proved advantageous, allowing it to capture and adapt to intricate and complex patterns effectively. The TT-RNN model’s superior performance while being able to reduce the number of parameters in the network underscores its potential to reshape the landscape of macroeconomic forecasting. This research leads to the conclusion that advanced neural network (NN) models like the TT-RNN may offer superior alternatives to traditional Big Data models in various forecasting scenarios, highlighting their adaptability and predictive power in complex and dynamic environments.
Thesis advisorKitti, Mitri
inflation forecasts, neural networks, recurrent neural networks, tensor-train decomposition