Forecasting Norwegian inﬂation with deep neural networks : the application and comparison of diﬀerent feedforward architectures
MetadataVis full innførsel
- Master Thesis 
This thesis investigates the feasibility of applying deep neural networks to macroeconomic forecastingintheNorwegianeconomy. Thethesisisintendedformacroforecasterscuriousaboutthe possibility of utilizing these approaches for macroforecasting. Deep Neural Networks is the most recent term coined for directed graphical models which act as universal nonlinear function approximators,optimizedthroughback-propagation. FocusingonmonthlyNorwegianyearonyear consumer price inﬂation, we design three diﬀerent network architectures, one representing the single hidden layer neural network ubiquitous in the literature and the remaining two representingrecentdevelopmentsintheﬁeld. Weapplythemodernandpragmaticapproachtodesigning network architectures, applying time-series cross-validation to tune network hyperparameters in a problem speciﬁc setting, giving ample time to the construction of optimal networks. Each network architecture is trained repeatedly to produce repeat ensemble forecasts of inﬂation, and the predictive acuity of these ensembles are evaluated against common linear benchmarks. We ﬁnd that both in the 2000 - 2009 period, used for network design, and in the 2010 - 2017 period, used for ﬁnal evaluation, at least one of our neural network architectures outperforms the best included benchmark for short term horizons. The forecasting improvements are generally found in times of high volatility. Further, we ﬁnd that the single hidden layer neural network is dominated by a deep multi-layer perceptron with residual connections. In the evaluation period, a deep convolutional neural network is the best overall forecast method, beating all benchmarks up to the six month horizon. At the three and six month horizons, the improvement over the best benchmark method is 18.2% and 10.6%, respectively. The convolutional neural network performs similar to the best benchmarks at longer horizons. While there exist barriers to the direct implementation of these networks in macroeconomic decision making, we argue, based on our results and recent literature, that including these methods in large statistical model suites, often applied by central banks, could indeed improve forecasting performance.