University of Twente Student Theses


Dutch Inflation Rate Forecasting Performance of Econometric and Neural Network Models

Plas, J.D.C. van der (2023) Dutch Inflation Rate Forecasting Performance of Econometric and Neural Network Models.

[img] PDF
Abstract:This thesis explores the importance of precise inflation forecasting in business valuation, recognizing the substantial influence of inflation on a company’s expected cash flows and risk, thereby shaping its overall value. The research specifically examines the performance of traditional econometric models alongside neural network models, undertaking a comprehensive comparison of their respective forecasting capabilities using the Dutch Consumer Price Index (CPI) as the key inflation measure. In order to anticipate future fluctuations in the general price level of goods and services, this study incorporates several macroeconomic factors as predictors of inflation, including historical inflation rates, money supply, GDP, interest rates, unemployment rates, and the price of gold. By investigating the performance of these models, this research aims to contribute valuable insights for businesses and decisionmakers, shedding light on the most effective methods for accurate inflation forecasting and ultimately enhancing the process of business valuation. The econometric models used to forecast inflation are ARIMA (Autoregressive Integrated Moving Average) and VAR (Vector Autoregression), which are both time series models. ARIMA models use a combination of autoregressive (AR) and moving average (MA) components to model the relationships between a dependent variable and its past values and error terms. VAR models, on the other hand, are used to model the joint behavior of multiple dependent variables. They assume that each variable is linearly dependent on its own past values, as well as the past values of other variables in the system. VAR models are particularly useful for modeling the dynamics of economic systems, where multiple variables influence each other. The three Neural Networks (NNs) that are investigated are Feedforward Neural Network (FFNN), Recurrent Neural Network (RNN) and Long Short-Term Memory (LSTM). A FFNN is a type of neural network where information flows in one direction, from input to output layer, through one or more hidden layers. RNNs work by using feedback connections that allow the output of a layer to be fed back into the input of the same layer, creating a loop that enables the network to remember previous inputs. Long Short-Term Memory (LSTM) is a type of RNN that can handle long-term dependencies by using a memory cell, which allows the network to selectively remember or forget previous inputs. Training these models involves optimizing the weights of the NNs to minimize the difference between predicted and actual values. The process involves an iterative procedure, where the model is fed with training data, the weights are adjusted based on the errors, and the process is repeated until the model reaches an acceptable level of accuracy. The training of FFNNs, RNNs, and LSTMs involves the use of backpropagation algorithm with gradient descent optimization, where the gradients of the error function with respect to the weights are computed and the weights are adjusted accordingly. The process of training neural networks is computationally intensive and requires careful selection of hyperparameters and optimization methods to avoid overfitting. The findings of this study are presented in Tabel 1 and indicate that the neural network models outperform the econometric models in the in-sample forecasts. Specifically, the LSTM model shows the best in-sample performance, suggesting that it is the most accurate model for predicting inflation based on historical data. However, when it comes to out-of-sample forecasting, the econometric models perform better, indicating that they can better generalize to future data. The RNN and LSTM model also did not perform better than the naive predictor out of sample. The challenging nature of the dataset, which includes the financial crisis, the COVID-19 pandemic and the invasion of Ukraine by Russia, posed significant difficulties for the RNN and LSTM models in their out-of-sample forecast. These extreme events caused sudden and significant shocks to the economy, resulting in rapid changes in market dynamics that the RNN and LSTM models struggled to adjust to. Furthermore, the RNN and LSTM models may not have been able to capture the complex interactions between economic variables during such unprecedented events. It is worth noting that the study highlights that a more complex model does not necessarily result in better and more accurate performance. This is evident from the performance of the RNN and LSTM models, which, despite their complexity, did not perform as well as the simpler econometric models in out-of-sample forecasting. Despite the poor performance of the neural networks, their inherent ability to capture complex patterns and relationships within data suggests that they hold great potential for future advancements, highlighting the need for continued research and development to unlock their full capabilities. Future research should explore the use of other machine learning models, examine the performance of the models on different data frequencies, inflation predictors, and forecast horizons, and test the performance of the models with different architectures.
Item Type:Essay (Master)
Faculty:BMS: Behavioural, Management and Social Sciences
Programme:Industrial Engineering and Management MSc (60029)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page