Evaluating the Impact of Similarity Measures on Transfer Learning in Economic Time Series Analysis

Abstract

This thesis investigated the impact of various similarity metrics on Transfer Learning effectiveness in economic time series analysis. The complexity of economic data presents unique challenges for predictive modeling. Leveraging Transfer Learning can potentially enhance model adaptability by transferring knowledge from one economic scenario to another, thus improving forecasting accuracy and robustness.

The study examined a range of similarity metrics, Dynamic Time Warping, Cross-Correlation, Pearson Correlation, Mutual Information, Shape-Based Distance, and Granger Causality, to determine their influence on Transfer Learning outcomes in economic forecasting. Long Short-Term Memory (LSTM) and Multilayer Perceptron (MLP) networks were implemented to assess the impact of these metrics on model performance.

Experimental results revealed that no single similarity measure uniformly predicts Transfer Learning success across all scenarios. Instead, the effectiveness of each metric largely depends on specific dataset characteristics. While LSTM models initially presented higher base error rates than MLP models, the application of Transfer Learning improved performance, reducing error rates and closely aligned the model architectures.