Browsing by Author "Cardoso, Edgar"
Now showing 1 - 1 of 1
Results Per Page
Sort Options
- Short-term exchange rate forecasting with machine learningPublication . Cardoso, Edgar; Tran, DanThis master thesis aimed at forecasting exchange rates returns at a weekly and monthly frequency, focusing on the United States Dollar, Euro, Japanese Yen and British Pound. Exchange rate movements are notoriously hard to predict, which leads the literature to consider the random walk without drift as the de-facto benchmark which is consistently hard to outperform. The main goal of the prediction task was motivated by the hypothesis that sovereign risk priced by the market in various instruments would provide information into the future course of exchange rates. Machine learning models (Elastic Net, Random Forest, XGBoost) were built and trained using this data with the goal of producing predictions of the currency movements. Faithful to its role of benchmark, the random walk outperforms our models during the test phase. This test period included events like the COVID pan- demic, the inflationary spikes that followed the reopening of the world economy as well as regional and global political tensions, a period which led to volatility spikes across asset classes and during which our models strongly underperformed. This unsuccessful attempt shines a light once again on the strenuous task of consistently predicting forecast exchange rates, especially at higher frequencies, and raised questions about the limitations of machine learning models when not purposely built with the limitations of the forecasted variable in mind.