Lag-Llama Forex Forecasting Thesis

Author: Nikoletta Protopapa Master’s Thesis – Cyprus University of Technology (CUT)

This repository contains the implementation of my Master’s thesis project, which focuses on time-series forecasting in the Forex market using two different model families:

Lag-Llama: a transformer-based foundation model for probabilistic forecasting LSTM: a classical baseline neural network

The project includes all stages of the pipeline: data preprocessing, exploratory analysis, model inference, and evaluation across multiple currency pairs.

Project Overview

The purpose of this thesis is to evaluate whether modern transformer-based architectures (specifically Lag-Llama) can outperform traditional deep learning models (LSTM) in predicting daily exchange rates for several EUR-based currency pairs, such as EURUSD, EURAUD, EURCAD, EURGBP, EURJPY, EURCHF, and EURCNY.

The project workflow includes:

->Downloading and cleaning daily Forex data ->Preprocessing and filtering each currency pair ->Exploratory Data Analysis (EDA) ->One-step-ahead forecasting ->Multi-pair inference ->Evaluation and performance comparison between models