Breaking the short-term myth for technical analysis in currency markets

Abstract The profitability of integrating machine learning with technical analysis is assessed through an examination of major floating currencies—specifically, the Australian dollar, Canadian dollar, Swiss franc, Euro, British pound, and Japanese yen—relative to the U.S. dollar, utilizing long-term historical data. Prior research has largely concentrated on short-term forecasts derived from technical indicators, resulting in limited investigation of machine learning-augmented technical analysis across varying time horizons and market conditions. To address this gap, predictive systems employing XGBoost classification models are utilized, incorporating an exceptionally large set of 750 variations of technical indicators across 116 types, a scale rarely observed in previous studies. Predictive performance is evaluated over forecast horizons ranging from 5 to 120 trading days between 2020 and 2023. This study challenges the prevailing assumption in both academic literature and practitioner communities that technical analysis benefits exclusively short-term trading, demonstrating instead that machine learning integration reveals exploitable market inefficiencies at extended horizons—a finding with direct implications for algorithmic trading strategy design and portfolio allocation decisions. The findings indicate that the highest and most stable predictive accuracies, along with superior trading performance measures, are achieved primarily for currencies such as the Euro and the British pound, particularly over the 120-day investment period. At the 120-day forecast horizon, EUR/USD and GBP/USD achieved AUC values of 95% and 98% respectively during the 2022–2023 test period, with corresponding annualized returns of 12.0% and 14.0%, substantially exceeding buy-and-hold benchmark performance.