A carregar...
Publicação
Evaluating time-series momentum against machine learning in commodity futures using multi-horizon and term-structure signals
| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 3.59 MB | Adobe PDF |
Autores
Orientador(es)
Resumo(s)
This thesis replicates the Moskowitz, Ooi, and Pedersen (2012) time-series momentum strategy in commodities, confirming strong pre-2009 performance and weaker results thereafter. Using a walk-forward backtest, it evaluates machine learning predictions based on individual lagged returns and compares linear and nonlinear models. Models using individual lags outperform TSMOM, with linear models performing best, while nonlinearities and interactions add no value. Time-series strategies based on fixed probability thresholds perform poorly, possibly due to miscalibration and limited directional forecasting ability. Adding term-structure and factor features improves Sharpe and accuracy, but gains are largely explained by factor exposure.
Descrição
Palavras-chave
Time-series momentum Commodity futures Machine learning Random forest XGBoost Logistic regression Term structure Basis Basis momentum Skewness SHAP interpretability