This figure compares forecasting performance metrics (R 2 OOS, overall accuracy, and F1) for benchmark models under two regimes: U.S.-only (left bar), and global (right bar) data. Bars of the same color correspond to the same rolling training window size (5, 21, 252, and 512 trading days). Each duplet of adjacent bars
illustrates the performance change attributable to expanding the training data from U.S.-only to Global. Benchmark models include linear (OLS, Lasso, Ridge, Elastic Net, and PCR), ensemble (XGBoost, CatBoost, and LightGBM), and neural network (NN-S and NN-L) models. ‘Overall Acc.’ denotes overall directional accuracy, and ‘F1’ refers to the macro-averaged F1 score. In the middle panel, the horizontal line indicates
the 50% overall accuracy. ‘H’ indicates that the model is estimated using the Huber loss.
This paper presents an empirical study of time series foundation models in global #financial markets, focusing on #forecasting daily excess returns across 94 countries over more than three decades.
Read: spkl.io/63329Ax5eU
Subscribe: spkl.io/63323Ax5et
#FinanceSky