Boosting usually outperforms a random-forest on clean tabular problems, which is why it dominates data-science competitions. That success does not transfer automatically: competition datasets have strong signal and independent rows, and financial datasets have neither.
It needs careful control. Low learning rate, shallow trees, strong subsampling, early stopping on a properly purged validation fold, and a hard cap on the number of features. Without these it will find whatever idiosyncrasies exist in your particular decade of history.
Sensible use: as a comparison against a regularised linear model on the same features. If boosting adds only a little, take the linear model, because you will be able to understand why it fails when it does.
Related: random-forest, regularisation, hyperparameter, overfitting