Predicting a rare event, say a 5% single-day drop occurring on 1% of days, a model that always says no is 99% accurate and completely useless. Accuracy is the wrong metric; precision, recall and the precision-recall curve are the right ones.
Standard remedies are class weights, resampling the minority class, or changing the decision threshold after fitting. Threshold adjustment is usually the cleanest, because it leaves the fitted probabilities intact and makes the cost trade-off explicit.
In trading, the asymmetry of cost matters more than the asymmetry of counts. Missing a crash costs far more than a false alarm that keeps you out for a day, so choose the threshold from the payoff table, not from a default of 0.5.
Related: label, supervised-learning, objective-function, type-i-error