Financial markets allocation is a difficult task as the method needs to dramatically change its behavior when facing very rare black swan events like crises that shift market regime. In order to address this challenge, we present a gradient boosting decision trees (GBDT) approach to predict large price drops in equity indexes from a set of 150 technical, fundamental and macroeconomic features. We report an improved accuracy of GBDT over other machine learning (ML) methods on the S&P 500 futures prices. We show that retaining fewer and carefully selected features provides improvements across all ML approaches. We show that this model has a strong predictive power. We train the model from 2000 to 2014, a period where various crises have been observed and use a validation period of 3 years to find hyperparameters. The fitted model timely forecasts the Covid crisis giving us a planning method for early detection of potential future crises.