In the multivariate time series forecasting problem i found out that fine-tuned xgboost (and its variants) performs much better than fbprophet, sarimax, RNN variations. Predicting time series with RNN is like killing a bird for bazooka.