A new approach to extended‐range multimodel forecasting: Sequential learning algorithms