Forecasting several steps ahead can follow different strategies: recursive (one model reusing its own forecasts, with error accumulation), direct (one model per horizon), or multi-output (one model producing the whole horizon). The choice trades off accuracy, computational cost, and coherence of the forecasts.