Exponential smoothing uses all time series values to generate forecasts, with lesser weights given to older observations. It calculates a smoothed level (Lt) at each period (t) as a weighted average of the current value (yt) and the previous smoothed level (Lt-1). This smoothed level then becomes the forecast for the next period. The smoothing constant (α) determines the weights, with lower α producing a "flatter" smoothed series that changes less over time.