Section 7 (ENS Products - Dealing with Uncertainty), and particularly Section 7.5 (Broad Guidelines), give suggestions on how to handle ensemble results from a series of forecast model runs.  Especially, it is unwise to follow a trend in forecast values.  This is because a change in evolution, due perhaps to a change in analysed upstream flow, can radically reposition location of features downstream (e.g. the position or development of a trough).  However, model forecasts can show increasing confidence in the range of forecast values as each model run becomes closer to the verifying time.

Fig9.2.1.5-1: Forecast evolution plot for 2-metre temperature around London-Heathrow VT 12UTC 26 May 2026.

Things to note in the temperature plot (Fig9.2.1.5-1):  

• IFS ensemble mean (black diamonds), AIFS ENS control (red circles), AIFS Single (cyan circles) are quite erratic during the period 10 to 16 May.  IFS ensemble mean extrapolation suggests reaching about 26C on 26 May.  IFS control and AIFS Single are particularly variable and no trend can be seen.  It is not until about 20 May that a sharper more steady but still erratic rise can be seen.  AIFS Single shows very high temperatures.  Temperature forecasts are generally more steady after 22 May.
• Box and whisker ranges (IFS blue AIFS ENS brown) are rather large implying low confidence until about 20 May.  Some ensemble members show temperatures below 15C while at the same time the upper extremity of the whisker is broadly about 27C for IFS, 30C for AIFS ENS.  Clearly there is uncertainty.  After 20 May the whiskers, boxes and range become smaller which allows greater confidence.
• No ensemble or single results should be discarded.   High forecast temperatures even from forecasts made as early as 10 May cannot be ruled out and this might be useful early information for heat sensitive customers even though with low probability.  Equally, temperatures as low as 10-15C are possible until about 16 May, even with some support from the lower bound of the boxes, and this possibility might be useful to customers even though this proved to be very incorrect.
• The latest model results cannot be taken as completely the best one without reference to earlier results from earlier model runs.  Confidence can grow as the box and whisker plots become shorter.

(FUG associated with Cy50r1)