Tim Hewson, Principal Scientist, Forecasts and Services Department
Have you ever browsed ECMWF’s sub-seasonal forecasts and wondered how to make good use of them? If so, this article is for you. The focal point is a newly invented forecast product – called 'quantile-based weekly guidance maps' – introduced in experimental form in June 2025.
The key aim of this new product is to improve upon previous offerings by providing users with very accessible map-format information about two key forecast aspects – the mid-range solution, and the range of possibilities around that, for each of the 5 or 6 one-week periods available.
We cover four key variables, including 2 m temperature and rainfall. Considerable attention was paid to creating a chart design that is both user friendly, from a graphical standpoint, and that provides a direct mapping onto narratives, for anyone needing to forward on forecast information to end users via the written or spoken word.
Historical background
ECMWF has been providing sub-seasonal forecasts for many decades now, but only occasionally have we made modifications to the product offerings. Yet improvements in the skill of our sub-seasonal forecasts mean that we can now say, with some confidence, that if, for example, the forecast is pointing to much warmer than average, and if there is relatively little spread in the ensemble member forecasts, it really is likely to be much warmer than average in the verifying week. The reason for this is that the (statistical) reliability of sub-seasonal forecasts is now much better than it used to be; so here we have our first motivation for these new products.
Enhanced supercomputer power has also facilitated recent structural changes in the sub-seasonal forecast suite that we wanted to take advantage of too. Key changes in IFS model cycles 48r1 (2023) and 49r1 (2024) were:
- An approximate doubling of the frequency of re-forecasts – improving the definition of model climate.
- 100 perturbed ensemble members rather than 50 – so lower probability outcomes have more credibility.
- A sub-seasonal ensemble running once per day rather than twice per week – giving users an earlier indication of a forecast signal.
The new products aim to exploit these advances, ultimately for the benefit of the end user, by plotting together the mid-range solution and spread information.
New products: design and structure
I have named the new products ‘quantile-based guidance maps’. This was with good reason, and can act as a subtle reminder for users! Product content is based on quantiles computed by ranking ensemble member solutions.
Inspiration first came from a pre-existing product: the sub-seasonal meteogram, a segment of which is shown in Figure 1.
Figure 1: Segment of a sub-seasonal forecast meteogram for Rome, showing precipitation (above) and mean sea level pressure (below). Data time 00 UTC 18 August 2025. Y-axis values show weekly mean anomaly, relative to the model climate. X-axis values denote the end of the 7-day forecast period. The red text is an example of the quantities we display on the new product: letters denote distances between lines, expanded on the inset. A = ‘anomaly of the forecast median’ (positive or negative) = colour shading on the new product. B divided by C (both positive) = inter-decile range ratio, or ‘spread metric’ = contours and transparent grey shading on the new product.
The box-and-whisker format (Figure 1) displays a set of forecast quantiles (minimum,10, 25, 50, 75, 90, maximum), whilst the background shading shows the equivalent (climatological) quantiles derived from the re-forecasts. All such values have been normalised to denote, on the y-axis, an anomaly relative to the model climate mean.
The red text in Figure 1 shows the quantities we display on the new maps. To explain further, let’s look at an example of the new product (Figure 2).
The colour shading in Figure 2 displays the anomaly of the forecast median, in units of the variable displayed, to signify the value which has a 50/50 chance of being surpassed in either direction (denoted by A in Figure 1).
The contours in Figure 2 show the forecast spread, relative to the model climatology. For this spread aspect we could have used measures such as standard deviation or inter-quartile range, but instead elected to use the inter-decile range (the separation of the 10% and 90% quantiles) represented as a ratio: forecast range divided by model climate range. This has a clear visual link with the sub-seasonal meteograms (B divided by C in Figure 1) and represents well the more extreme solutions without straying into outlier territory where sampling becomes more of an issue. Contours are coloured thick purple/thin green when the forecast spread is respectively larger/smaller than the climatological spread, whilst transparent grey shading shows where model and climatological spread are similar (B divided by C close to 1).
Figure 2: Example segment of the new product, centred on northern Africa, for 7-day mean 2 m temperature. Data time is 00 UTC Thursday 14 August 2025, valid time is 00 UTC Monday 1 to 24 UTC Sunday 7 September, so about 3 weeks ahead. Colour shading shows the 2 m temperature anomaly (⁰C). Contours show the inter-decile range ratio 'spread metric', with semi-transparent grey shading for values between 0.9 and 1.1. Green contours denote low (relative) spread, purple denote high (relative) spread. Purple contours are usually much less common. Four added boxes show, for the sites Catania (CA), N'Guigmi (NG), Faya-Largeau (FA) and an Atlantic point (AT), the corresponding segments of sub-seasonal meteograms (as on Figure 1, but for 2 m temperature anomaly, with dashed grey lines labelled in ⁰C) for these dates and times.
Examples and usage
Figure 2 shows a forecast for mean 2 m temperature for a period about 3 weeks into the future. The inset boxes are extracts from sub-seasonal meteograms; each illustrates a different type of forecast behaviour.
In Catania (CA) the forecast is very similar to the model climate, so we get grey shading to denote ‘climatological spread’ and white underneath to signify the forecast median has a very small anomaly; in other words, this could be described as a ‘no signal’ forecast.
At the Atlantic point (AT), the anomaly is also about zero, but the green contours signify an inter-decile range ratio that is very small, so we can say quite confidently that temperatures will be very close to normal here.
At both N'Guigmi (NG) and Faya-Largeau (FA), the forecast favours a cold anomaly, but whilst at N'Guigmi the anomaly has a relatively high confidence level, at Faya-Largeau there is unusually large uncertainty in what the anomaly will be.
Typically, on such charts, as lead time advances, we see anomaly magnitudes reduce, so charts become less colourful, and also spread increases, so green contours tend to get replaced by black, with increasingly large areas of transparent grey shading.
Purple contours (denoting a forecast spread larger than the climatological spread) are uncommon, but when present would urge a user to be particularly cautious and to not over-interpret a forecast. Purple contours can appear, for example, as a result of climate-change-related shifts in the cryosphere, or with forecasts of very wet conditions.
Figure 3: Quantile-based guidance map for 7-day precipitation, for a European domain, valid 17–23 August 2025, from data time of 00 UTC 20 July 2025, so looking about 5 weeks ahead. Shading and spread are as on Figure 2, but for precipitation totals, with different scales.
Forecasts for rainfall tend to collapse towards a climatological signal relatively quickly, compared to temperature parameters. However, because the climatological distribution for 7-day rainfall tends to be much less Gaussian than for other parameters, more care is needed when interpreting these new rainfall charts. At longer lead times for example, colour-fill often shows white, or brown shades to denote a median drier than average. This is because, climatologically speaking, you are generally more likely to see below average rainfall than above average. This is useful to note in drought situations! Figure 3 is a case in point. Here the week-5 forecast predicted dry anomalies over the southern half of Europe, but, significantly, that dry signal was reinforced by relatively low spread (green contours). So, this forecast does have more information content than we usually see for rainfall at such long leads.
How do the new quantile-based maps improve on previous sub-seasonal products?
Historically ECMWF products have tended to show anomalies of the mean forecast, or particular quantile values. However, using such charts, it is rather laborious to elucidate the useful information of the type shown by the new products. Now that is of course much quicker, whilst, in addition, the 'click on a chart' facility intrinsic to Open Charts allows users to delve deeper for particular sites, by showing plots like Figure 1 in a pop-up window. Indeed, it can be quite rewarding using this facility to explore strange or unexpected features.
Some of ECMWF’s older products also apply statistical tests, to try to discern when there is a forecast signal. These can be perceived as incomplete: white shading for example can have meanings other than ‘average values expected’. Moreover, these plots can be disconcertingly noisy, and the statistical significance levels quoted are vulnerable to misinterpretation. For example, a user might incorrectly attach the displayed confidence level to the forecast anomaly. So, on the quantile-based guidance maps we elected to not apply significance tests.
Some might argue that forecast quantiles should be compared with a climatological median rather than a climatological mean. However, our approach does have benefits, as in the drought example given above, and is not new. ECMWF’s pre-existing 'anomaly>0' products perform a similar comparison, but display probabilities instead of a dimensioned quantity.
Looking ahead
Current sub-seasonal products, including the quantile-based guidance maps, tend to represent forecasts in ‘anomaly space’. Soon we will introduce maps to show users what the sub-seasonal re-forecast system computes to be the actual climate at different times of year, and at different lead times. We will display the mean (climatological) value, that can be cross-referenced with the shaded anomalies on the new maps, and the climatological inter-decile range, to compare with the contoured spread metric. Both will be dimensioned quantities, having units of the displayed parameter. For example, in early September, Reading would have a mean 2 m temperature of about 16°C, and a climatological inter-decile range of about 5°C.
In the meantime, we would strongly encourage users to explore the new forecast products, and the related documentation, and provide ECMWF with feedback.
Find out more
You can find more information, including how to convert the new product into textual guidance, in our guidance for forecasters.