The challenge of seasonal weather prediction

By Hannah Nissan, Research Assistant in Regional Climate Modelling, Physics

In April 2009 the UK Met Office issued their now infamous forecast: “odds-on for a BBQ summer”. By the end of August, total precipitation since June had climbed to 42% above average levels for 1971-2000 (UKMO, 2014). Why is it so challenging to provide seasonal forecasts several months ahead?

A question which arises often in conversations about climate change is “how can we predict the climate in 50 years when we can’t even get the weather right next week?” While we have no skill in forecasting the weather on a particular day in 2060, we can make much more confident projections about the average conditions at that time. Some explanation of this comes from mathematician and meteorologist Ed Lorenz, who made an accidental discovery while rerunning some calculations inspired by weather predictions on his computer (Lorenz, 1963). Taking a seemingly harmless shortcut, he rounded a number to a couple of decimal places, and the result was a completely different weather forecast. His finding was coined the “butterfly effect”, which invokes a powerful analogy: tomorrow’s weather forecast depends so strongly on getting today’s weather right (a significant challenge) that forgetting to account for the effect of a butterfly’s flapping wings is enough to derail it. In contrast, climate is an average of the weather over a few decades, and doesn’t suffer as strongly from this debilitating “initial conditions” problem.

The distinction becomes rather blurred when we turn to seasonal prediction. When looking at longer term forecasts, say for this coming summer or next winter, correctly characterising the initial conditions of parts of the climate system still matters. This is particularly true for slower varying fields like soil moisture and the upper ocean levels, and to a lesser extent for the atmosphere. However, on these timescales other challenges become increasingly important. As we move from forecasting days ahead, to weeks, months, seasons and decades, the number and complexity of physical processes that must be well described by the computer models used to simulate the weather increases. Delivering better forecasts requires improvements on both these fronts, which compete for limited computer resources.

The benefits of developing more sophisticated prediction models must be balanced against the critical need for an indication of the uncertainty related to a forecast. To do this meteorologists create not just one weather forecast, but a whole ensemble of predictions starting from slightly different initial conditions. Other ensembles to capture the uncertainty in the physics of the model itself are also simulated. The spread of the ensembles tell us something about the range of possible outcomes and their likelihoods. We can never have perfect knowledge of the current state of the weather, nor a perfect model, so running lots of ensembles to develop a probabilistic weather forecast in this way is really important.

To illustrate the trade-offs involved between running ensembles and improving model complexity, consider an example from the UK Met Office. The wetness and storminess of UK winters are strongly linked with the North Atlantic Oscillation (NAO), a pattern of sea surface pressure between Iceland and the Azores islands. By investing in research to uncover some of the important physical drivers for the NAO and including these in their model, the Met Office have recently achieved a significant improvement in NAO forecasts (Scaife, 2014a,b). At the European Centre for Medium Range Weather Forecasting (ECMWF), improving the way convection is modelled delivered better tropical precipitation forecasts and greater skill in predicting seasonal European weather (Bechtold, 2013).

Predictability itself is not a fixed quantity, but varies with each situation. Some weather systems are strongly influenced by larger scale phenomena happening over longer periods of time for which forecasts can be quite reliable. For example, weather in much of the tropics depends on the El Nino Southern Oscillation (ENSO), a pattern of sea surface temperature and atmospheric pressure in the tropical Pacific which persists for several months. Others may be dominated by brief, local processes that are much harder to predict, like convection.  In general, models are able to simulate the cascade of energy downwards, from large movements like the jet stream down to small waves and turbulence. The reverse case is not so easy: it is a major challenge to represent the effects of processes occurring at small physical scales and short time periods, like turbulence and convection, on the wider weather. Consistent with the butterfly effect, some of the effects of small-scale processes are inherently unpredictable and must be represented by random noise.

A good test for the usefulness of a seasonal forecast is whether it offers an improvement over simply looking at the average conditions. In other words, can we do a better job if we simply say that summer temperatures will be the same this year as they have been on average over the last 30 years? Weather prediction models beat statistical forecasts in the tropics, where the influence of ENSO is strong and fairly predictable. This has not in general been the case over Europe and in other higher latitude regions, where lots of different phenomena interact (Buizza, 2014).  However, the latest forecast systems are starting to show some skill even here (Scaife, 2014b).

Temperature forecasts several months ahead are often better than looking at long term data. Predictive skill for precipitation, however, is much worse. This is because rainfall is driven partly by local processes, right down to how individual raindrops are formed. Temperature on the other hand, tends to be controlled by larger, more predictable features (Buizza, 2014). That said, the disastrous floods in Pakistan in 2012 were well forecast a week ahead by ECMWF because, in that particular situation, the rainfall was controlled by large air movements that were relatively well understood (Hoskins, 2012).

The challenging reality is that predictability varies from case to case according to the physical factors controlling each phenomenon. Extracting predictive information across both space and time scales can allow us to unpick these convoluted problems and make real improvements in seasonal prediction (Hoskins, 2012).

With thanks to Brian Hoskins for his helpful review comments.


Lorenz, 1963. Deterministic non-periodic flow. JAS 20:130-141.

UKMO, 2014. Summer 2009. Accessed 20/03/2014.

Bechtold, P., N. Semane, P. Lopez, J.-P. Chaboureau, A. Beljaars, N. Bormann, 2013: Representing equilibrium and non-equilibrium convection in large-scale models. ECMWF RD TM 705, available at

Buizza, R., 2014. Coupled prediction: opportunities and challenges. Seminar, Imperial College London, 18th March.

Scaife, A., 2014a. Forecasting European winters: present capability and potential improvements. Willis Research Network Seminar: Forecasting, Flood & Fortitude: An Afternoon with Willis, 18th March.

Scaife, A., 2014b. Skilful long range prediction of European and North American winters. GRL, in press.

Hoskins, B., 2012. The potential for skill across the range of seamless weather-climate prediction problem: a stimulus for our science. QJRMS 139(672):573-584.

[1] Convection is when local heating causes air to rise

Leave a Reply

Your email address will not be published. Required fields are marked *