After last year's wonderfully mild April, Germans expected to face a sizzling-hot summer. But German food retailer REWE bet money it knew better.

Already in April, REWE, which operates a number of well-known German retail chains, unloaded onto the market a product that wasn't exactly part of its ordinary range: around three gigawatt hours of electricity, or enough to meet the annual energy needs of about 700 households. To put it another way, REWE sold its options to buy the energy packet, which it was entitled to exercise in June and July. At the time, many electricity dealers thought it was a strange move.

Electricity -- like pork bellies, telephone minutes and other commodities -- is traded on exchanges. REWE has a subsidiary called EHA set up specifically for this purpose, and REWE buys electricity for its own needs through EHA. As a company that uses about as much electricity as the large German city of Hanover, which has half a million inhabitants, it sometime buys the commodity years in advance. Large amounts of electricity are needed to power REWE's refrigeration rooms, where perishable products are stored, as well as its freezers and refrigerated shelving in supermarkets.

Every summer, the demand for electricity to keep food fresh rises, especially during hot weather. When that happens, prices go up on the exchanges. Astute users stock up on electricity when the prices are low; buying electricity when summer temperatures are already high can get very costly.

So, given all of that, why on earth was REWE unloading its electricity options in the warmest April on record?

The Supercomputer Trump Card

The rest is now history. Summer 2007 was unusually cold, and electricity prices dropped as a result. REWE, for its part, raked in a tidy additional profit. As EHA spokesman Arnd Kiesselbach puts it: "We sold our surplus when people were still willing to pay good money for it."

The scouting for the crafty deal took place in the British city of Reading, where a supercomputer constantly runs weather simulations months in advance. According to the computer's calculations, even by as early as April, things weren't looking too good for summer 2007.

The computer is part of the European Center for Medium-Range Weather Forecasts (ECMWF). For the last few years, researchers there have been trying to do something that the weather's inherently random nature would normally preclude: forecast entire seasons. The results have become so reliable that Germany's National Meteorological Service (DWD), a member organization of ECMWF, now officially offers them for sale. For €150 ($231), companies can now purchase the center's six-month forecast.

For example, according to the center, this summer will be a little warmer than usual in Germany, although rain levels will remain about average. The reports from Reading are oracle-like in their simplicity, merely pointing out broad tendencies -- warmer, colder, wetter or dryer than average -- and for good reason. More information cannot be gleaned this far in advance. The forecasts only reveal the extent to which actual values could diverge from the averages of many years.

The computer still makes significant mistakes, and the information it provides is too vague for purposes like planning vacations. But the data can be of use to business, and the DWD is already advising power companies and municipal utilities. If they know, for example, that the upcoming winter is likely to be colder than usual, these companies can stockpile natural gas reserves for their power plants early on.

The number of economic sectors sensitive to the weather is surprisingly large, from ice cream makers to the textile industry. In the latter case, knowing whether summer temperatures are more likely to create a need for bikinis or windbreakers can be critical.

The DWD's list of customers even includes a manufacturer of mosquito repellant, because mosquitoes are also subject to the weather. The dominant mosquito strains in cool, wet summers are different from those in warm summers, and combating each strain requires a slightly different cocktail of chemicals.

Cloning Chaos

The DWD receives its forecasts from Reading, where the supercomputer is fast enough to digest millions of pieces of observation data. The weather model the computer uses is similar to the one employed for daily forecasts. But a single model would eventually become useless and would only spit out nonsense after a week or two. The meteorologists use a trick to overcome this hurdle: They clone their model and then run simultaneous simulations with multiple clones.

The cast of clones currently consists of 41 members, and each one begins with the weather data for the present. However, because current conditions can never be captured perfectly, each model receives slightly modified initial values. Then the computer simulates an imaginary worldwide weather process. Pressure zones move around the globe hour by hour, snow falls and melts again, and monsoon clouds gather over India at the customary time -- everything is simulated almost exactly as it would be on the real planet.

Of course, the models only coincide at the beginning. They are often noticeably different by the third day, and after three weeks each model is computing away in its own, unique world. In Model 5, all of Germany may be inundated with nonstop rain in June, while Model 12 is predicting the hottest June in years. Meanwhile, Model 23 may have concocted a moody low pressure zone over the British Isles.

Things just get more and more complicated. The meteorologists could watch the models and lose themselves in an endless array of possibilities, but no one is interested in that. "We don't even look at individual calculations," says meteorologist Renate Hagedorn, who works on seasonal prognoses for the DWD in Reading.

Finding Order in Chaos

As a rule, none of the weather models makes even vaguely accurate predictions. But this changes once the scientists consider the overall ensemble. On average, this organized chaos produces quite reasonable data. Of course, all day-to-day fluctuations have to be erased first. Any detail would only pretend to be an accuracy that cannot exist. After the outliers have been removed, all that remains are the monthly averages.

The measure of variance, though, is also always known. The farther apart the 41 models are, the less reliable the average becomes. "That's the real breakthrough," says Hagedorn. What it does is enable the researchers to determine, rather accurately, just how inaccurate the forecast is.

For companies, this information can actually be very useful. For example, for the operator of a wind farm, it is really advantageous to know that one-fifth of the models are predicting such powerful storms that he will have to shut down his wind turbines. As soon as the residual risk is quantifiable, he can include it in his calculations.

However, sometimes the uncertainty is so great that guessing becomes just as effective. If the models happen to be correct sometimes, this is because of the skillful pre-selection of data. The computer pays little attention to the temperamental atmosphere, which -- as we all know -- changes its mind on a daily basis. Instead, it makes its calculations using mostly variables that only change slowly, such as the global dispersal of snow and ice, soil moisture and ocean temperatures.

The oceans, in particular, have a lasting effect on complex weather patterns. As a temperature sink, they essentially act as the long-term memory of weather patterns. As a rule, a high pressure system in the atmosphere disappears after a few days, almost as if it had never existed. But excessive warming of the western Pacific Ocean is something that the weather doesn't forget for months.

The researchers in Reading have just discovered that additional factors should be taken into account, such as the carbon cycle and the greenhouse gas effect associated with it or the distribution of suspended particles in the atmosphere that filter sunlight. Until now, these processes have only been simulated in climate models that compute the course of several centuries. But, as scientists are learning, they also play a role in shaping the current weather. "That's why we are trying to incorporate them into the forecasts," says Hagedorn. "Besides, climate and weather models are becoming more and more alike."

Fine-Tuning the Forecast

The ECMWF could also have used a more fine-tuned climate simulation to develop its seasonal forecasts. "But weather models happen to be our great strength," says Hagedorn. In fact, the center in Reading is the leader in the exclusive world league of major weather forecasting centers. In global performance comparisons, the European model usually comes out on top.

Nevertheless, the supercomputer in Reading also has its limits. The weather model used to develop the daily short-term weather forecast divides the atmosphere into a grid network consisting of individual cells 25 kilometers (15.5 miles) wide. But to perform the half-year forecast, the grid had to be modified to make the cells 125 kilometers (77.7 miles) wide, so that the computer can handle 41 models simultaneously. However, this is still more precise than climate simulation, which often operates using cells 200 kilometers (124.3 miles) wide. But satisfactory it is not.