Saturday, November 12, 2016

A Discussion of the Young Science of Seasonal Forecasting

Every fall, clients expect private sector meteorologists to make forecasts for the upcoming winter season. These forecasts are intended to provide useful information about the average outcome over the season, such as whether it is likely to be warmer or colder than normal. Such forecasts can be difficult to make, and many forecasters make predictions based on tenuous claims.

Seasonal forecasting is complicated because the average weather in the upcoming season depends both on signals evolving on seasonal or longer timescales, such as El Niño and climate change, and on signals that were originally acting on short timescales but that cause longer lasting changes in the state of the atmosphere or ocean. Phenomena like El Niño provide a kind of anchor pattern to which the weather can often return throughout the season, occasionally allowing skillful predictions to be made on a seasonal level.

In contrast, relatively brief weather events like major midlatitude storms, hurricanes, or Madden-Julian Oscillation (MJO) events can displace warm air from its normal locations to higher latitudes, leading to blocking patterns that can sometimes hold on for weeks at a time, influencing weather around the world. Wind stress associated with weather events can also give the ocean a "kick", thereby moving warm or cold water around. Such "kicker" events can make conditions after them different from conditions if they had not occurred (such as by amplifying or breaking down El Niño). They can move the climate into a new state. Most such transitional events are not usually predictable more than a couple of weeks in advance. Events that can alter the subsequent mean state of the seasonal climate are characteristic of nonlinear or chaotic systems. The best we can hope for in predicting such events beyond part of the next month is to correctly predict whether such events are more or less likely under certain background conditions.


Seasonal forecasters use both global climate models and statistical methods in their quest. Climate models are usually not good at predicting weather event-driven changes in the mean state, but they are good at maintaining the present state. Statistical methods, including data mining in large climate datasets, allow us to find relationships between different signals across the climate system. For example, strong El Niño conditions might be correlated with reduced snow cover in Canada, which then affects the temperature in Canada and the United States. Correlated signals like Canadian snow cover extent and eastern United States temperature can sometimes be applied to make predictions. When the ground is covered with snow, sunlight gets reflected back into space. Snow also creates cold air by emitting longwave radiation into space.

With many interacting variables, however, it can be difficult to determine cause and effect. Usually real weather events respond to multiple interacting causes, some contributing more than others. Snow cover may be reduced because the background pattern favors import of warm air over the continent, which can sometimes continue even after a snow event. If an assumed cause (such as extensive snow cover) is present, but another major factor such as strong southerly wind is also present, using the snow cover alone to predict upcoming temperature patterns would lead to a bad forecast. During strong El Niño events, a rogue storm event might increase Canadian snow cover, contributing toward lower temperatures, but import of warm air characteristic of El Niño can overcome these cooling effects and even melt the snow. Without taking the likelihood of import of warm air into account, a forecaster might suggest a long-term shift in outcomes across the season due to increased snow cover when one might be unlikely.

It is easy and occasionally useful to blame a warm or cold weather event on climate change, El Niño, abnormal Arctic Ice cover, snow extent in Siberia, the distribution of rainfall in the tropics, the North Atlantic Oscillation, sudden stratospheric warmings, tropical cyclone recurvature, the Pacific decadal oscillation, the Madden Julian Oscillation, or many other possible mechanisms. Many forecasters have their favorite indicator. My favorite is organized rainfall patterns in the tropics. I emphasize this indicator not because I think that it causes every major outcome in the atmosphere, but because it constitutes a well-defined energy source that leads to stationary or propagating waves that communicate outcomes to different parts of the world. Tropical rainfall also includes many signals that tend to favor certain sizes, propagation speeds, and lifetimes, implying some level of consistency from event to event. Yet, although rainfall in the tropics does drive changes in patterns in atmospheric circulation, it also responds to that circulation, complicating forecasting because the quantity often labeled as the "effect" can lead to changes in the thing we labeled as the "cause".

Another scientist might prefer to track snow cover in Siberia as a favored indicator in a seasonal temperature forecast. When cold air is available in Siberia, it can be displaced across the frozen Arctic Ocean into Canada, leading to an available source of cold air for the eastern United States. Yet, the presence of snow cover in Siberia does not necessarily imply that cold air will ultimately reach the eastern United States for any extended period of time. The winter of 2015-2016 was an excellent example of an exception (at least considering the season as a whole). Snowcover accumulated early in Siberia, but that winter was the mildest on record across much of the eastern United States. That winter, the structure of the El Niño response pattern favored both extensive snow cover in Siberia in October and warm outcomes through most of the winter in eastern North America.

Many seasonal forecasters fall into the trap of identifying a tracer signal of future outcomes, while not understanding the different pathways whereby different outcomes can occur. Scientists like simple explanations. This type of reductionist thinking often leads to good understanding, even in climate science. In a linear system, the total is just the sum of the parts. On the other hand, in a nonlinear system like the atmosphere, reductionism can occasionally lead to large errors. Brief weather events can sometimes kick off long lasting changes. Interacting signals in the climate system can lead to outcomes that are different from the simple sum of the signals.

Bottom line: When considering a seasonal forecast, think about the arguments being made, the level of confidence in those arguments, and alternate pathways to different solutions. Skill of even the best forecast techniques is relatively poor. In the end, no matter how well informed your seasonal forecast might be, anticipate that occasionally, weather events will alter the signal over the course of the season, leading to seasonal averages that are different from what you anticipated. That outcome does not necessarily mean that the forecast was not based on good information--it might just mean that unpredictable signals in the climate system led to a different outcome.