On a notorious fall day in 1987, a British weather forecaster named Michael Fish told viewers not to worry about an incoming storm. The storm hit South East England, and was one of the worst ever recorded there. At least 19 people died. While Fish did in fact forecast strong winds that day, and while he also had a long and successful forecasting career, he is largely remembered though terms like “Michael Fish Moment”, where a forecaster makes an embarrassingly bad prediction
Two decades later, in 2009, the “L’Aquila Earthquake” struck Abruzzo, devastating the region and killing at least 308 people. In the aftermath of the quake, six Italian scientists were tried and convicted of involuntary manslaughter for failing to predict the earthquake based on preceding tremors. Though they were eventually acquitted after a lengthy appeal, the story highlights how precarious the job of forecasting is, and how serious the ramifications of a Michael Fish Moment can be.
History abounds with cautionary examples like these—an important lesson for aspiring forecasters. Forecasting has even been outlawed at various times in the past (cf Hindman & Athanasopoulos 2018). But today, as the COVID-19 pandemic sweeps across the planet, there has perhaps never been such a thirst for forecasts. In fact, grim forecasts played a key role in convincing the US administration to take the pandemic seriously. From elections to sports, people have come to expect complex, information-rich forecasts at their fingertips. With the stakes so high, forecasters are rushing to fill this demand. The CDC has even issued an open call for COVID-19 forecasts through a forecasting challenge. While it’s great to have so many minds working towards this admirable goal, we should be wary of our own Michael Fish Moment.
Enter the ecological forecaster. The quantitative skills ecologists have acquired to model complex ecosystems can also be applied to disease forecasting. There has been a flood of armchair forecasts from the far corners of the internet (often betrayed by their excel-style graphs). Such forecasts are often produced with little to no training in disease dynamics or forecasting but possessing a belief that their expertise at something else (e.g. physics, statistics) makes them qualified to wade into the problem (see Dunning-Kruger effect). Should ecologists lend their expertise to pressing problems like COVID-19? Or should we “stay in our lane”? Just because you can build a forecast model does not always mean you should. To be able to answer this question it is critical to explore ethical considerations about how to create, communicate, and interpret forecasts. Recently, Hobday et al. (2019) outlined some of the ethical considerations involved in ecological forecasting. A few of these, highlighted here, are particularly relevant to the rush to produce COVID-19 forecasts, and should be considered before entering the fray.
Conflicts of interest – Forecasts are information, and information can be powerful. As Hobday et al. write, “a forecast advantage for one group may be a disadvantage for another”. In ecological forecasting, this case can often arise for harvested or protected species. For example, forecasts of certain protected species, such as Atlantic sturgeon, might be used for illegal targeting of that species. Even for legal harvest, accurate forecasts can influence supply and demand dynamics, benefitting certain members of the supply chain to the expense of others. There are implicit value judgements in the way a forecast is designed, the choices of what to forecast, and what information to use.
For COVID-19, conflicts are even more pronounced, as everyone is a stakeholder. Resources are limited, from toilet paper* to PPE, and supply and demand dynamics are at play here as well. Even if a forecast is accurate, and even if a corresponding redistribution of PPE were to benefit many people, other groups of people would be at the losing end of the deal. Regardless of how accurate a forecast is, its output is likely to influence decisions, potentially favoring one group of people over another, or one objective over another. At the outset of a forecasting project, one should take some time to consider who might be negatively impacted by a forecast and why.
Uncertainty – For a decision-maker to be able to use a forecast, it must contain information about uncertainty, otherwise it is impossible to know how much trust to put in the forecast. Should a line without a confidence interval be interpreted as exactly what’s going to happen, or something that could easily be off by +/- 500%? In decision science, uncertainty translates to risk, and without knowing the uncertainties involved, we are liable to make decisions that are either overly risky or excessively cautious. Forecast uncertainty also tells us how far into the future we can trust the forecast. Intuitively, forecasts should become less confident as they project farther into the future, and forecast uncertainty estimates that don’t show this pattern should send up warning flags Most reputable COVID forecasts do include error bars that say something about the likely range of predicted outcomes, such as the number of expected cases or deaths over time. Clearly communicating what those error bars represent can be tricky. Is the concept of the confidence interval explained in a way that a politician or lay person can understand?
A more challenging problem is that error bars seldom tell the whole story about uncertainty. They do not say anything about sources of uncertainty not included in the model. For example, the widely used IHME projection (referred to as a “forecast” in the methods paper) does show an uncertainty range, but the initial model also had COVID-19 deaths going to 0 ±0 in mid-summer for every state. While they were following the good practice of including a range of uncertainty in their plots, the ±0 should raise a red flag. This artifact comes from the choice of statistical model used to fit the data**–essentially structural uncertainty. Explaining how these sources of uncertainty might impact the forecast is even harder. We have seen model assumptions listed in the fine print, where most consumers of the forecast are unlikely to read them, let alone understand their implications. A lot of responsibility falls on the forecaster to carefully communicate how uncertainty should be interpreted for a particular forecast product, and if a component of a forecast could be easily misinterpreted, perhaps that component should be excluded.
One way to address critical assumptions and structural uncertainty is by focusing on forecasts from multiple models–”ensembles” in the jargon (here is a nice example for COVID-19). As long as the different models represent a range of assumptions, approaches, or training data, the variation in the forecasts they produce can help communicate the magnitude of uncertainty from these sources. We should not expect every forecaster to produce an ensemble, or even collect and present forecasts from multiple models, but perhaps it is not too much to ask forecasters to discuss how their model relates to the rest of the pack. Is their forecast more or less optimistic than others? Why?
Unintended Consequences – On November 3, 1948, the Chicago Daily Tribune posted the now infamous headline, “Dewey Defeats Truman”.*** On one level, this could be thought of as a Michael Fish Moment–a public prediction that turned out to be profoundly mistaken. But many historians go a step further, and attribute Truman’s victory in part to a response to the forecasts themselves. For weeks leading up to the election, forecasts based on polls broadcasted confidence in a Dewey victory, and these forecasts likely motivated the Democrats to turn out in large numbers. Under this interpretation, the forecasts themselves influenced the outcome.
When human behavior is involved in a prediction, a weird circularity occurs. If we predict high COVID-19 mortality rates, and people respond by quarantining, thereby reducing mortality, then was the prediction wrong? Or did it serve its purpose? Forecasters often distinguish between a prediction and a projection in cases like this. A projection is when a certain scenario is assumed–usually an assumption about human behavior. Climate projections are a familiar example, where each scenario assumes a certain greenhouse gas emissions level. This approach is also common in making market projections, where humans are an integral part of the system. Most forecasters want their forecasts to be useful, to have an impact, or to affect human behavior. But these effects are often the hardest things to predict.
Sins of omission vs. commission
The Michael Fish Moment could be seen as a sin of commission: an inaccurate forecast that caused harm. The L’Aquila Earthquake case could be seen as a sin of omission: forecasters failing to provide any information.**** Forecasters must balance these opposing risks. Rushing a poorly calibrated and vetted forecast out the door may be a recipe for a sin of commission. Fiddling forever to perfect a model can lead to a sin of omission. What poses a greater risk, the release of flawed information or the absence of information? The answer will vary case by case, but relevant considerations might include urgency (how soon must a decision be made?), the shape of the “cost curve” (does under/overestimating risk cause similar harm, or are the costs asymmetric?), and the audience for the forecast (we might be more comfortable passing incomplete or flawed information to a sophisticated audience with experience interpreting forecasts). Clear communication of uncertainty, and how uncertainty translates to risk, will be especially important for novel or risky forecasts.
So how do we avoid our own Michael Fish Moments? As the pandemic unfolds, there is a wide range of forecasts being developed and used. The US Center for Disease Control is currently using at least 15 different forecast models. Nicholas Reich, a biostatician who works on an ensemble projection using these models, described it this way: “We’ve been sort of building the car as we’re driving it at 90 miles an hour down the highway. And we’re learning as we go.”
For those interested in joining the forecasting game, the Hobday paper is a good starting point. The authors go on to describe a number of other ethical considerations, ranging from managing expectations to user equity. As ecological forecasters, if we are sitting there with a warm cup of coffee (or other beverage), our data laid out before us, some solid scripts, and a fancy live-update web app–all adding up to a top-notch forecasting product–we need to do our due diligence and step through the ethical considerations before unleashing the forecast on the world.
* There is plenty of toilet paper. Right?
** To be fair, the IHME projections were designed to estimate peaks and peak hospital resource use, not to estimate the duration of the outbreak. But, to be even fairer, they were still broadcasting the outbreak-duration component of the projection, and people were basing decisions on these projections.
*** Truman won, at least in this realized universe.
****In both of these cases, the forecasters were likely providing the best information they had.
Hobday, A.J., Hartog, J.R., Manderson, J.P., Mills, K.E., Oliver, M.J., Pershing, A.J. and Siedlecki, S., 2019. Ethical considerations and unanticipated consequences associated with ecological forecasting for marine resources. ICES Journal of Marine Science, 76(5), pp.1244-1256.
Hyndman, R.J. and Athanasopoulos, G., 2018. Forecasting: principles and practice. OTexts.