What do words about probability actually mean?

Every day the political, sporting and other media, along with lots of other Australians, use words about probability – how likely is rain, when’s the best time, perhaps, pretty likely, a dead set certainty. Yet the words are often less a form of prediction or probability assessment but more ways of providing a sense of safety to those who speak them.

Indeed, they are words which are becoming increasingly frequent in media – particularly print media – coverage where many of the employees are fearing for their safety.

Some years ago, writing in Crikey, the journalist and former political staffer Richard Farmer, wrote about an epiphany he experienced one morning when he collected all the newspapers delivered to his front door. They do things differently in Canberra where he lives obviously. In Melbourne, where the blog lives, they get delivered somewhere or other in the garden either under shrubs or up trees.

But nevertheless Farmer’s epiphany was that most of what he saw in the papers was not news. Instead news was stuff he had already picked up in news broadcasts and online. The blog, having worked in journalism and PR back when both were largely focussed on print media, empathised with the epiphany.

Obviously the Farmer insight explains one of the reasons much of the print media is in decline. But it is also a major cause of a fundamental shift in what the print media reports and how it reports it. Some media are trying to address the problem through long reads which give much more detail about issues in the way the work of the recently late and much lamented Evan Whitton did. But mainly (with some notable exceptions such as Ross Gittins or John Silvester) print media ‘news’ is about pundits’ predictions about what might or could happen; lengthy analysis of phrases and situations about which the fundamental facts are already known; or, reflections or tirades based on their proprietor’s prejudices.

In the US the problem has been accentuated by the Trump phenomenon in which tweets can be analysed for instant bulletins interspersed by reports of reactions. The key to effective reporting about Trump, in contrast,  would be through a more fundamental way of producing news – following the money. The Weekend FT produced a fine example in its July 14/15 issue when it devoted two full broadsheet pages to a detailed analysis of the financial history (complete with details of Russian oligarch links) of the Toronto Trump Tower. To their credit Adele Ferguson and a few others have done similar things in Australia about the financial industry.

One of the fundamental problems with the predictive tone of print media is that the language used is imprecise and often misleading. You never read that something is 65% probable or 5% likely and instead you read that something is ‘likely’ or ‘probable.’ The blog has written often about probabilities and predictions referring frequently to the work of Philip Tetlock and Nate Silver. They are not always right but there is a degree of accuracy about the probabilistic nature of their analyses and their reliance on Bayesian techniques. Silver, of course, also has a nice turn of phrase in explaining the odds. When asked how he had got his prediction of the Trump-Clinton election wrong he replied by asking whether anyone would have taken the probability of a Trump win he calculated as comforting odds for playing Russian roulette.

Two authors – Andrew Mauboussin at Twitter and Michael J. Mauboussin, an investment banker – have contributed an article to Harvard Business Review (3 July 2018) which looks at this language imprecision. They quote from much of the research about probabilities and encourage readers to improve their forecasting by consulting Tetlock’s The Good Judgement Open which helps you practise forecasting; other tools such as the Brier score metric; the Metaculus practice question site; and, prediction markets such as Predictit which let you put your money where your mouth is.

But their central argument is about what we mean when we say ‘likely’ and how likely people think it is. To illustrate it they cite intelligence estimates such as a March 1951 CIA estimate that a Soviet attack on Yugoslavia within a year was a “serious possibility”. A professor of history, brought in as a new co-head of the CIA Office of National Estimates, was puzzled by this and thought the phrase meant a chance of attack was about 65%. But when he asked more people in the Office what they thought it meant the estimates ranged from 20% to 80%.

The Mauboussins did some research about how people see probabilistic words such as maybe, possibly, always, certainly, slam dunk, likely, rarely and never and found that “always doesn’t always mean always”. They also found that women tend to interpret “certain probabilistic words or phrases more positively” than men – that is as indicating that something would happen.

The authors suggest three main lessons can improve your understanding of all of this: use probabilities instead of words to avoid misinterpretation; use structured approaches to set probabilities; and, seek feedback to improve your forecasting.

Of course, nobody can ever have a perfect forecasting record, but some people are better than others. But the worse way to go about it is to make predictions simply because you have to talk about the future because the present and the immediate past have already been communicated online.