The surprising power of simple predictions - FT中文网
登录×
电子邮件/用户名
密码
记住我
请输入邮箱和密码进行绑定操作:
请输入手机号码,通过短信验证(目前仅支持中国大陆地区的手机号):
请您阅读我们的用户注册协议隐私权保护政策,点击下方按钮即视为您接受。
卧底经济学家

The surprising power of simple predictions

Experts can be sidetracked by elaborate ideas, overlooking the direct approach
00:00

{"text":[[{"start":5.8,"text":"For an academic researcher who first trained as a philosopher, then as a psychologist, Robyn Dawes was a practical fellow. He would tell a story from his time working in a psychiatric ward in the late 1950s. “There was a client who had this delusion, and the delusion was that he was growing breasts.” The man was locked in a secure ward while the psychiatrists pondered the reason for this fascinating delusion; they suspected that it was most likely the traumatic impact of the recent and shocking death of a parent. "}],[{"start":36.449999999999996,"text":"Six weeks later, someone asked the man to take off his shirt. He had a genetic condition, not a delusion. “It was in fact true: he was growing breasts.”"}],[{"start":46.5,"text":"It was a lesson to learn: even experts — no, especially experts — can become sidetracked by elaborate ideas, overlooking the simple and direct approach. No surprise, then, that Dawes would become fascinated by the research of Ted Sarbin and Paul Meehl, psychologists who had studied the surprising power of simple statistical predictions in areas such as clinical diagnoses or academic performance. "}],[{"start":71.7,"text":"Sarbin, for example, used a linear regression — almost the simplest statistical rule imaginable — to predict the college grades (GPA) of graduating students based on their high-school class rank and their score on the entrance test. That method was more accurate than the opinion of clinical psychologists armed with the same data and vastly more besides. Meehl found many more examples of cases where simple statistical rules beat the diagnosis or forecasts of the experts."}],[{"start":100.35,"text":"But just how simple could those rules be? A standard linear regression predicts an output based on a combination of various inputs. For example, the probability that an offender would be rearrested might be a function of their age, sex, number of previous convictions and severity of previous convictions. How much weight each factor gets is determined by a mathematical formula to most closely fit the historical data. "}],[{"start":126.39999999999999,"text":"Instead, Dawes suggested what he called “improper” linear regression, where the weights weren’t optimised, but chosen arbitrarily — perhaps equally weighted, or even chosen at random."}],[{"start":138.29999999999998,"text":"To pick an example suitable for the Financial Times, think of the choice of an optimal investment portfolio. The output is the portfolio return; with the right allocation across different assets, we could maximise the predicted return for any given level of risk. Harry Markowitz, who shared a Nobel memorial prize in 1990, showed in the 1950s how to choose the weights in such a perfect portfolio. The Robyn Dawes school of thought says not to bother. Instead, follow a rule such as “Just invest your money equally in the 50 largest public companies” — or maybe even “half in stocks, half in bonds”."}],[{"start":178.54999999999998,"text":"That can’t possibly work, can it? Well, in the first job Markowitz took after publishing his theory, he had to decide how to allocate his pension contributions. He plumped for half in stocks, half in bonds. That’s an improper weight for you. But it’s not clear that Markowitz was wrong, even if he was self-contradictory: a 2009 paper by Victor DeMiguel, Lorenzo Garlappi and Raman Uppal found that the simple strategy of investing equally across a bunch of assets is surprisingly effective."}],[{"start":209.89999999999998,"text":"Over drinks after an academic conference discussion, a fellow panellist challenged Dawes: “Could you . . . use one of your improper linear models to predict how well my wife and I get along together?”"}],[{"start":221.79999999999998,"text":"Dawes thought he could. He had colleagues who had been gathering data on sex and relationships, and he proposed the following improperly weighted predictor: that couples would probably describe their relationship as “happy” if they had sex more often than they had fights, and “unhappy” if the frequency of fights exceeded the frequency of sex. "}],[{"start":241.49999999999997,"text":"Two variables, equally weighted — surely there’s a more accurate model than that? Yet the absurdly simple theory fit the evidence. A colleague had data on 12 unhappy couples; all of them fought more often than they had sex. Of 30 happy couples, 28 had sex more often than they had arguments. Subsequent small studies reached the same conclusion. "}],[{"start":264.15,"text":"(An important caveat: ask the couples about the quality of the relationship first, and the quantity of sex and arguments afterwards, otherwise the totting-up may provoke a self-fulfilling crisis. One woman counted the sex, and the fights, and decided it was time to file for divorce.)"}],[{"start":283.25,"text":"“The conclusion is that if we love more than we hate, we are happy; if we hate more than we love, we are miserable,” wrote Dawes in a 1979 article, “The Robust Beauty of Improper Linear Models in Decision Making”, adding: “This conclusion is not very profound, psychologically or statistically. The point is that this very crude improper linear model predicts a very important variable.”"}],[{"start":309.6,"text":"Why do these almost laughably simple models work? One answer is that while the weights are arbitrary, there is already some expertise smuggled into the choice of variables to throw into the mix. Dawes might have claimed that marital happiness was a function of average monthly rainfall in Nigeria; another simple model but not a very good one. "}],[{"start":328.1,"text":"Another answer is that complicated-seeming outcomes often reflect fairly straightforward combinations of variables. It’s almost always a bad sign if an offender has a string of previous convictions, no matter what else might be true. And regardless of the psychodramas surrounding a couple’s relationship, it’s probably a good sign if they’re having plenty of sex."}],[{"start":349,"text":"But a third factor is that the typical dataset only captures a slice of what’s really going on. Marital happiness is hard to measure precisely. Risk is hard to measure precisely. Even the frequency of sex is harder to measure than it might seem — who’s counting, and what do they think counts? Alongside all this noise is the fact that everything changes."}],[{"start":370.65,"text":"As a result, the optimal-seeming estimate may prove overconfident as time goes by and more data arrives. A simpler, cruder method may be a bit more robust. Victor DeMiguel and his colleagues reckoned that in order for the allegedly optimal estimates to reliably outperform the simple equal-shares rule for a 50-asset portfolio, the analyst would need a dataset five centuries long."}],[{"start":395.15,"text":"The point is not that simple — “improper” — analysis is always best, just that it’s a surprisingly strong baseline. It doesn’t mess about or claim too much. It can be performed on a napkin in a bar, or jotted on a doctor’s notepad. Before unrolling a great analytical edifice, sometimes it’s worth asking to check under the shirt."}],[{"start":418.29999999999995,"text":"Find out about our latest stories first — follow FT Weekend Magazine on X and FT Weekend on Instagram"}],[{"start":432.49999999999994,"text":""}]],"url":"https://audio.ftcn.net.cn/album/a_1780051985_9355.mp3"}

版权声明:本文版权归FT中文网所有,未经允许任何单位或个人不得转载,复制或以任何其他方式使用本文全部或部分,侵权必究。

太空探索技术公司IPO大胆踏入AI经济学的未知领域

对于前沿模型最基本的商业模式,我们仍有许多有待了解之处。

教宗搅动硅谷

不同于美国总统,这位教宗选择直面AI带来的严峻挑战。

科学家发现:鸟类凭“直觉”导航

研究发现,肝脏中富含铁的免疫细胞像一个“内置罗盘”,帮助鸽子感知地球磁场。

法拉奇正面临来自右翼的威胁

“复兴英国”的崛起或将迫使改革党领导人调整策略。

如果问题根源在远程办公而非AI,这对初级岗位招聘意味着什么?

最新证据显示,居家办公的兴起让企业对聘用初级员工的意愿下降。

被罢免的英国石油董事长阿尔伯特•马尼福德在遭撤职前曾与公司秘书发生冲突

在这家石油巨头发生剧变后,本•马修斯已请假离岗。
设置字号×
最小
较小
默认
较大
最大
分享×