Superforecasting – the thing that TalkTalk didn’t do

I have just been reading Superforecasting: The Art and Science of Prediction by Philip Tetlock and Dan Gardner. The book has attracted much attention and enthusiasm in the press. It makes a bold claim that some people, superforecasters, though inexpert in the conventional sense, are possessed of the ability to make predictions with a striking degree of accuracy, that those individuals exploit a strategy for forecasting applicable even to the least structured evidence, and that the method can be described and learned. The book summarises results of a study sponsored by US intelligence agencies as part of the Good Judgment Project but, be warned, there is no study data in the book.

I haven’t found any really good distinction between forecasting and prediction so I might swap between the two words arbitrarily here.

What was being predicted?

The forecasts/ predictions in question were all in the field of global politics and economics. For example, a question asked in January 2011 was:

Will Italy restructure or default on its debt by 31 December 2011?

This is a question that invited a yes/ no answer. However, participants were encouraged to answer with a probability, a number between 0% and 100% inclusive. If they were certain of the outcome they could answer 100%, if certain that it would not occur, 0%. The participants were allowed, I think encouraged, to update and re-update their forecasts at any time. So, as far as I can see, a forecaster who predicted 60% for Italian debt restructuring in January 2011 could revise that to 0% in December, even up to the 30th. Each update was counted as a separate forecast.

The study looked for “Goldilocks” problems, not too difficult, not to easy but just right.

Bruno de Finetti was very sniffy about using the word “prediction” in this context and preferred the word “prevision”. It didn’t catch on.

Who was studied?

The study was conducted by means of a tournament among volunteers. I gather that the participants wanted to be identified and thereby personally associated with their scores. Contestants had to be college graduates and, as a preliminary, had to complete a battery of standard cognitive and general knowledge tests designed to characterise their given capabilities. The competitors in general fell in the upper 30 percent of the general population for intelligence and knowledge. When some book reviews comment on how the superforecasters included housewives and unemployed factory workers I think they give the wrong impression. This was a smart, self-selecting, competitive group with an interest in global politics. As far as I can tell, backgrounds in mathematics, science and computing were typical. It is true that most were amateurs in global politics.

With such a sampling frame, of what population is it representative? The authors recognise that problem though don’t come up with an answer.

How measured?

Forecasters were assessed using Brier scores. I fear that Brier scores fail to be intuitive, are hard to understand and raise all sorts of conceptual difficulties. I don’t feel that they are sufficiently well explained, challenged or justified in the book. Suppose that a competitor predicts a probability p for the Italian default of 60%. Rewrite this as a probability in the range 0 to 1 for convenience, 0.6 If the competitor accepts finite additivity then the probability of “no default” is 1- 0.6 = 0.4. Now suppose that outcomes f are coded as 1 when confirmed and 0 when disconfirmed. That means that if a default occurs then f ( default ) = 1 and f (no default ) = 0. If there is no default then f ( default ) = 0 and f (no default ) = 1. It’s not easy to get. We then take the difference between the ps and the fs, calculate the square of the differences and sum them. This is illustrated below for “no default” which yields a Brier score of 0.72.

Event p f ( pf ) 2
Default 0.6 0 0.36
No default 0.4 1 0.36
Sum 1.0 0.72

Suppose we were dealing with a fair coin toss. Nobody would criticise a forecasting probability of 50% for heads and 50% for tails. The long run Brier score would be 0.5 (think about it). Brier scores were averaged for each competitor and used as the basis of ranking them. If a competitor updated a prediction then every fresh update was counted as an individual prediction and each prediction was scored. More on this later. An average of 0.5 would be similar to a chimp throwing darts at a target. That is about how well expert professional forecasters had performed in a previous study. The lower the score the better. Zero would be perfect foresight.

I would have liked to have seen some alternative analyses and I think that a Hosmer-Lemeshow statistic or detailed calibration study would in some ways have been more intuitive and yielded more insight.

What the results?

The results are not given in the book, only some anecdotes. Competitor Doug Lorch, a former IBM programmer it says, answered 104 questions in the first year and achieved a Brier score of 0.22. He was fifth on the drop list. The top 58 competitors, the superforecasters, had an average Brier score of 0.25 compared with 0.37 for the balance. In the second year, Lorch joined a team of other competitors identified as superforecasters and achieved an average Brier score of 0.14. He beat a prediction market of traders dealing in futures in the outcomes, the book says by 40% though it does not explain what that means.

I don’t think that I find any of that, in itself, persuasive. However, there is a limited amount of analysis here on the (old) Good Judgment Project website. Despite the reservations I have set out above there are some impressive results, in particular this chart.

The competitors’ Brier scores were measured over the first 25 questions. The 100 with the lowest scores were identified, the blue line. The chart then shows the performance of that same group of competitors over the subsequent 175 questions. Group membership is not updated. It is the same 100 competitors as performed best at the start who are plotted across the whole 200 questions. The red line shows the performance of the worst 100 competitors from the first 25 questions, again with the same cohort plotted for all 200 questions.

Unfortunately, it is not raw Brier scores that are plotted but standardised scores. The scores have been adjusted so that the mean is zero and standard deviation one. That actually adds nothing to the chart but obscures somewhat how it is interpreted. I think that violates all Shewhart’s rules of data presentation.

That said, over the first 25 questions the blue cohort outperform the red. Then that same superiority of performance is maintained over the subsequent 175 questions. We don’t know how much is the difference in performance because of the standardisation. However, the superiority of performance is obvious. If that is a mere artefact of the data then I am unable to see how. Despite the way that data is presented and my difficulties with Brier scores, I cannot think of any interpretation other than there being a cohort of superforecasters who were, in general, better at prediction than the rest.

Conclusions

Tetlock comes up with some tentative explanations as to the consistent performance of the best. In particular he notes that the superforecasters updated their predictions more frequently than the remainder. Each of those updates was counted as a fresh prediction. I wonder how much of the variation in Brier scores is accounted for by variation in the time of making the forecast? If superforecasters are simply more active than the rest, making lots of forecasts once the outcome is obvious then the result is not very surprising.

That may well not be the case as the book contends that superforecasters predicting 300 days in the future did better than the balance of competitors predicting 100 days. However, I do feel that the variation arising from the time a prediction was made needs to be taken out of the data so that the variation in, shall we call it, foresight can be visualised. The book is short on actual analysis and I would like to have seen more. Even in a popular management book.

The data on the website on purported improvements from training is less persuasive, a bit of an #executivetimeseries.

Some of the recommendations for being a superforecaster are familiar ideas from behavoural psychology. Be a fox not a hedgehog, don’t neglect base rates, be astute to the distinction between signal and noise, read widely and richly, etc..

Takeaways

There was one unanticipated and intriguing result. The superforecasters updated their predictions not only frequently but by fine degrees, perhaps from 61% to 62%. I think that some further analysis is required to show that that is not simply an artefact of the measurement. Because Brier scores have a squared term they would be expected to punish the variation in large adjustments.

However, taking the conclusion at face value, it has some important consequences for risk assessment which often proceeds by broadly granular ranking on a rating scale of 1 to 5, say. The study suggests that the best predictions will be those where careful attention is paid to fine gradations in probability.

Of course, continual updating of predictions is essential to even the most primitive risk management though honoured more often in the breach than the observance. I shall come back to the significance of this for risk management in a future post.

There is also an interesting discussion about making predictions in teams but I shall have to come back to that another time.

The amateurs out-performed the professionals on global politics. I wonder if the same result would have been encountered against experts in structural engineering.

And TalkTalk? They forgot, pace Stanley Baldwin, that the hacker will always get through.

Professor Tetlock invites you to join the programme at http://www.goodjudgment.com.

The Iron Law at Volkswagen

So Michael Horn, VW’s US CEO has made a “sincere apology” for what went on at VW.

And like so many “sincere apologies” he blamed somebody else. “My understanding is that it was a couple of software engineers who put these in.”

As an old automotive hand I have always been very proud of the industry. I have held it up as a model of efficiency, aesthetic aspiration, ambition, enlightenment and probity. My wife will tell you how many times I have responded to tales of workplace chaos with “It couldn’t happen in a car plant”. Fortunately we don’t own a VW but I still feel betrayed by this. Here’s why.

A known risk

Everybody knew from the infancy of emissions testing, which came along at about the same time as the adoption of engine management systems, the risks of a “cheat device”. It was obvious to all that engineers might be tempted to manoeuvre a recalcitrant engine through a challenging emissions test by writing software so as to detect test conditions and thereon modify performance.

In the better sort of motor company, engineers were left in no doubt that this was forbidden and the issue was heavily policed with code reviews and process surveillance.

This was not something that nobody saw coming, not a blind spot of risk identification.

The Iron Law

I wrote before about the Iron Law of Oligarchy. Decision taking executives in an organisation try not to pass information upwards. That will only result in interference and enquiry. Supervisory boards are well aware of this phenomenon because, during their own rise to the board, they themselves were the senior managers who constituted the oligarchy and who kept all the information to themselves. As I guessed last time I wrote, decisions like this don’t get taken at board level. They are taken out of the line of sight of the board.

Governance

So here we have a known risk. A threat that would likely not be detected in the usual run of line management. And it was of such a magnitude as would inflict hideous ruin on Volkswagen’s value, accrued over decades of hard built customer reputation. Volkswagen, an eminent manufacturer with huge resources, material, human and intellectual. What was the governance function to do?

Borrowing strength again

It would have been simple, actually simple, to secret shop the occasional vehicle and run it through an on-road emissions test. Any surprising discrepancy between the results and the regulatory tests would then have been a signal that the company was at risk and triggered further investigation. An important check on any data integrity is to compare it with cognate data collected by an independent route, data that shares borrowing strength.

Volkswagen’s governance function simply didn’t do the simple thing. Never have so many ISO 31000 manuals been printed in vain. Theirs were the pot odds of a jaywalker.

Knowledge

In the English breach of trust case of Baden, Delvaux and Lecuit v Société Générale [1983] BCLC 325, Mr Justice Peter Gibson identified five levels of knowledge that might implicate somebody in wrongdoing.

  • Actual knowledge.
  • Wilfully shutting one’s eyes to the obvious (Nelsonian knowledge).
  • Wilfully and recklessly failing to make such enquiries as an honest and reasonable man would make.
  • Knowledge of circumstances that would indicate the facts to an honest and reasonable man.
  • Knowledge of circumstances that would put an honest and reasonable man on enquiry.

I wonder where VW would place themselves in that.

How do you sound when you feel sorry?

… is the somewhat barbed rejoinder to an ungracious apology. Let me explain how to be sorry. There are three “R”s.

  • Remorse: Different from the “regret” that you got caught. A genuine internal emotional reaction. The public are good at spotting when emotions are genuine but it is best evidenced by the following two “R”s.
  • Reparation: Trying to undo the damage. VW will not have much choice about this as far as the motorists are concerned but the shareholders may be a different matter. I don’t think Horn’s director’s insurance will go very far.
  • Reform: This is the barycentre of repentance. Can VW now change the way it operates to adopt genuine governance and systematic risk management?

Mr Horn tells us that he has little control over what happens in his company. That is probably true. I trust that he will remember that at his next remuneration review. If there is one.

When they said, “Repent!”, I wonder what they meant.

Leonard Cohen
The Future

First thoughts on VW’s emmissions debacle

It is far too soon to tell exactly what went on at VW, in the wider motor industry, within the respective regulators and within governments. However, the way that the news has come out, and the financial and operational impact that it is likely to have, are enough to encourage all enterprises to revisit their risk management, governance and customer reputation management policies. Corporate scandals are not a new phenomenon, from the collapse of the Medici Bank in 1494, Warren Hastings’ alleged despotism in the British East India Company, down to the FIFA corruption allegations that broke earlier this year. Organisational scandals are as old as organisations. The bigger the organisations get, the bigger the scandals are going to be.

Normal Scandals

In 1984, Scott Perrow published his pessimistic analysis of what he saw as the inevitability of Normal Accidents in complex technologies. I am sure that there is a market for a book entitled Normal Scandals: Living with High-Risk Organisational Structures. But I don’t share Perrow’s pessimism. Life is getting safer. Let’s adopt the spirit of continual improvement to make investment safer too. That’s investment for those of us trying to accumulate a modest portfolio for retirement. Those who aspire to join the super rich will still have to take their chances.

I fully understand that organisations sometimes have to take existential risks to stay in business. The development of Rolls-Royce’s RB211 aero-engine well illustrates what happens when a manufacturer finds itself with proven technologies that are inadequately aligned with the Voice of the Customer. The market will not wait while the business catches up. There is time to develop a response but only if that solution works first time. In the case of Rolls-Royce it didn’t and insolvency followed. However, there was no alternative but to try.

What happened at VW? I just wonder whether the Iron Law of Oligarchy was at work. To imagine that a supervisory board sits around discussing the details of engine management software is naïve. In fact it was the RB211 crisis that condemned such signal failures of management to delegate. Do VW’s woes flow from a decision taken by a middle manager, or a blind eye turned, that escaped an inadequate system of governance? Perhaps a short term patch in anticipation of an ultimate solution?

Cardinal Newman’s contribution to governance theory

John Henry Newman learned about risk management the hard way. Newman was an English Anglican divine who converted to the Catholic Church in 1845. In 1850 Newman became involved in the controversy surrounding Giacinto Achilli, a priest expelled from the Catholic Church for rape and sexual assault but who was making a name from himself in England as a champion of the protestant evangelical cause. Conflict between Catholic and protestant was a significant feature of the nineteenth century English political landscape. Newman was minded to ensure that Achilli’s background was widely known. He took legal advice from counsel James Hope-Scott about the risks of a libel action from Achilli. Hope-Scott was reassuring and Newman published. The publication resulted in Newman’s prosecution and conviction for criminal libel.

Speculation about what legal advice VW have received as to their emissions strategy would be inappropriate. However, I trust that, if they imagined they were externalising any risk thereby, they checked the value of their legal advisors’ professional indemnity insurance.

Newman certainly seems to have learned his lesson and subsequently had much to teach the modern world about risk management and governance. After the Achilli trial Newman started work on his philosophical apologia, The Grammar of Assent. One argument in that book has had such an impact on modern thinking about evidence and probability that it was quoted in full by Bruno de Finetti in Volume 1 of his 1974 Theory of Probability.

Supposes a thesis (e.g. the guilt of an accused man) is supported by a great deal of circumstantial evidence of different forms, but in agreement with each other; then even if each piece of evidence is in itself insufficient to produce any strong belief, the thesis is decisively strengthened by their joint effect.

De Finetti set out the detailed mathematics and called this the Cardinal Newman principle. It is fundamental to the modern concept of borrowing strength.

The standard means of defeating governance are all well known to oligarchs, regulator capture, “stake-driving” – taking actions outside the oversight of governance that will not be undone without engaging the regulator in controversy, “whipsawing” – promising A that approval will be forthcoming from B while telling B that A has relied upon her anticipated, and surely “uncontroversial”, approval. There are plenty of others. Robert Caro’s biography The Power Broker: Robert Moses and the Fall of New York sets out the locus classicus.

Governance functions need to exploit the borrowing strength of diverse data sources to identify misreporting and misconduct. And continually improve how they do that. The answer is trenchant and candid criticism of historical data. That’s the only data you have. A rigorous system of goal deployment and mature use of process behaviour charts delivers a potent stimulus to reluctant data sharers.

Things and actions are what they are and the consequences of them will be what they will be: why then should we desire to be deceived?

Bishop Joseph Butler

 

FIFA and the Iron Law of Oligarchy

Йозеф Блаттер.jpgIn 1911, Robert Michels embarked on one of the earliest investigations into organisational culture. Michels was a pioneering sociologist, a student of Max Weber. In his book Political Parties he aggregated evidence about a range of trade unions and political groups, in particular the German Social Democratic Party.

He concluded that, as organisations become larger and more complex, a bureaucracy inevitably forms to take, co-ordinate and optimise decisions. It is the most straightforward way of creating alignment in decision making and unified direction of purpose and policy. Decision taking power ends up in the hands of a few bureaucrats and they increasingly use such power to further their own interests, isolating themselves from the rest of the organisation to protect their privilege. Michels called this the Iron Law of Oligarchy.

These are very difficult matters to capture quantitavely and Michels’ limited evidential sampling frame has more of the feel of anecdote than data. “Iron Law” surely takes the matter too far. However, when we look at the allegations concerning misconduct within FIFA it is tempting to feel that Michels’ theory is validated, or at least has gathered another anecdote to take the evidence base closer to data.

But beyond that, what Michels surely identifies is a danger that a bureaucracy, a management cadre, can successfully isolate itself from superior and inferior strata in an organisation, limiting the mobility of business data and fostering their own ease. The legitimate objectives of the organisation suffer.

Michels failed to identify a realistic solution, being seduced by the easy, but misguided, certainties of fascism. However, I think that a rigorous approach to the use of data can guard against some abuses without compromising human rights.

Oligarchs love traffic lights

I remember hearing the story of a CEO newly installed in a mature organisation. His direct reports had instituted a “traffic light” system to report status to the weekly management meeting. A green light meant all was well. An amber light meant that some intervention was needed. A red light signalled that threats to the company’s goals had emerged. At his first meeting, the CEO found that nearly all “lights” were green, with a few amber. The new CEO perceived an opportunity to assert his authority and show his analytical skills. He insisted that could not be so. There must be more problems and he demanded that the next meeting be an opportunity for honesty and confronting reality.

At the next meeting there was a kaleidoscope of red, amber and green “lights”. Of course, it turned out that the managers had flagged as red the things that were either actually fine or could be remedied quickly. They could then report green at the following meeting. Real career limiting problems were hidden behind green lights. The direct reports certainly didn’t want those exposed.

Openness and accountability

I’ve quoted Nobel laureate economist Kenneth Arrow before.

… a manager is an information channel of decidedly limited capacity.

Essays in the Theory of Risk-Bearing

Perhaps the fundamental problem of organisational design is how to enable communication of information so that:

  • Individual managers are not overloaded.
  • Confidence in the reliable satisfaction of process and organisational goals is shared.
  • Systemic shortfalls in process capability are transparent to the managers responsible, and their managers.
  • Leading indicators yield early warnings of threats to the system.
  • Agile responses to market opportunities are catalysed.
  • Governance functions can exploit the borrowing strength of diverse data sources to identify misreporting and misconduct.

All that requires using analytics to distinguish between signal and noise. Traffic lights offer a lousy system of intra-organisational analytics. Traffic light systems leave it up to the individual manager to decide what is “signal” and what “noise”. Nobel laureate psychologist Daniel Kahneman has studied how easily managers are confused and misled in subjective attempts to separate signal and noise. It is dangerous to think that What you see is all there is. Traffic lights offer a motley cloak to an oligarch wishing to shield his sphere of responsibility from scrutiny.

The answer is trenchant and candid criticism of historical data. That’s the only data you have. A rigorous system of goal deployment and mature use of process behaviour charts delivers a potent stimulus to reluctant data sharers. Process behaviour charts capture the development of process performance over time, for better or for worse. They challenge the current reality of performance through the Voice of the Customer. They capture a shared heuristic for characterising variation as signal or noise.

Individual managers may well prefer to interpret the chart with various competing narratives. The message of the data, the Voice of the Process, will not always be unambiguous. But collaborative sharing of data compels an organisation to address its structural and people issues. Shared data generation and investigation encourage an organisation to find practical ways of fostering team work, enabling problem solving and motivating participation. It is the data that can support the organic emergence of a shared organisational narrative that adds further value to the data and how it is used and developed. None of these organisational and people matters have generalised solutions but a proper focus on data drives an organisation to find practical strategies that work within their own context. And to test the effectiveness of those strategies.

Every week the press discloses allegations of hidden or fabricated assets, repudiated valuations, fraud, misfeasance, regulators blindsided, creative reporting, anti-competitive behaviour, abused human rights and freedoms.

Where a proper system of intra-organisational analytics is absent, you constantly have to ask yourself whether you have another FIFA on your hands. The FIFA allegations may be true or false but that they can be made surely betrays an absence of effective governance.

#oligarchslovetrafficlights

Proposition 65

WarningPoster1I had break from posting following my recent family vacation to California. While I was out there I noticed this rather alarming notice at a beach hotel and restaurant in Santa Monica. After a bit of research it turned out that the notice was motivated by California Proposition 65 (1986). Everywhere we went in California we saw similar notices.

I stand in this issue not solely as somebody professionally involved in risk but also as an individual concerned for his own health and that of his family. If there is an audience for warnings of harm then it is me.

I am aware of having embarked on a huge topic here but, as I say, it is as a concerned consumer of risk advice. The notice, and I hesitate to call it a warning, was unambiguous. Apparently, this hotel, teeming with diners and residents enjoying the pacific coast, did contain chemicals emphatically “known” to cause cancer, birth defects or reproductive harm. Yet for such dreadful risks to be present the notice gave alarmingly vague information. I saw that a brochure was available within the hotel but my wife was unwilling to indulge my professional interest. I suspect that most visitors showed even less curiosity.

As far as discharging any legal duty goes, vague notices offer no protection to anybody. In the English case of Vacwell Engineering Co. Ltd v B.D.H. Chemicals Ltd [1969] 3 All ER 1681, Vacwell purchased ampules of boron tribromide from B.D.H.. The ampules bore the label “Harmful Vapour”. While the ampules were being washed, one was dropped into a sink where it fractured allowing the contents to come into contact with water. Mixing water with boron tribromide caused an explosion that killed one employee and extensively damaged a laboratory building. The label had given B.D.H. no information as to the character or possible severity of the hazard, nor any specific details that would assist in avoiding the consequences.

Likewise the Proposition 65 notice gives me no information on the severity of the hazard. There is a big difference between “causing” cancer and posing a risk of cancer. The notice doesn’t tell me whether cancer is an inevitable consequence of exposure or whether I should just shorten my odds against mortality. There is no quantification of risk on which I can base my own decisions.

Nor does the notice give me any guidance on what to do to avoid or mitigate the risk. Will stepping foot inside the premises imperil my health? Or are there only certain areas that are hazardous? Are these delineated with further and more specific warnings? Or even ultimately segregated in secure areas? Am I even safe immediately outside the premises? Ten yards away? A mile? I have to step inside to acquire the brochure so I think I should be told.

The notice ultimately fulfils no socially useful purpose whatever. I looked at the State of California’s own website on the matter but found it too opaque to extract any useful information within the time I was willing to spend on it, which I suspect is more time than most of the visitors who find their way there.

It is most difficult for members of the public, even those engaged and interested, to satisfy themselves as to the science on these matters. The risks fall within what John Adams at University College London characterises as risks that are known to science but on which normal day to day intuition is of little use. The difficulty we all have is that our reflection on the risks is conditioned on the anecdotal hearsay that we pick up along the way. I have looked before at the question of whether anecdote is data.

In 1962, Rachel Carson published the book Silent Spring. The book aggregated anecdotes and suggestive studies leading Carson to infer that industrial pesticides were harming agriculture, wildlife and human health. Again, proper evaluation of the case she advanced demands more attention to scientific detail than any lay person is willing to spare. However, the fear she articulated lingers and conditions our evaluation of other claims. It seems so plausible that synthetic chemicals developed for lethal effect, rather than evolved in symbiosis with the natural world, would pose a threat to human life and be an explanation for increasing societal morbidity.

However, where data is sparse and uncertain, it is important to look for other sources of information that we can “borrow” to add “strength” to our preliminary assessment (Persi Diaconis’ classic paper Theories of Data Analysis: From Magical Thinking through Classical Statistics has some lucid insights on this). I found the Cancer Research UK website provided me with some helpful borrowing strength. Cancer is becoming more prevalent largely because we are living longer. Cancer Research helpfully referred me to this academic research published in the British Journal of Cancer.

Despite the difficulty in disentangling and interpreting data on specific risks of alleged pathogens we have the strength of borrowing from life expectancy data. Life expectancy has manifestly improved in the half century since Carson’s book, belying her fear of a toxic catastrophe flowing from our industrialised society. I think that is why there was so much indifference to the Santa Monica notice.

I should add that, inside the hotel, I spotted five significant trip hazards. I suspect these posed a much more substantial threat to visitors’ wellbeing than the virtual risks of contamination with hotel carcinogens.