Judgment

Outline

• Incorporation of judgment in forecasting
  • Roles of judgment in quantitative forecast methods
  • Judgmental Adjustment
  • Purely Judgmental Forecasts
• Behavioral limitations
• Processes for improving judgment

The Role of Judgment

• By now we have seen a large class of formal forecasting methods
• Decision theory defines principles by which to decide how to evaluate a forecast or forecasting method
  • Choose a way to measure results using a loss function
  • Choose a principle by which to assess forecasting methods: like (average or max) Risk or Regret
  • Choose a model, rule class, or comparison set with which to apply principles
  • Apply a method for constructing a forecast from the chosen set which is supported by the principles
• Most of the approaches are quantitative, exact, and are often automatic once these decisions are made
• But we still need to make these decisions, in a way reflecting our objectives, values, and beliefs
  • This is the role for judgment in forecasting: choices that require human decision making
• Judgmental forecasting examines how humans make these decisions, and suggests ways to do so in accordance with values, beliefs, and goals

Judgmental Forecasting - Approaches

• A purely quantitative forecasting approach is one which can be specified exactly
  • Can write the function down, or implement it on a machine
  • Human input goes into deciding which approach to use, but once decided, result is automatic
• A purely judgmental forecast is one where \(\widehat{y}\) is chosen directly
  • Individual knowledge and motivations used to produce a number
• A “judgmental adjustment” takes a quantitative approach and changes some aspect
  • Forecast produced differs by amount decided subjectively
• Performance and foundations of pure judgmental approach depend on thought process to produce it
  • This could exactly replicate a quantitative approach, sharing its benefits and pitfalls
  • More generally, properties depend on psychology of individual and group decision making
• Reason one might want to use judgment would be to incorporate knowledge or perspectives which can’t easily be expressed in formal way
  • Maybe you have sources of data which are hard to incorporate explicitly
  • Or you suspect that some features of implementation of quantitative method are inadequate for problem actually faced
    • Many “exact” models are “exact” for the wrong problem

Where do choices enter into quantitative methods? - Goals

• The starting point for a forecast should be a clearly defined goal
• What do we want to know that we don’t yet know?
  • Choice of what to forecast is most consequential component of forecasting process
• What are the consequences of getting it right or wrong?
• These questions should be answered based on the particular business context
  • Is there some action we are going to take depending on what we think will happen?
  • Can we quantify the effects of this action in terms of profit or other first order business priorities?
• Express this goal mathematically as a loss function
  • If the result happens to be \(y_{t+h}\) and I guessed \(\widehat{y}_{t+h}\), what are the consequences?
• In practice, methods often use “default” loss function (MSE, 0-1 loss, etc)
  • Whether result is genuinely useful requires good alignment of actual consequences with loss chosen
  • Be sure to at least evaluate using criterion relevant to goals

Choices in Quantitative Methods - Principles

• Choice between forecasting principles usually harder than choice of loss
  • Average vs Maximum, Risk vs Regret, all sensible ways of describing outcomes
• Difficult because you have to compare across sets of multiple outcomes
  • Need to know not only consequences of decisions, but how they depend on situation
• Worry about maximum risk or regret when you can’t afford chance of persistently bad performance
  • This is especially the case if you are providing forecasting tool to someone else and don’t know what they will do with it
  • Example: inventory forecasts sent to each outlet in large company, for use by each local manager
  • Example: automated web user responses set by prediction algorithm
• Average case methods like Bayes allow specifying tradeoffs of what kinds of states performance is valued in
  • Use both knowledge and values to determine these

Choices - Model Class

• Even with a well-defined goal, forecasting approaches require many choices in terms of model
  • In statistical or online learning approach, need a class of predictors; in probability approach, need a probability model of the environment
  • Even with model combination, need a set of models or methods to include
• Class of predictors or models is consequential: by no free lunch principle, no one class is universally better
  • Need some knowledge of features of environment to ensure model captures them
• Visualizing and exploring data useful, as are explicit model selection methods like cross validation & information criteria
  • But can only help choose between set of models you have in mind already (or can find software to implement)
• Process of turning expert knowledge of features of environment you face into model not easy to formalize
  • Knowledge of, say, medical billing or collateralized loans or SMEs in Mozambique may be helpful for figuring out model features
• Can try to account for unexpected possibilities by using complicated models, but overfitting becomes critical problem
  • Also which type of complicated model requires choices: hundreds of kinds of neural network
• Can try to get around this by using simple off-the-shelf models (ARIMA/ETS) in cases where they are good enough
  • Provide a simple description of some ways series behaved in the past, which usually tells you something about the near future
• Can’t apply mindlessly: future pattern can completely change in ways that are predictable if paying attention
  • “This event has never happened in my data set” \(\neq\) “This event has never happened before and won’t again”

Choosing A Prior

• Bayesian inference often called subjective because prior encodes beliefs
  • This is not really a distinction: so does any other feature of a model or method. Forecast class, selection method, etc, should all reflect beliefs
• Priors can and should be chosen to reflect known characteristics of environment
  • Some parameter values are obviously implausible, and this can be encoded by prior
• Consider autoregressive model \(\log y_t=b_0+\sum_{j=1}^{J}b_j\log y_{t-j}+\epsilon_t\) and suppose \(y_t\) is dollars of sales of a small company
  • US GDP is \(\approx 20\) trillion = \(2\times 10^{13}\)
  • If prior on \(b_0\) is \(N(0,\sigma^2)\), \(\sigma>7\) means about 3% chance of sales equal to entire economy
  • Prior on \(\sigma\) should probably put near 0 weight on values above 7
• More generally, can draw parameters from prior and simulate from likelihood, making sure that samples reasonable
  
• Priors are useful place to incorporate numerical adjustments not seen in data directly
  • But only if you can express this as a parameter of your model: If you can’t, maybe you need a new parameter in your model?
• E.g., suppose you have a pretty good additive model for web traffic, but you figure out people stop visiting your website when a new episode of some TV show is airing
  • A prior adjustment to an hour of day or day of week effect will change outcome at that time, but also at times when you don’t want it to
  • Adding TV schedules as a covariate is change in likelihood that will encode this knowledge

Judgmental Adjustment - Motivation

• After figuring out your objectives and building model or method to suit your needs, may still not be satisfied
  • Learn about something relevant for prediction but not in your model
  • Examining past performance find that there is systematic pattern of mistakes
• One approach to handling these issues is to update model to incorporate additional information or correct mistake
  • Add a covariate or predictor, change a functional form, adjust a tuning parameter
• Making such model changes “adaptively”, in response to data, affects prediction properties
  • Increases chance of overfitting, and much more than by number of parameters
  • There are many changes you could and would have made had things been different
• May also not be feasible to incorporate info or change in systematic way
  • If you are tracking corporate performance and see newspaper report showing that company lost a huge lawsuit, where do you put that?
  • Supposing something like this only happens once in your data, adding it as predictor same as just changing today’s or future forecasts
• Model output looking wrong doesn’t always mean you have a way to fix it easily
  • Fix may require going from simple, easy to use, computationally fast model to something that you just can’t code up or run

Judgmental Adjustment - Approaches

• A long established method for “fixing” a forecast that needs nonsystematic changes is an add factor
  • Take the quantitative method output, and change it by a bit to reflect non-systematic patterns
  • Difference from pure judgement is only in view that quantitative forecast is good starting point
• Can also include add factors in model or method parameters, changing them by some amount after fitting
  • Useful when parameters have interpretation and natural units, because you can use common sense about these parameter size to make sure change is reasonable
  • If making sequence of forecasts, this can provide way to jointly adjust both current and future forecasts
• Intercept Corrections are used by professional forecasters to account for large and persistent changes
  • Especially useful if using linear model, like VAR, as shifts not accounted for by other parameters
  • Working with \(\Delta y_t\) (so a permanent shift only affects forecast one period) or explicitly adding shifts to model is an alternative
• When adjustment is tied to specific qualitative information, use checks to make sure adjustment size reasonable
  • Even if this never happened before, maybe you know possible scale or can compare to similar events

Cases For Judgmental Forecasting

• When you have very little data, the choices that go into building a forecast matter a lot more
• Most extreme case of this is Cold Start Problem: prediction for outcome that hasn’t happened yet
  • Want to know sales for new product not yet introduced, or new hire, or new customer, etc
• Even more extreme, may have one shot prediction task: event hasn’t happened, and will only happen once
  • Who will win the next election? What is chance Argentina will default on this (particular) bond issue?
• In these cases, can’t use data on past performance to construct own forecast
• Can still apply formal modeling frameworks and loss criteria, but not much advice on what to do
  • Statistical learning approach guarantees performance only when available data is abundant
  • Online learning approach acts along a full sequence, but guarantees usually on average over the sequence
• Bayesian approach is unchanged in principle even when performed with 0 data
  • But prior dominates forecast, so needs to be chosen carefully, reflecting any informal information
  • Cannot rely on Bernstein von Mises phenomenon which makes prior mostly irrelevant with large data
• In low data situations, even with a formal model, any forecast determined almost completely by judgment

Judgmental Forecasting - Pitfalls

• Behavioral psychology literature has studied how people make predictions, in laboratory and real life settings
• Patterns and quirks have been noted in decision making which seem to correspond to simple rules - heuristics
• Evaluated in simple contexts, performance shows deviations that appear suboptimal relative to benchmarks - biases
• Kahneman and Tversky examined simple problems in which Bayesian updating and expected risk minimization provide clear recommendations for correct decisions
  • In these situations, documented a range of patterns which cannot be justified by any prior or loss function
  • Certain patterns appear systematically across individuals and situations, reflecting generalizable characteristics of human decisions
• This literature suggests that self-reported beliefs do not follow laws of probability including Bayesian updating
  • Also that decisions do not reflect principle of risk minimization
• Many choices made seem to be mistakes: decisions reflecting cognitive limitations
  • Kahneman refers to process rapidly generating these initial, biased decisions as System 1
  • Given time and mental effort for deep consideration and deliberation, System 2 thinking, most individuals would change their decisions
• Not just that people make mistakes, but that they make similar mistakes repeatedly
  • Even when informed of biases, need active thinking to avoid them

Biases in Judgment

• A non-comprehensive list of well-documented decision-making biases includes
• Overconfidence: Decision makers tend to underestimate degree of uncertainty in outcomes
• Probability Weighting: For low probability events, reported chance either set to 0 or much too large
• Stereotypes/Base Rate Neglect: Asked to express conditional probabilities of event \(P(B|A)\), may report high value when \(P(A|B)\) large, even if \(P(B)\) known to be tiny
  • Bayes rule requires \(P(B|A)\propto P(A|B) P(B)\), but ignore \(P(B)\) even when it is known to be rare
  • Can even result in guesses implying \(P(A \cap B)>P(A)\), impossible for any genuine probability
• Saliency bias: Perceived probability affected by order or emphasis with which information is presented
  • More recently or prominently displayed outcomes are considered more likely
• Reference Dependence: In a decision between action A with payoffs \((A_1,A_2)\) in state 1 and 2, respectively, and \(B\) with constant payoff, choice of A vs B depends on whether decision is framed as chance of loss or of gain
  • This difference cannot be expressed by a loss function
• In general, decision making often neglects important components of the problem, and responds strongly to “superficial” aspects which correspond to mere relabeling of the same problem

Biases and Alternative Approaches

• Evidence is strong that Bayesian approach does not come naturally to people
• Maybe shouldn’t be a surprise given that it isn’t even easy for computers
  • Although differences not well understood, cognition also limited by memory and speed
    
• One systematic bias noted early on in literature is in response to decisions where probability not explicitly given
  • Choose between bet with an explicit probability distribution, like a coin flip, and bet with no explicit distribution
  • Ambiguity Aversion: Decision makers likely to favor choice with explicit probability
• This is consistent with a min max risk criterion, since choice with no distribution could be worse than with
  • Could justify this as reflecting principles, though explanations based on cognitive limitations also possible
• Regardless of source of decision making biases, should be aware of limitations of intuitive judgment
• Ability to address limitations of formal models and incorporate qualitative information should be traded off against chance of systematic biases in decisions
  • Need to consider possibility that own choices make guess systematically worse

Improving Judgmental Forecasts

• Aside from switching to quantitative models, can process of judgment be formalized and improved?
• Tetlock et al, sponsored by IARPA, ran forecasting tournaments for one-shot binary predictions of geopolitical events
  • Encouraged research and active effort by participants and studied in detail behaviors associated with good performance
• Found that with time and effort, many though not most participants could consistently achieve high performance
• Several features of their process seemed to play role in encouraging good performance
• Calibration to a reference class: May have similar but not identical events to compare to
  • Among similar events, what was distribution of outcomes? Start with this guess, then adjust
  • Bias research suggests that predictions exhibit anchoring to salient numbers, so initial guess should be genuinely reasonable
• Fermi Problems: Break into subcomponents, each contributing to outcome, for which one can gather info
  • Ex: If asked “How many piano salesmen are there in Chicago”, break up into: population of Chicago, percent of people who own a piano, how frequently might one be bought, how centralized the market might be, what might be employment in each establishment…
• Frequent, small updates: Seek out additional info and change in response to information by sensible amount
• Eclecticism: Strong ideology tends to reduce performance by forcing predictions to match perceived theory
  • Even if broadly right, lots of other things affect outcomes, and may go in opposite direction of big trend for individual events
  • Studies of nominal experts and pundits found little evidence that quantitative predictions outperformed those of informed laypeople

Judgment in Groups and Teams

• If you are not working alone, or can rely on external advice, can obtain the judgment of a group
• This has benefits in ability to obtain extra information and ideas, which can be combined to improve forecasts
  • Model combination using several experts can outperform any individual
  • Tetlock et al find that if you have several good predictors, can help to extremize their forecasts
    • If 5 people independently guess 60%, if each had different information to find this, better to guess number higher than 60%
  • Even if many are biased, can adjust for this quantitatively: use each as one predictor in, e.g., affine model combination
    
• This also has costs in that additional sources of bias can arise in interpersonal settings
• Those making predictions may have reason to misreport to others
  • They may have concerns for reputation or standing, and so, e.g. prefer to be wrong at same time as others
  • They may simply have different goals than you have and want to influence you, as in financial advisers who “talk their book”
• If multiple forecasters communicate as a team, info is shared directly
  • Adds information for each forecaster, but possibly also reinforces errors
  • Working together initially, can be driven to consensus which concentrates on highly biased prediction: groupthink
  • Hyndman suggests iterating, starting by recording independent forecasts, then discussing, then getting new forecasts before aggregating

Steps for Effective Forecasts, Quantitative or Not

• Decide on a goal: and express it quantitatively
• Gather data: Past outcomes, analogous events, possible external predictors
• Build a (class of) models: figure out what features might be relevant and how they might affect outcomes
  • This can be mathematical, or informal, as in the Fermi approach
• Simplify: based on goals and assumptions, choose procedure that should achieve those goals given available info
  • This need not be complicated or incorporate every aspect of problem considered
  • Instead it should be a process which is justifiable given what you know, which should be simple
  • E.g., can just use naive predictor for a series that mostly stays in place, even if you know a lot more about it
• Validate: Get honest feedback about performance and faults to update procedure
  • Can use quantitative measures with external validation data
  • And/or external feedback from informed but impartial third party
• Document process and performance: maintain careful record of procedures and evaluate ex post performance relative to goals
  • Documented process (including code if automated) needed for critical reevaluation to improve future performance

Conclusions

• There is a substantial role for human decision-making even in automated forecast procedures
  • Decide on goals, principles, models, priors using knowledge of situation
• Cases with limited or hard-to-incorporate data call for direct application of judgment
  • Can adjust quantitative forecasts or parameters to reflect information external to model
  • In extreme case, may derive number directly using personal judgment
• Human prediction and decision making shows pervasive biases, especially without careful systematic reevaluation
  • Subjective judgmental forecasts should be subject to scrutiny, evaluation, and modification to improve performance
• By following disciplined process and performing honest evaluation, can help to ensure that knowledge and judgment contribute to achieving goals

References

• Daniel Kahneman (2011) Thinking, Fast and Slow Farrar, Straus and Giroux: New York.
  • Overview of psychology of decision making
• Philip Tetlock and Dan Gardner (2015) Superforecasting: The Art and Science of Prediction Crown Publishers: New York.
  • Findings from experiments in forecasting competitions