Modelling behaviour using @RISK and PrecisionTree Why probability estimates aren’t always what they seem Palisade Risk Conference 2011 © 2011Captum Capital Limited Chris Brand Modelling Behaviour The psychology of decision making Examples of judgement errors How to model human behaviour 2 People Aren’t Perfect Humans are not perfectly rational Errors in judgement are predictable People are often poor at making probability judgements 3 Modelling Behaviour The psychology of decision making Examples of judgement errors How to model human behaviour 4 The Monty Hall Problem 5 The Game Three doors; one hides a car, the others hide goats The host opens a door you haven’t chosen to reveal a goat Should you stay with your initial choice, or change your preferred door? 6 Monty Hall in @Risk 7 The logic of the problem Swap Stick £ 0 0 £ 0 0 0 £ 0 0 £ 0 0 0 £ 0 0 £ 1:3 chance to win 2:3 chance to win 8 PrecisionTree Model 9 A Controversial Game A newspaper column received several thousand complaints about this solution Experiments show ~80% think there is no difference between staying or switching Even after training in probability, ~70% still choose the wrong answer 10 Base Rate Neglect Another error in probabilistic judgements Considerably easier to account for than the error demonstrated by the Monty Hall problem 11 Taxi Experiment 85% of taxis are yellow, and 15% are black 12 The Problem A taxi is involved in a hit and run incident A witness testified that the cab involved was black Testing of the witness reveals his vision is accurate 80% of the time 13 What do you think? What is the probability that the witness’ claim is accurate? 14 Witness Testimony Model 15 The Error 29% (12% + 17%) chance that witness will testify that a black cab was responsible But this judgement will only be accurate 12% ÷ 29% = 41% of the time Most people believe accuracy is >50%, and many think it to be 80% In reality, the witness is only 41% accurate 16 Why is this interesting? Judgement and decision making involves reasoning about probabilities, both explicit and implicit Human behaviour does not follow the laws of probability, but is somewhat predictable Can we model behaviour? 17 Modelling Behaviour The psychology of decision making Examples of judgement errors How to model human behaviour 18 The Probabilistic Turn Psychology and cognitive science increasingly influenced by Bayes’ theorem Probabilistic models could be implemented in Palisades’ Decision Tools Suite 19 The Iowa Gambling Task 20 The Setup Deck A is high risk, high reward with a low chance of loss Deck B is high risk, high reward with a 50% chance of a loss Deck C is low risk, low reward with a low chance of loss Deck D is low risk, low reward with a 50% chance of a loss 21 The Goal Free to sample from each deck, for a total of 100 card choices The aim of the task is to maximise the amount of money you possess 22 Standard behaviour Most participants will sample from the various decks, before mainly choosing the high-risk and high-reward decks – decks A and B After several large losses – by approximately trial 40 – they will instead begin to prefer the lower risk decks 23 Some notes… Loss aversion; losses are experienced as more significant than gains of equal value All models have some degree of error to them – especially models of human behaviour 24 Single trial model 25 A (very) simple model Model assumes prior knowledge of card value distributions Recommends deck which most participants eventually settle for Normative model, rather than descriptive 26 Extend model using @Risk PrecisionTree alone does not grant models the flexibility to adequately account for behaviour Using @Risk in conjunction with PrecisionTree aids matters 27 Possible applications Visualisation of theories of decision making Bayesian models Management decision making training tool Consumer behaviour 28 Summary Human behaviour is predictably irrational This behaviour can be modelled using probabilistic decision trees Such models may have applications in teaching and other areas 29 About Captum… Formed in 2004 Transatlantic presence Life science sector consulting: Business development, valuation, partnering MasterClasses: Valuation Masterclass attended by over 500 executives in UK and Europe Internet virtual communities Sensor100 30 Contact Captum Capital Limited Chris Brand e: cmb@captum.com t: +44 (0) 115 988 6154 m: +44 (0) 7800 829 012 Cumberland House 35 Park Row Nottingham NG1 6EE United Kingdom www.captum.com 31