### abstract ###
we analyze behavior in two basic classes of decision tasks  description-based and experience-based
in particular  we compare the prediction power of a number of decision learning models in both kinds of tasks
unlike most previous studies  we focus on individual  rather than aggregate  behavioral characteristics
we carry out an experiment involving a battery of both description- and experience-based choices between two mixed binary prospects made by each of the participants  and employ a number of formal models for explaining and predicting participants' choices  prospect theory pt CITATION   expectancy-valence model evl CITATION   and three combinations of these well-established models
we document that the pt and the evl models are best for predicting people's decisions in description- and experience-based tasks  respectively  which is not surprising as these two models are designed specially for these kinds of tasks
furthermore  we find that models involving linear weighting of gains and losses perform better in both kinds of tasks  from the point of view of generalizability and individual parameter consistency
we therefore  conclude that  overall  when both prospects are mixed  the assumption of diminishing sensitivity does not improve models' prediction power for individual decision-makers
finally  for some of the models' parameters  we document consistency at the individual level between description- and experience-based tasks
### introduction ###
all our lives we have to make decisions
we have to choose where to go on vacation  when to replace our old car  which pair of shoes to buy
in fact  each step we make in our life is a result of a decision we have made
even if we do nothing  this is probably our personal choice
of course  we would be happy if we could always make correct decisions in order to maximize the resulting utility  yet we sometimes fail as a result of objectively insufficient information or subjective behavioral biases
therefore  understanding and systematically describing people's behavior is extremely important both for predicting their future decisions and for potentially improving decision quality
decisions we make may be classified into two main categories
when we study newspaper daily weather forecasts  drug package inserts and mutual funds' brochures  we enjoy convenient descriptions of the risky prospects  including the probabilities of possible outcomes
respectively  decisions based on such statistical descriptions are called description-based decisions
when we decide whether to backup their computer's hard drive  cross a busy street  go on a blind date  put on a belt during driving  we are typically denied a benefit of convenient descriptions of the possible outcomes for example  the probabilities of a hard disk failure  of an accident  or of meeting a desirable partner in a blind date are never explicitly provided
in many such decisions  all we can rely on is our own past experience
respectively  decisions based on past personal experience are called experience-based decisions
tasks typically studied under the description-based paradigm tend to focus on one-shot decisions that are based on detailed information concerning the relevant outcome distributions
tasks typically studied under the experience-based paradigm do not provide objective prior information concerning the payoff distributions  and this could be these tasks' drawback
on the other hand  the decisions are repeated  and thus decision-makers get a chance to learn from experience
the distinction between risky description-based and experience-based decisions has attracted recent attention because the ostensibly same information can lead to different choices depending on how the information is acquired CITATION
this difference  sometimes referred to as description-experience gap  is usually attributed to the difference in treatment of rare outcomes in the two paradigms
on the one hand  in description-based tasks  according to  presentation effect   low-probability events are likely to be overweighted if their probabilities are explicitly presented CITATION
on the other hand  in experience-based tasks  low-probability outcomes may be underweighted either because of a recency effect-since low-probability events are not likely to occur recently the availability of personal experience tends to reduce the weighting of these events CITATION -or simply because of a small sample size  which may cause rare events not to occur at all during relatively short intervals of time CITATION
as an example of a description-experience gap  we may cite a classic example given by weber CITATION   consider the decision of whether to vaccinate a child against diphtheria  tetanus  and pertussis dtp
parents who research the side effects of the dtp vaccine by consulting the national immunization programweb site or a brochure provided by their pediatrician will learn that up to  NUMBER  child out of  NUMBER   NUMBER  will suffer from high fever and about  NUMBER  child out of  NUMBER   NUMBER  will suffer from seizures as a result of immunization
an increasing number of parents  after reading such information  decide not to immunize their child
although doctors have the same statistics at their disposal  they also have access to information not available to parents-namely  personal experience gathered across many patients
this information tells them that vaccination is very unlikely to result in side effects
few doctors will have encountered one of the rare cases of high fever or seizures
if they have encountered one  the experience is dwarfed by hundreds of memories of side-effect free immunizations
one sources of differences between doctors and vaccine resisters may result from differences in the weight given to rare events like the likelihood of a seizure as a function of whether this likelihood is acquired through experience or statistical description
in naturally occurring situations  decision-makers often base their decisions both on descriptions and on their own experience
in laboratory settings it is possible to construct choice tasks based purely on descriptions or experience
our major goal is to determine which models best explain behavior in these two classes of decision tasks
unlike previous studies focusing on aggregate behavioral characteristics of groups of people CITATION   our study analyzes choice evaluation parameters for individual decision-makers
moreover  we examine the connection between the choices of the same decision-makers in description- and experience-based tasks
we carry out an experiment involving a battery of description-based and experience-based choices
so that our decision tasks are more representative of real-world situations  each of the choices is made between two binary prospects involving probabilities of both gains and losses
to examine the common behavioral characteristics of these kinds of tasks  we employ a number of formal decision learning models  incorporating a number of factors or parameters  namely  loss aversion  diminishing sensitivity  probability weighting  choice consistency or the degree of randomness in choice  and recency the tendency to relate past information to the current choice
first  we use two well-established models  i prospect theory pt CITATION   which was developed mainly for description-based tasks  and ii expectancy-valence model evl CITATION   which was designed specially for experience-based tasks
in addition  we create three combinations of these basic models to find out what parameters would be most suitable for describing human behavior at the individual level by comparing same person's choices in the two types of tasks  i pt-no-s model  which is similar to the pt without the assumption of diminishing sensitivity to gains and losses  ii evl-s model  which is similar to the evl  but with the assumption of diminishing sensitivity  and iii evl-pt model with utility function similar to that of the pt  except for the loss aversion parameter
in description-based tasks  in order to account for the presentation effect  a learning model should contain a possibility of non-linear weighting of explicitly stated probabilities
therefore  we employ models possessing the probability weighting parameter  that is  pt and pt-no-s  for this kind of tasks
on the other hand  in order to reflect the recency effect  learning models dealing with experience-based decisions should contain a recency parameter
the evl  evl-s and evl-pt models meet this criterion
first of all  we expect that the classical pt and evl models that were specially designed for description- and experience-based tasks  respectively  will prove more suitable than their possible combinations for the respective tasks
our findings  based on the models' fit  support this expectation
furthermore  we expect that people tend to behave consistently in terms of decision-making model parameters CITATION
we thus expect that the better fitting models should also provide higher individual parameter consistency
in other words  we hypothesize that an individual who  according to a certain model  reveals relatively high low values of certain decision parameters in a certain task should reveal  according to the same model  relatively high low values of the same parameters in other tasks of the same choice paradigm
additionally  we expect that the contribution of the diminishing sensitivity assumption to explaining decision-makers' behavior in our experiment may be rather limited  due to the mixed nature of prospects we employ
the former surmise is partially supported  since in description-based tasks the combined pt-no-s model is found to outperform the pt model in terms of parameter consistency
this finding is in line with other results indicating that models involving linear weighting of gains and losses perform better in both kinds of tasks
finally  we expect that individuals' behavior should also be consistent between different choice paradigms  that is  individual parameters having similar functionality in description- and experience-based tasks should be positively correlated
this hypothesis is only partially supported  suggesting that these widely-used learning models may still need further improvement
the rest of the paper is structured as follows
in section  NUMBER   we present the models of decision-making we employ
in section  NUMBER   we describe our experimental design and research approach
section  NUMBER  describes the empirical tests we perform and provides the results
section  NUMBER  concludes and provides a brief discussion
