ECON3124 What I Learned in Behavioral Economics: Complete Notes from Core Theories to Final Review

Machine-translated with NiuTrans. Read the Chinese original.

ECON3124 is Behavioural Economics. Instead of taking "human irrationality" as a conclusion, it systematically asks three questions above the standard economic model: How exactly do human preferences deviate from the standard model? Why do people's beliefs go systematically wrong? When these people compete with each other, how will market outcomes and public policy change?

In the process of completely reviewing this course, I gradually organized the content scattered in the handouts, tutorials, and mock exams into a set of interconnected frameworks, and deepened my understanding by using a cycle of concept recognition, formula modeling, independent exercises, error diagnosis, and targeted retesting. Below is both a course knowledge map and a summary of how to systematically review behavioral economics.

I. The main theme of the entire course: Where the standard model fails

I ended up understanding the class as a four-tier structure:

1. standard benchmarkProbability, Bayesian update, expected utility, exponential discounting, Nash equilibrium, and inverse induction.
2. preference biasReference dependence, loss aversion, present bias, projection bias, belief utility, and social preference.
3. Belief and Reasoning BiasRepresentativeness, availability, anchoring, finite-level reasoning, cursed thinking, and erroneous learning.
4. Market and policy consequencesInformation avoidance, the winner's curse, path dependence, internalization, and pushing and soft parenting.

This main thread is important because new situations in exams often don't directly tell you "this is a loss aversion question" or "this is a current bias question." What really needs to be done is to first write down the predictions of the standard model, then determine why the data or behavior is inconsistent with it, and finally choose the behavioral model that can explain the difference.

A very practical framework for answering English questions is:

Define the concept → explain the mechanism → derive its prediction → connect the prediction to the evidence.

That is:Define concepts - explain mechanisms - derive predictions - link evidence。

II. Risk, Bayesian Update, and Reference Dependence

1. Bayesian formulas are not something you can just memorize.

The Bayesian formula is the foundation for the first part of the course:

What is really easy to get wrong is not the formula itself, but the mixing of false positives and false negatives. For example, the prevalence is , sensitivity is The false positive rate is 。 Assume 10,000 people: 100 people get sick, of which 99 are positive; 9,900 people are unaffected, of whom 99 have also tested positive. Therefore, the probability of a positive individual actually developing the disease is only...

I would also mess up the denominator at first. Later, the most effective method was not to continue memorizing symbols, but to draw a four-cell table and remember:The numerator is true positive, and the denominator is all positive, that is, true positive plus false positive.。 Parenthesizing any long denominator can avoid many basic calculation errors.

2. Expected utility, certainty equivalence, and risk premium

Standard Expected Utility Model Writing:

deterministic equivalence Satisfy The risk premium is

A concave utility function implies diminishing marginal utility, so risk-averse individuals are willing to sacrifice a portion of their expected wealth to eliminate risk. A distinction must be made between "loss aversion” and "risk aversion”: the former relies on reference points, while the latter is explained by the curvature of wealth utility in the Standard Model.

3. Prospect theory: The outcome depends on where you look at it.

The utility with a reference point can be written as:

A simplified expression of prospect theory is:

It contains three core elements:

• reference dependencyPeople often evaluate not ultimate wealth, but gains and losses relative to a certain point of reference.
• Loss aversionThe pain caused by the same amount of loss is greater than the joy caused by the gain.
• Decreasing sensitivity and probability weightingThe further away from the reference point, the weaker the psychological impact of the additional changes; People may also overestimate low probabilities and underestimate medium to high probabilities.

This may explain the endowment effect. If the slope of the profit range is 1, the slope of the loss range is 2, and the consumption value brought by one cup is 4, then the people who do not have a cup are willing to pay a maximum of about 2, while the people who already have a cup are willing to pay a minimum of about 8. The difference between the buy price and the sell price is not the change in income, but the transaction is coded as a gain or loss under different reference points.

Taxi drivers' "income targets" are also a typical example: after daily targets become a reference point, they are more willing to continue working on low-income days and finish work earlier on high-income days, which is different from the standard labor supply model that only looks at hourly wages.

III. Time Preference: Why do we always do it tomorrow even though we know we should do it?

1. Index Discounting and Dynamic Consistency

The exponential discounting model is:

Its key property isdynamic consistencyAs long as the time interval between outcomes remains constant, people's relative preferences for the same set of future options will not reverse simply due to the passage of time. While reviewing, I initially wrote about dynamic consistency too vaguely, but later realized that the answer must explicitly include two parts: "the same future option" and "it will not reverse over time."

2. – Model and current bias

Quasi-hyperbolic discounting writing:

Indicates a conventional level of patience. Then give "now" extra weight. It can explain procrastination, impulse spending, addiction, and insufficient savings.

Two types of people must be distinguished here:

• **naive** knows they will procrastinate now, but mistakenly believes their future self will follow the original plan.
• **sophisticated** foresees that they will still make current mistakes in the future, so they will generalize backwards and may also actively seek commitment devices.

A good task to solve this problem is: The costs of completing the task in four phases are 8, 12, 18, and 27 respectively, and 。 The naive type kept thinking that "it would be more cost-effective to do it in the next phase," and eventually dragged it down to a cost of 27; Mature types predict their future choices and work backward, potentially bearing the next outcome of the cost of 12 in the first phase, rather than constantly comparing superficial adjacent numbers. When doing this type of question, you can't just compare the nominal cost of the next period; you must compare...The path that actually happens after the delay。

More subtly, there is harmful consumption. If each purchase yields 4 in the current period, 2 in the price paid, and 3 in the future, and The marginal return I currently see is

Therefore, consumption may be chosen in each period; But any self, when evaluating the entire sustainable consumption path in advance, may wish it had never happened. This is why "voluntary every time" does not necessarily equal "the whole path is in line with long-term welfare".

IV. Projection Bias, Belief Utility, and Information Avoidance

1. Projection deviation

When people are hungry, thirsty, in pain, or excited, they overestimate the extent to which they will remain in the same state in the future. One expression in the course is:

In the agreement of this set of lecture notes, It indicates an accurate prediction. Indicates complete projection. However, different textbooks sometimes switch the order of the two items, so a more reliable method than rote memorizing "the larger the parameter, the more biased it is" is to read the formulas directly during the exam:Which weight is multiplied by the current state to represent the projection strength?。

I also made a typical mistake when doing problems: confusing state variables with utility. The future state of hunger may be However, the effectiveness of eating hamburgers in this state may be 2; These two numbers have completely different meanings.

2. Beliefs can also have direct effects.

In the standard model, beliefs are only used to choose actions; Behavioral models allow "believing yourself to be healthy, smart, or liked" to be useful in itself. This can lead to information avoidance and self-deception.

If the belief before obtaining the information is Post-information beliefs are Psychological utility function For a concave function, then the Jensen inequality may give:

Even if information has instrumental value, the belief fluctuations it causes may reduce the expected psychological utility, so people choose not to test, not to look at bills, or not to check exam results. When calculating psychological costs, good news and bad news must be weighted by probability; We cannot compare only one bad outcome with the current state.

The self-deception question also requires the use of two probabilities simultaneously:objective probabilityDetermine the material rewards of action.Subjective and more optimistic probabilityDetermine the effectiveness of beliefs. If a person does not take the test, he usually cannot take the right action in both real situations. This is the core of judging the value of information.

V. Judging Heuristics, Selecting Frameworks, and Psychological Accounts

There are many concepts in this part, but it can be summarized as "judging how the presentation is changed".

1. Three classic heuristics

• representativeBecause the description looks like a certain type, the underlying probability is ignored. The conjunction fallacy in the Linda problem violates 。
• availabilityThe easier an event is to remember, the more common it is judged to be. Intensive media coverage of rare accidents amplifies risk perception.
• anchorThe numbers that appear first, even if they are unrelated, will drive subsequent estimates.

The most effective answer here is not to simply write the bias name, but to point out neglected canonical benchmarks, such as underlying probabilities, set inclusion relationships, or true information independent of the anchor.

2. Preferences are not a fixed list to carry with you.

Bait effects, compromise effects, selection overload, and default option specifications indicate that preferences are constructed at the selection site. The standard model predicts that adding a disadvantage option should not change the order of the original two items; If the order changes, the menu itself plays a role in shaping preferences.

3. Narrow frameworks and mental accounts

When evaluating two gambles separately, people may choose A and D; When they are combined, B+C increases by 10 in each state. The problem isn't that we don't calculate expectations, but rather that we only assess each risk in local accounts without integrating the total wealth results. This mechanism is frequently used in insurance, investment, coupons, and sinking topics.

VI. Social Preferences and Behavioral Game Theory

1. First, master standard game theory benchmarks.

There is only one criterion for judging Nash equilibrium:Stick to other people's strategies; no player is willing to deviate unilaterally.。 It cannot be replaced by "everyone is happier" or "total returns are higher". In sequential games, reverse induction is performed starting from the last node.

2. Outcome fairness: Charness–Rabin type preference

Taking Player 2 as an example, one segmentation utility is:

Before calculating, you must first see whether player 2 is leading or trailing, and then select the corresponding branch. Social preferences may alter the optimal response, thereby altering Nash equilibrium.

3. Same result, different intention

Allocation preferences only depend on the final payment; The intention and reciprocity models also depend on how the other party could have chosen. In the end, it was all However, if the proposer can fairly distribute and Between choices, the rejection rate is often higher than when he simply has no fairer option. When you see "the same result, different menus, or unselected options" on an exam, you should prioritize intentions, reciprocity, moral wiggle room, or face-saving, rather than just unequal results.

VII. Finite Strategic Reasoning, Cursed Thinking, and Learning

1. Beauty contests cannot be adjusted directly only once.

If the proportional rule is Percentage of those who are completely non-strategic And choose 50, all rational participants choose , the overall average is

The optimal choice of a rational person satisfies ,therefore

When When, the answer is

During my review, I found that the easiest thing to miss here is the feedback between rational players: you can't just multiply 20% of irrational choices directly, but you also need to bring your own choices back into the overall average.

Level- The model makes the finite inference hierarchy more intuitive: if ,then

for example At that time, each level was 50, 40, 32, and 25.6 in sequence.

2. Information projection, cursed thinking, and the winner's curse

In games involving private information, actions themselves leak signals. Cursed players underestimate the "correlation between actions and private information" and are therefore underupdated. A very short answer is:

Actions reveal signals; cursed agents under-update from those actions.

The winner's curse in shared value auctions stems from the selection effect: winning the bid often means one's own valuation is relatively the most optimistic, therefore

Rational bidders must revalue and lower their bids on the condition that "I win," rather than directly bidding on the unconditional valuation.

3. Learning does not necessarily converge to the same result.

The course compares mechanisms such as reinforcement learning, belief learning, and EWA. Reinforcement learning relies more on the actual returns of chosen actions, belief learning updates judgments about other people's strategies, and EWA attempts to unify the two.

Path-dependent experiments further demonstrate that small initial differences can push the population towards different stability points through positive feedback, such as ultimately concentrating at 3 or 12. Short-term history is not just noise; it can permanently change long-term outcomes.

8. Behavioral Welfare Economics: When to Intervene

The core of behavioral policy is not to immediately prohibit deviations upon seeing them, but to identify them.internalityToday's choices incur costs for the future that are not fully factored in today.

In a simplified model, if future damage is The current error parameter is , the discount factor is The corrective tax can be written as:

It is not a traditional externality tax: the victim and the decision-maker are different temporal selves of the same person.

Policy tools can be understood in terms of intervention intensity:

• Free will paternalism: Retain freedom of choice, such as joining by default but allowing exit.
• asymmetric paternalismIt primarily helps people who make mistakes, while minimizing hindrance to rational choices.
• Steady parenting: Even if the estimation of behavioral parameters is inaccurate, it can improve welfare under a wide range of conditions.

Tools in the course include default options, active selection, cooling-off periods, commitment devices, disclosures, and durable selection. A complete policy answer should unfold along the lines of "bias—intrinsics—tools—welfare costs," rather than simply saying "just set the defaults." Defaults can also mislead heterogeneous groups, and taxes can punish inherently rational people, so collateral damage and exit costs must be discussed.

9. The most worthwhile formulas to include in the Cheat Sheet and their analysis

module	Minimum formula or judgment	high frequency trap
Bayes	Posterior = True positive / All positive	Confusing false positives and false negatives
expected utility		Directly replace expected wealth with expected utility
prospect theory	Results relative to reference point coding	Equate loss aversion with general risk aversion
Now error	None in the current period All future items	Forget the difference between naive and sophisticated expectations
projection deviation	Prediction of Future Utility of Current State Pollution	Utility in confusing state values and states
information avoidance	Comparative instrument value and psychological cost	Forgot to give probability weights to good and bad news
Nash equilibrium	Fixing others' strategies and checking for unilateral deviations	Replacing equilibrium conditions with Pareto optimality
social preferences	First, determine who is leading, then select the piecewise function.	Confusion between outcome fairness and intention reciprocity
Beauty Contest	Simultaneous solution with	Ignore feedback between rational players
Winner's Curse	Valuation updated on the condition of "winning the bid"	Direct bidding at unconditional value
Conduct Policy	Deviation → Intrinsic → Tools → Welfare	They only talk about promotion, without discussing heterogeneity and costs.

Several other types of purely computational errors are particularly worth noting: the consecutive probabilities of independent events must be multiplied; for example, two consecutive successes are... ； When transferring items, check for inequality signs step by step, for example The launch is ； Multilevel denominators must be enclosed in parentheses. These mistakes are unrelated to theory, but they are the easiest to lose points under time pressure.

10. My complete review method: not only to "finish reading," but also to "be able to recognize."

Complete review can be conducted in stages according to the connections between knowledge: first cover risk, time preferences, beliefs and heuristics, then deal with game theory and social preferences, and finally complete limited reasoning, learning, policy analysis and simulation questions. This sequence allows subsequent behavioral models to be based on clear standard model benchmarks, and also facilitates the connection of multiple modules in a comprehensive question.

The most efficient loop for me is:

1. Explain the normative benchmarks and deviation directions of the concept in one sentence.
2. Look at a complete example and observe how to recognize models from text.
3. Do a similar question independently without looking at the answers.
4. Label the error as one of four categories: "concept recognition, formula setting, algebraic calculation, and mechanism expression".
5. Do another question just for the type of error, instead of rereading the entire chapter from scratch.

Bayesian is a typical example: the first mistake was a false positive, and it was only after drawing a four-bar table and retesting that it was truly mastered. Beauty Contest made the mistake of missing equilibrium feedback and only understanding the model after reassembling the equations. The problem with dynamic consistency is not a lack of concepts, but rather an insufficiently precise definition; The problem with the social preference question is that the utility function branch was not examined first. After categorizing the errors, the review speed is significantly faster than repeatedly reading the lecture notes.

When prioritizing review, I think the following order can be used:

1. First priorityBayes, expected utility, prospect theory,–Projection bias, information avoidance.
2. Second priority: Nash equilibrium, reverse induction, social preference, intention, Beauty Contest, Level-。
3. Third priorityCursed thinking, winner's curse, learning models, intrinsics, and policy comparisons.
4. Final addition: Experiment name and scattered facts; They should be recorded on the mechanism, not recited in isolation.

Mock exams cannot be mechanically treated as real weights. The Mock Final we have is clearly shorter than the official exam, with only two major questions; What is truly worth migrating is the question structure: giving an unfamiliar situation, and then continuously requiring calculations, explanations of mechanisms, comparison of models, and proposal of policy implications.

I will use three sets of short templates in the exam room:

• Theory questionsStandard model predictions remain unchanged; The observed changes support a behavioral mechanism because...
• Calculation problemWrite formulas - substitute numbers - compare candidates - explain the economic meaning in one sentence.
• Policy questions: Identifying biases - explaining intrinsics - choosing tools - discussing benefit gains, collateral injuries, and exit costs.

11. What did I really take away after completing ECON3124?

The most valuable aspect of this course is not providing a "list of human biases," but rather teaching people to discuss irrationality in a testable way. A behavioral explanation must explain what the baseline model is, where the deviation enters utility or belief, what different predictions it will produce, and whether experimental and real-world data can distinguish between these models.

It also changed the way I look at policy. A person making a choice at a certain moment does not guarantee that that choice represents long-term welfare; However, finding behavioral deviations does not automatically give policymakers unlimited power to intervene. Good behavioral policies need to simultaneously respect heterogeneity, freedom of choice, and the benefits of error correction.

The same applies to review itself: real progress doesn't come from "how many pages I read today," but from being able to identify mechanisms in new questions, write correct models, explain the results, and know exactly where I went wrong last time. For me, this is the experience that is most worth retaining after reviewing ECON3124 in its entirety.

This article is based on my ECON3124 lecture notes, tutorials, mock exams, and personal review dialogues from the semester I studied. The formula symbols and assessment focus may change with the semester; please refer to the current course materials for the most accurate information.

Cover photography:Kaboompics.com / Pexels