Predicting loss aversion behavior with machine-learning methods

Behavioral economics, particularly Prospect Theory and Cumulative Prospect Theory by Kahneman and Tversky, highlights cognitive biases like loss aversion. Loss aversion, the greater emotional impact of losses compared to equivalent gains, has been extensively studied in various contexts, including financial decision-making, risk-taking, and consumer behavior. This study aims to leverage machine learning to predict loss aversion. Empirical findings in behavioral and neuroeconomics inform social engineering tools, consumer tendencies, risk appetite, macroeconomic policy trade-offs, tax/penalty rates, and election prediction. Machine learning offers a powerful tool to address the complexity of these predictions, particularly considering the intricacies of raw data, variable determination, transmission mechanisms, and the influence of the human factor. Policymakers can use machine learning to improve behavioral models and parameter efficiency, refining models based on decision-makers' risk profiles. This allows for a better understanding of expenditure patterns, policy effectiveness, and international trade policy dynamics, all crucial elements of economic growth analysis. The study's objective is to investigate the impact of demographic and psychological factors on loss aversion using hybrid machine learning models, furthering understanding and improving the design and implementation of these models.

The literature extensively covers loss aversion, initially introduced by Kahneman and Tversky (1979). Studies explore the concept's implications for various decision-making contexts and its neural basis. The relationship between loss aversion and other factors has also been examined. For instance, research explores the interplay between loss aversion and overconfidence (where overconfidence might lead to riskier choices resulting in losses), and the connection between loss aversion and hopelessness (where significant losses could contribute to hopelessness). The association between loss aversion and financial literacy (where greater financial literacy might lead to better risk management) is also discussed. Existing studies explore potential gender and age differences in loss aversion and whether income level influences loss aversion. Several studies utilize different methodologies to investigate loss aversion and its effect on various aspects, like trade policy, macroeconomic models, and investor behavior, highlighting the need for advanced prediction models. Previous research also uses machine learning models, including Random Forests, demonstrating their effectiveness in predicting human decision making based on behavioral theories.

This study employs a hybrid machine learning approach that combines behavioral theories with machine learning algorithms. The loss aversion coefficient (λ) is a central concept, representing the ratio of the pain of loss to the pleasure of gain (Kahneman and Tversky, 1979). The study utilizes an experimental design involving 28 student participants making decisions in a mixed gamble simulation with four options: "Reject," "Strictly Reject," "Accept," and "Strictly Accept." Reaction times (milliseconds) were recorded. The data included objective (gamble details), naïve (milliseconds, gain/loss differences), sociodemographic (gender, age, department, income), and psychological (self-confidence, hopelessness, financial literacy) features. Machine learning classifiers (Decision Tree Classifier, Random Forest Classifier, Kernel SVC, k-NN Classifier) predicted the probability of accepting a gamble (0 or 1), used as an input for regression algorithms (Decision Tree Regressor, Random Forest Regressor, Kernel SVR, k-NN Regressor). The data was split into training (80%) and testing (20%) sets. Accuracy and mean squared error (MSE) were used to evaluate the model performance. The analysis was performed using Python (Keras, Pandas, NumPy, Matplotlib, Plotly) and MATLAB.

Random Forest proved superior to other algorithms, achieving the lowest MSE in both classification and regression tasks. The median loss aversion coefficient (λ) was determined to be 3.1 (range: 0.5-6), higher than the typical range of 2-2.5 reported in previous studies. This difference might be attributed to the experimental design and participant scale values. An interesting finding is the "irresistible impulse of gambling," where decision-making time decreased as the gain/loss ratio approached the median loss aversion value. Decision times were significantly longer when the gain/loss ratio was near zero or when the predicted probability of acceptance was around 50%. Decision times were shorter for high-probability acceptance gambles with large gain/loss differences. The average decision time was 1600 ms, with shorter times for accepted gambles (1564 ms) than rejected gambles (1670 ms). These findings align with Kahneman's "Fast and Slow Thinking" model. The study highlights the effectiveness of integrating behavioral insights (like loss aversion) within machine learning models to improve predictive power. The 'diff' feature (gain/loss ratio) was a highly influential factor in the model.

The findings address the research question by demonstrating the predictive power of hybrid machine learning models in capturing loss aversion. The higher λ value than typically reported in the literature may be due to the specific experimental setup and participant characteristics. The "irresistible impulse of gambling" finding suggests a dynamic interaction between loss aversion and decision-making speed. The superior performance of the Random Forest algorithm aligns with prior research on machine learning in behavioral economics. This study emphasizes the potential of integrating behavioral insights into machine learning for improved prediction accuracy. The integration of reaction time and psychological factors enhances the model's capability to capture the complexities of human decision-making under risk.

This study demonstrates the value of hybrid machine learning models for predicting loss aversion behavior. Random Forest's superior performance highlights the efficacy of integrating behavioral insights into data-driven models. The findings provide valuable insights into the dynamics of loss aversion and decision-making, including the novel "irresistible impulse of gambling" phenomenon. Future research could explore other biases and expand the model's application to broader contexts.

The study's ecological validity might be limited due to the laboratory setting. The simplicity of the decision tasks may not fully reflect real-world complexities. The use of self-reported measures introduces potential biases. The relatively small sample size of 28 participants limits the generalizability of the findings. Potential confounding factors not controlled for in the study may influence the results.