This thesis provides a multi-disciplinary approach to decision modeling using finance and neural networks. Neural modeling simulates biological decision-making mechanisms. Modern finance formalizes human decision-making. Combining them illuminates how to better emulate intelligent decision processes.

Three projects address key aspects of decision processes heretofore neglected in automated decision models: selective attention, working memory, and social learning. Exploring each aspect from the perspective of real world traders yields quantifiable enhancements to existing neural models.

Financial decision-making first involves gathering evidence to consider before rendering an informed decision. Basing decisions on unfiltered and conflicting evidence raises confounds that reduce the reliability of the judgment. Clustering based on professional trader interviews can implement selective attention that filters and cleans the evidence. If similar lines of evidence are treated as a single unit, this can clarify and improve the decision-making process. Results from neural networks enhanced with this selective attention process show significant improvements on standard benchmark tasks relative to alternative approaches.

Financial decision-making next considers the context under which the evidence arrives. For example, investing in a bankruptcy-prone company is far riskier than investing in a company with growth prospects, all else equal Financial and accounting industry best practices connect past and present evidence to establish the context, which implements a form of working memory. Working memory enables a neural network to aggregate multiple observations in time to form a temporal contextual pattern. Results show the enhancement provides significant improvements.

Finally, financial decision-makers often bias their actions by observing others' actions and inferring the relevant evidence. Industry best practices allow auditors and analysts to partially leverage others' recommendations to reduce the information costs of decision making. Allowing a population of heterogeneous neural networks to observe each others' decisions simulates markets of constantly learning individuals. The simulations allow empirical exploration of novel market dynamics, and results corroborate prior research predictions on the nature of markets.