♠ ♥ ♦ ♣

Designing CogCards — a behavioral assessment for evaluating new hires

A novel, game-based cognitive assessment I built from scratch to guide hiring and improve investment performance, deployed firmwide.

Role
End-to-end ownership
Stakeholders
CIO, Human Capital, Portfolio Managers at a leading global hedge fund
Timeline
5 months, research → deployment
Overview

A multidimensional assessment in the form of a game

CogCards is an online, game-based cognitive assessment that measures risk appetite and other key cognitive dimensions to guide hiring decisions and improve investment team performance.

Instead of a simple personality test, candidates play a complex, card-based gambling game, resulting in a metric-rich snapshot of their decision-making under stress and uncertainty.

Development involved constructing and validating a novel measurement framework that triangulated behavioral and survey-based data to predict real-world investment performance.

The Problem

Risk-taking was a blind spot

Stakeholders at a leading global hedge fund agreed that current risk-taking among the investment team was suboptimal — but had no reliable way to assess risk appetite and other key cognitive dimensions in current or prospective hires. It's well known in finance that optimal risk-taking is critical to investing skill.

Project goal

Build a tool to assess risk appetite and decision-making ability for current and prospective employees that's engaging, valid, and ab bele to deployed at scale.

Research questions

Can we identify behavioral patterns that impact investment performance — and what methodology, validated rigorously, can deliver that quickly with limited engineering resources?

Process

Four stages of iterative development

Month 1

Research & Planning

Competitive analysis found no real precedent — just personality tests or small scientific instruments. Our objectives pointed toward combining a behavioral task with a brief self-report survey.

Months 2–3

Development & Iteration

Built and piloted a game prototype and results dashboard, cycling through stakeholder feedback and analysis between rounds.

Month 4

Deployment

Shipped the final platform (JavaScript + Claude Code) with a custom-built Python-based interactive results dashboard, live with candidates and current analysts.

Month 5+

Analysis & Refinement

Validated the measurement framework internally and externally, then mined behavioral insights from the results.

Methodology

Behavior, self-report, and a dashboard

Behavioral

A card-based gambling game where players try to win as much money as possible. Decision-making patterns as the game unfolds reveal risk appetite and other cognitive abilities.

Survey

Closed- and open-ended questions to triangulate against behavioral data and surface additional qualitative insights.

Deliverables

A polished cloud-deployed game platform, admin panel, and a dashboard surfacing player results for hiring teams.

Key Findings

Suboptimal Risk, Overconfidence, and Cognitive Biases

Validating the framework internally (behavioral × self-report) and externally (against portfolio metrics) surfaced several surprising patterns: risk-taking was suboptimal, the worst performers overestimated their own performance (and the best performers underestimated theirs), and there were surprisingly-high rates of cognitive biases impacting decision-making.

Behavioral Data

The current investment team was missing a key range of risk appetite scores entirely.

Self-report Data

An unexpectedly high rate of cognitive biases, including gambler's fallacy.

Combined

Clear evidence of a Dunning-Kruger effect — overconfidence concentrated among weaker performers.

Qualitative

Open-ended responses (analyzed with Claude) surfaced recurring player archetypes.

Evidence for Gambler's Fallacy

“If I lost several card draws in a row, I would keep hitting that deck because I knew a win would be coming…”

Evidence for the Dunning–Kruger Effect

Scatter plot of 50 data points illustrating the Dunning-Kruger effect. Points below the diagonal line indicate underestimation of skill; points above indicate overestimation.
Individual respondent Perfect calibration (AB line)

Overconfidence is a widespread and consequential problem in investing. The relationship between actual and self-perceived performance here illustrates a classic Dunning–Kruger pattern: lower-performing participants tended to overestimate their performance, while higher-performing participants were more likely to underestimate it.

Recommendations & Impact

From insight to measurable change

+5%
Improvement in risk appetite scores, comparing the six months before deployment to the six months after
Reflections & Next Steps

What I'd carry into the next project

Behavioral + self-report measures, externally validated against portfolio metrics, gave a framework that could survive scrutiny — and leaning on LLMs for prototyping and qualitative analysis let a single owner ship something this rigorous in five months.

Some findings have been generalized or omitted, and all visuals are mockups, to preserve confidentiality. They remain faithful to the original results and content.