[Paper] Human-AI Co-Mentorship in Project-Based Learning: A Case Study in Financial Forecasting

Source: arXiv - 2605.05144v1

Overview

This paper reports on a summer‑long, project‑based learning experiment where high‑school and early‑undergraduate students built an ETF price‑forecasting model under the guidance of graduate mentors and AI‑powered development assistants. The authors show that AI tools can shift the learning focus from low‑level syntax to higher‑order problem formulation, letting novices contribute meaningfully to a real‑world finance task.

Key Contributions

• Human‑AI co‑mentorship framework: Introduces a workflow where human mentors and generative AI tools jointly guide students through each stage of a data‑science project.
• Workflow‑first pedagogy: Demonstrates that teaching students to design the end‑to‑end pipeline (data collection → preprocessing → modeling → evaluation) before exposing them to individual algorithms accelerates comprehension.
• Empirical case study: Documents a concrete financial‑forecasting project (ETF price prediction) completed by students with minimal prior AI or finance knowledge.
• Toolchain insights: Evaluates the effectiveness of specific AI assistants (code generation, documentation, debugging, and hyper‑parameter tuning) in a classroom‑like setting.
• Scalable mentorship model: Shows how daily stand‑up meetings combined with AI support can replace many traditional lecture hours while preserving conceptual depth.

Methodology

1. Team composition – 4 high‑school students, 2 early‑undergraduate volunteers, and 3 graduate mentors.
2. Project definition – Students chose “ETF price prediction” as a tangible, data‑rich problem.
3. Workflow design – In the first stand‑up, the group mapped out the required steps: data sourcing (Yahoo Finance API), cleaning, feature engineering, model selection, training, and evaluation.
4. AI‑augmented development – Participants used large‑language‑model (LLM) assistants (e.g., GitHub Copilot, ChatGPT) for:
   • Generating boilerplate code (API calls, pandas pipelines)
   • Suggesting feature‑engineering ideas (technical indicators, lag features)
   • Prototyping models (ARIMA, LSTM, Gradient Boosting)
   • Debugging runtime errors and explaining statistical concepts on demand
5. Iterative mentorship – Daily 15‑minute stand‑ups focused on high‑level questions (why a model is chosen, interpretation of results) while AI handled routine implementation.
6. Evaluation – Model performance was measured with mean absolute percentage error (MAPE) on a held‑out test set; student learning was assessed via self‑reported confidence surveys and code‑review rubrics.

Results & Findings

Practical Implications

• Accelerated onboarding – Companies can use LLM‑driven code assistants to bring junior developers up to speed on domain‑specific pipelines (e.g., fintech, health‑tech) without exhaustive classroom training.
• Mentor bandwidth – AI tools can offload routine coding and debugging, letting senior engineers spend more time on architecture, design reviews, and strategic guidance.
• Project‑based curricula – Educational programs (bootcamps, corporate up‑skilling) can adopt a workflow‑first approach, letting learners define the problem space before diving into algorithms, which improves engagement and retention.
• Rapid prototyping – Teams building predictive models for market data can leverage AI‑generated feature pipelines and model scaffolding to iterate faster, shortening time‑to‑insight.
• Personalized learning tracks – By letting AI suggest resources tailored to each learner’s interest (finance vs. ML), organizations can foster deeper expertise without a one‑size‑fits‑all curriculum.

Limitations & Future Work

• Tool dependence – The study relied on proprietary LLMs; performance may vary with open‑source alternatives or newer models.
• Small sample size – Only one project team was evaluated, limiting generalizability across domains or larger cohorts.
• Evaluation depth – While model accuracy was comparable to academic baselines, the study did not explore robustness to regime shifts or live‑trading constraints.
• Future directions – The authors propose scaling the co‑mentorship model to multiple concurrent projects, integrating automated evaluation dashboards, and investigating long‑term retention of concepts after the AI‑assisted phase ends.

Authors

• Freyaa Chawla
• Ahan Chawla
• Rishi Singh
• Joe Germino
• Grigorii Khvatskii

Paper Information

• arXiv ID: 2605.05144v1
• Categories: cs.LG, cs.CY
• Published: May 6, 2026
• PDF: Download PDF