Potential Final Project Ideas#
Here are ideas that you might consider for your final project. These are ideas that are meant as inspiration. You should feel free to choose something entirely different, as long as it relates to generative or agentic AI and finance.
Auto-Research for Financial Forecasting#
Autoresearch is a framework by Andrej Karpathy that automates the ML experimentation loop. An AI agent reads your training code, forms hypotheses for improvement, modifies the code, runs fixed-length training experiments, and evaluates whether the changes helped — all autonomously. It can run ~100 experiments overnight on a single GPU.
Andrej Karpathy’s announcement of autoresearch. View original tweet.
For this class, you could apply auto-research to financial time series forecasting using the Financial Time Series Forecasting Repository (FTSFR). FTSFR is an open benchmark that evaluates a suite of classical and modern forecasting methods on financial datasets — including arbitrage basis spreads, banking indicators, and asset returns. You would choose one of the FTSFR datasets and one of the forecasting methods, then use auto-research to systematically improve model performance.
Startup Pitch with Working Prototype#
Another class of project is something you might pitch to a VC firm — a product idea for an AI tool in finance, backed by a working prototype. The prototype should be the smallest possible case that demonstrates the idea is feasible. Think of it as a proof of concept you could demo in a meeting.
For inspiration on what makes a compelling AI startup idea, see this talk:
Benchmarks#
Design and build a benchmark harness, then run evaluations across models and report results. The goal is to create a reproducible evaluation framework for a specific AI capability relevant to finance, run it against multiple models, and report the results. Example benchmark ideas include: quant interview question performance (visual and non-visual), SEC filings information retrieval, financial time series prediction, academic paper table or plot extraction, and PDF-to-markdown conversion accuracy.
For an example of what professional AI benchmarking looks like, see Artificial Analysis. For more context on how benchmarks work and why they matter, see these videos:
Agent Use Cases#
Build an AI agent that solves a specific problem in finance workflows. Examples include: a fact-checking agent, a plot format or common-sense checker (e.g., verifying that financial time series are smooth with no spurious breakpoints), or tools that enhance developer workflows like test-driven development with hold-out evaluation.
Web App#
Build a finance-focused AI web application. This could be a dashboard, an internal tool, or a consumer-facing product prototype that leverages generative or agentic AI.
More details and specific project options will be provided soon.