ReplicaVire

Inspiration

Investing is one of the best ways to build long-term wealth, but professional research and analysts are out of reach for people without disposable income. We wanted to see if a network of machine learning models could do an analyst's job, reading the news, weighing the numbers and explaining its reasoning, at almost no cost.

What it does

ReplicaVire is an AI research and paper-trading platform for 55 large-cap stocks across all 11 market sectors. It combines 15 kinds of market evidence with an LLM agent that reads news and SEC filings to judge how sentiment is changing for each company. It then forecasts the expected move over six timeframes, from 6 hours to 5 days, and explains every recommendation in plain English. Users choose their sectors, set minimum and maximum dollar amounts per trade, and decide whether the system can buy, short, or both. They can get popup recommendations or turn on auto-trading for paper trades.

How we built it

We pull price, news, filing, analyst and short-volume data from Alpaca, Finnhub, SEC EDGAR and FINRA, and store it in Tiger Data. Fifteen ridge regression models each analyze one type of evidence. A Snowflake Cortex RAG agent tracks how sentiment shifts, using an open-source Llama 3.1 model. A PyTorch neural network fuses everything into forecasts and explanations, and it is trained on USC's Theia HPC cluster with strict checks against using future data. The dashboard is React and TypeScript on Vercel with a Python FastAPI backend, and trades run through Alpaca's paper-trading API.

Challenges we ran into

The hardest part was making sure the model couldn't accidentally "peek" at the future. We timestamped every data point by when it actually became public and wrote tests to prove nothing leaks. We also had to handle forecasts that land while the market is closed, combine models without them over-trusting each other, and work within free-tier data limits. We rebuilt the trading dashboard in the final 45 minutes.

Accomplishments that we're proud of

We built the full pipeline, from raw data to explained trade recommendations, with working sector filters, position sizing, long and short modes, day-trade protection and auto-trading. The system is honest by design. Its performance page compares results against a simple baseline and the S&P 500 instead of showing made-up accuracy numbers.

What we learned

Markets are mostly noise, so honest measurement matters more than model complexity. How sentiment changes tells you more than its level. We also gained hands-on experience with time-series databases, stacked models, RAG agents and HPC computing in one weekend.

What's next for ReplicaVire

Next, we'll train on the full multi-year history and publish the results honestly. Then we plan to expand to more stocks, add product reviews and personalized risk profiles, and explore the regulatory path beyond paper trading. All of it serves our goal of giving everyday investors access to professional-grade research.