Inspiration
In recent years, quantitative trading programs in China’s A-share market have consistently achieved stable and considerable beta returns that outperform the market. For the large number of retail investors in the A-share market, however, the barrier to entry for quantitative trading is extremely high. Meanwhile, discretionary trading by individuals is often prone to challenges influenced by a variety of subjective and objective factors—including trading psychology, the maturity of one’s trading system, the level of trading experience, and so on. Against this backdrop, low-threshold, iterable, convenient, and user-friendly AI quantitative trading assistants based on large language models have become particularly important for retail investors in China’s A-share market at this stage.
What it does
We believe that the strategies developed by cutting-edge financial AI and quantitative firms in the market are comprehensive and sophisticated. Even their foundation models have been specially trained through multiple rounds of iterations, and their performance based on feedback from the A-share market has been quite impressive. However, for individual retail investors, the services provided by quantitative firms are relatively expensive, and standardized products lack flexibility. Retail investors typically prefer a limited number of familiar quantitative strategies that they can iterate on themselves with a low barrier to entry. Against this background, we have decided to focus on deeply refining individual effective trading strategies, mining additional factors that influence strategy performance, presenting them to users through engineering visualization, and providing users with access to preset prompts for strategy iteration. Ultimately, AI agents will help users automatically execute validated strategies to generate returns.
How we built it
Our team, consisting of one Product Manager and two Developers, adopted an AI-first development paradigm. We didn't just use AI to write code; we placed Gemini 3 at the heart of our product as the primary reasoning engine.
The Brain (Gemini 3): We integrated Gemini 3 to act as the "Quantitative Strategist," processing market data and sentiment to make real-time decisions.
Collaborative Architecture: While the PM defined the strategy guardrails, the developers built a seamless pipeline where Gemini 3's insights are directly translated into trade executions.
Challenges we ran into
Domain Knowledge Gap: As newcomers to the stock market, we had to rapidly bridge the gap between financial theory and algorithmic execution, using AI to help us interpret complex market signals.
Engineering the Decision Loop: Integrating Gemini 3 for autonomous decision-making was a major hurdle. We had to ensure the model could not only analyze data but also trigger precise Buy/Sell signals within our custom-built backtesting framework.
Data Reliability: We struggled with unstable data sources, requiring us to build a robust pre-processing layer so that Gemini 3 received clean, high-quality inputs for its reasoning.
Accomplishments that we're proud of
Three-Pillar Proprietary Architecture: We successfully engineered three core modules from the ground up: Stock Selection, Backtesting, and Execution. Unlike many projects that rely on black-box libraries, we built our own infrastructure to ensure full control over the trading logic.
Gemini 3 Powered Decision-Making: We achieved a sophisticated integration where Gemini 3 acts as the "Autonomous Trader." It doesn't just provide text analysis; it actively drives the Backtesting and Execution phases—evaluating historical performance to refine strategies and making real-time Buy/Sell decisions based on live market dynamics.
A Truly Autonomous Loop: We are proud of creating a seamless transition from the PM's strategic vision to an AI-driven system that scans the market, validates its own hypotheses via backtesting, and executes trades without manual intervention.
Full-Stack Synergy in Record Time: Delivering a high-fidelity FinTech dashboard integrated with a cutting-edge LLM within a single hackathon sprint, proving the efficiency of our 1-PM & 2-Dev team structure.
What we learned
LLMs in Finance: We discovered the immense potential (and the necessary constraints) of using Gemini 3 for financial decision-making, moving beyond simple chatbots to functional agents.
Advanced AI Orchestration: We mastered the "Human-in-the-loop" design, learning how to structure data so that an AI agent can reliably handle high-stakes logic like asset allocation.
Agile Quant Development: We learned that the combination of PM vision and AI-assisted coding can compress months of traditional financial engineering into a single hackathon.
What's next for China A-shares popular stock quantitative agent
We will continuously iterate and optimize based on the actual A-share trading performance over a period of time, starting from identified problems. Potential future technical and product iteration directions include: Consider adopting a RAG (Retrieval-Augmented Generation) approach to improve the accuracy of model judgment and decision-making. Consider adopting a multi-agent architecture; for example, assigning the task of selecting high-liquidity, high-popularity stocks to a dedicated agent. Consider encapsulating strategies proven effective in live trading as strategy skills, and iteratively developing more such skills. Attract more users to our product, and establish a flywheel mechanism for demand and data through their feedback, so as to continuously optimize product functions, as well as the accuracy and stability of model decision-making and execution.
Log in or sign up for Devpost to join the conversation.