Overview
AlphaGPT is an open-source, deep-reinforcement-learning-based automated factor factory for crypto quant trading, released by imbue-bit, a 15-year-old developer who self-describes as managing a ~5M CNY quant fund. The system does not predict prices directly. Instead, it generates interpretable mathematical formulas, evaluates them via backtesting, and trains a Transformer to produce higher-scoring formulas over time.
Architecture
The codebase is organized in clear functional layers:
data_pipeline/— pulls token metadata and OHLCV from Birdeye and DexScreener into Postgres / TimescaleDB.model_core/— encodes market data into features, defines an operator language, and uses a Transformer to generate formula token sequences.strategy_manager/— periodically loads data, scores tokens using the best generated formula, applies risk filters, and routes orders.execution/— wraps Solana RPC and Jupiter aggregator for quoting, signing, and submitting swaps.dashboard/— Streamlit UI for positions, market snapshots, and logs.- Main factors (6):
ret,liq_score,pressure,fomo,dev,log_vol. - Extension factors (12):
vol_cluster,momentum_rev,rel_strength,hl_range,close_pos,vol_trend, and others. - GitHub:
imbue-bit(real name 栀染) - Age: 15, female
- Manages ~5M CNY quant fund
- AlphaGPT repo: https://github.com/imbue-bit/AlphaGPT
- CATREADME: https://github.com/imbue-bit/AlphaGPT/blob/main/CATREADME.md
- no_JIT repo: https://github.com/imbue-bit/no_JIT
- Author profile: https://github.com/imbue-bit
- SourcePulse project page: https://www.sourcepulse.org/projects/22459563
- Copy fork: https://github.com/yuz0101/AlphaGPTCopy
Main flow: data ingestion → formula generation → scoring Top N tokens → risk filter → Jupiter execution → portfolio update → dashboard.
Core Idea: Discover Formulas, Do Not Predict
Traditional quant strategies rely on human researchers hand-writing factor formulas (for example "RSI < 30 with volume expansion") and backtesting them. AlphaGPT treats formula discovery as a sequence generation problem in the spirit of AutoML and symbolic regression. A Transformer proposes candidate token sequences, a StackVM executes each sequence into a signal per timestep, a backtest engine simulates trading, and policy-gradient reinforcement learning uses backtest rewards (Sharpe, max drawdown, win rate) to update the Transformer.
Factors and Operators
The system defines 18 features covering trend, momentum, volatility, volume, and mean reversion:
The 12 operators form a minimal programming language:
| Operator | Function |
|---|---|
| ADD, SUB, MUL, DIV | Arithmetic |
| NEG, ABS, SIGN | Sign operations |
| GATE | Conditional select |
| JUMP | Z-score > 3 spike detection |
| DECAY | Weighted lag aggregation |
| DELAY1 | Lag by one period |
| MAX3 | Max of current and last two lags |
Example: the token sequence [ret, fomo, MUL, liq_score, DIV, DECAY] is pushed onto a stack, operated on step by step, and yields a fully interpretable factor signal. The same underlying principle applies to DeepMind's AlphaTensor, which used Transformers to discover faster matrix multiplication algorithms.
Execution Layer
Signals flow from best_meme_strategy.json through risk filters (liquidity floor, slippage cap, duplicate holding skip) and into execution via Jupiter, Solana's largest DEX aggregator. Risk is layered: signal-level (formula can encode gates such as GATE(liq_score>0.5, signal, 0)), strategy-level (filters based on liquidity and slippage), and execution-level (position caps, daily loss caps in config). The README warns that the default configuration is not production-ready.
no_JIT: A Companion Uniswap V4 Hook
The same author released no_JIT, a Uniswap V4 Hook defending against Just-In-Time (JIT) liquidity attacks. JIT attackers observe large pending swaps in the mempool, insert concentrated liquidity in the same block, capture most of the fee, then withdraw. no_JIT models this as a predator-prey differential game and derives a Nash equilibrium under the Hamilton-Jacobi-Isaacs (HJI) framework, implementing the equilibrium on-chain so JIT becomes ex-ante unprofitable. Key design choices: atomic deterrence within a single block, no oracle dependency (uses on-chain pool state only), and gas-efficient execution.
Author note: the underlying paper, *Defense in Predatory Markets: A Differential Game Framework for AMM Liquidity via Uniswap V4 Hooks*, has not been submitted to a conference.
Author Profile
Published research across deep learning theory, LLM inference and alignment, hyperparameter optimization, autonomous driving, and quant finance. Engineering skill set includes C10K-scale services, Kubernetes, Prometheus, React/Vue, CTF problem authoring, and fluent use of C++, Python, and TypeScript.
Strengths and Limitations
Strengths: clean architectural layering, fully interpretable formulas, elegant StackVM design, Solana-native meme-coin focus, strong academic grounding.
Limitations: no .env template, no default best_meme_strategy.json (must be trained), incomplete risk module, placeholder documentation, and the fundamental quant constraint that widely-adopted factors decay as arbitrage closes the edge.
Broader Significance
AlphaGPT exemplifies the broader shift toward automating strategy discovery, the same direction as AutoML and Neural Architecture Search: humans define the search space, algorithms explore it, and feedback loops compress iteration cycles from weeks to hours. Its open-sourcing may itself reflect strategy decay: when an edge is widely known, it disappears.