Production-grade, event-driven quantitative backtesting engine with deep learning signal generation. Built from scratch in Python + PyTorch.
生产级事件驱动量化回测引擎，集成深度学习信号生成。基于 Python + PyTorch 从零构建。
A complete quantitative trading infrastructure covering the full pipeline: data ingestion -> feature engineering -> model training -> signal generation -> order execution -> portfolio management -> performance analysis. Designed around a central EventBus architecture with pluggable components.
完整的量化交易基础设施，覆盖全链路：数据采集 → 因子工程 → 模型训练 → 信号生成 → 订单执行 → 组合管理 → 绩效分析。以中央事件总线（EventBus）为核心架构，所有组件可插拔替换。
The project was developed iteratively across 10 versions (v1–v10), each addressing critical flaws discovered in the previous version — from data leakage bugs to unrealistic execution assumptions to cross-validation methodology. The final v10 uses Combinatorial Purged Cross-Validation (CPCV) across 15 splits on 1M+ bars of real Binance market data, with adversarial execution modeling, achieving Sharpe 0.38 and +58% return on purely out-of-sample data.
项目经历了 10 个大版本的迭代（v1–v10），每个版本都在解决上一版暴露出的致命缺陷——从数据泄露、不切实际的撮合假设到交叉验证方法论漏洞。最终 v10 在 100 万+ 条真实 Binance 市场数据上使用组合净化交叉验证（CPCV, 15 splits），含逆向选择撮合模拟，在纯样本外数据上达到 Sharpe 0.38、收益率 +58%。
┌─────────────────────────────────────────────────────────────┐
│ Data Layer 数据层 (data/) │
│ Binance Archive Downloader ──→ Parquet Data Lake │
│ CCXT Multi-Exchange Feed ──→ SQLite Cache │
│ WebSocket Daemon ──→ Avro/Parquet Stream │
└──────────────────────┬──────────────────────────────────────┘
│
┌──────────────────────▼──────────────────────────────────────┐
│ Factor Layer 因子层 (factors/) │
│ Plugin Factor Library (10 hot-loadable .py files) │
│ FactorRegistry: auto-discover + @register_factor │
│ Causal rolling z-score normalization (no look-ahead) │
└──────────────────────┬──────────────────────────────────────┘
│
┌──────────────────────▼──────────────────────────────────────┐
│ Model Layer 模型层 (model/) │
│ CrossAssetGRUAttention (GRU temporal + cross-asset attn) │
│ QuantTransformer (Encoder-Decoder, 3 presets) │
│ CrossSectionalTransformer (4D [B,A,T,F] + ListMLE) │
│ Dual Loss: ListMLE + Focal + Uncertainty Weighting │
└──────────────────────┬──────────────────────────────────────┘
│
┌──────────────────────▼──────────────────────────────────────┐
│ Engine Layer 引擎层 (engine/) │
│ CPCV: Combinatorial Purged Cross-Validation (15 splits) │
│ EventBus (pub/sub, 7 event types) │
│ Adverse Selection Simulator + TWAP Executor │
│ Kelly Criterion Sizing + Drawdown Circuit Breaker │
└──────────────────────┬──────────────────────────────────────┘
│
┌──────────────────────▼──────────────────────────────────────┐
│ Paper Trading 模拟盘 (paper_trading/) │
│ Live WebSocket → Model Inference → Simulated Execution │
│ SQLite Logger (signals / fills / equity snapshots) │
└─────────────────────────────────────────────────────────────┘
| Module 模块 | Description 描述 |
|---|---|
cpcv.py |
Combinatorial Purged Cross-Validation with purge + embargo / 组合净化交叉验证，含净化+隔离 |
events.py |
Typed EventBus with 7 event types / 类型化事件总线，7 种事件类型 |
order_book.py |
LOB matching, adaptive cost model (A-share / Crypto) / 限价指令簿撮合，自适应成本模型 |
adverse_selection.py |
Micro-execution: 80% favorable reject, 100% adverse fill / 逆向选择模拟器 |
twap_executor.py |
TWAP split-order execution / TWAP 拆单执行器 |
execution.py |
Kelly Criterion dynamic position sizing / Kelly 公式动态仓位管理 |
portfolio.py |
Position tracking, equity curve, Sharpe/Calmar/MaxDD / 持仓跟踪、权益曲线 |
risk.py |
Max drawdown circuit breaker / 最大回撤熔断器 |
| Module 模块 | Description 描述 |
|---|---|
base.py |
BaseFactor ABC + @register_factor decorator + FactorRegistry / 基类+装饰器+注册表 |
log_return.py |
Log returns / 对数收益率 |
sma_ratio.py |
SMA5 and SMA20 price ratios / SMA5/SMA20 价格比率 |
ema_ratio.py |
EMA10 price ratio / EMA10 价格比率 |
rsi.py |
Relative Strength Index / 相对强弱指标 |
macd.py |
MACD histogram / MACD 柱状图 |
bollinger.py |
Bollinger Band position / 布林带位置 |
volume_zscore.py |
Volume z-score / 成交量 z-score |
trade_imbalance.py |
Trade-based order imbalance (OBI) / 基于成交的订单不平衡度 |
price_impact.py |
Amihud illiquidity ratio / Amihud 非流动性比率 |
funding_rate.py |
Funding rate proxy (direction × volume) / 资金费率代理 |
btc_dominance.py |
Relative strength vs own mean / 相对自身均值的强弱 |
volume_momentum.py |
Short/long volume acceleration / 短期/长期成交量加速度 |
qlib_pack.py |
8 Qlib-inspired factors: kmid, klen, kup, klow, roc10, corr_pv, std20, max20_ratio |
multi_timeframe.py |
Multi-TF wrapper: factors at 1h+4h+24h scales / 多时间尺度因子封装 |
| Module 模块 | Description 描述 |
|---|---|
cross_asset_attention.py |
GRU temporal + cross-asset self-attention / GRU 时序 + 跨资产自注意力 |
transformer.py |
Encoder-Decoder Transformer (3 presets, CUDA) / 编解码 Transformer |
cross_sectional.py |
4D [B, A, T, F] + ListMLE ranking loss / 4D 横截面 + 排序损失 |
features.py |
Feature pipeline (delegates to factor registry) / 因子管线 |
obi_features.py |
Order Book Imbalance features / 订单簿不平衡度因子 |
| Module 模块 | Description 描述 |
|---|---|
archive_downloader.py |
Bulk download data.binance.vision → Parquet / Binance 归档批量下载 |
async_feed.py |
CCXT concurrent feed → SQLite / CCXT 并发拉取 |
ws_daemon.py |
WebSocket daemon + heartbeat + exp backoff / WebSocket 守护进程 |
avro_writer.py |
Avro streaming serialization for real-time data / Avro 实时流序列化 |
lake_loader.py |
Parquet data lake reader / 数据湖加载器 |
| Module 模块 | Description 描述 |
|---|---|
schema.py |
8 typed dataclasses: Data, Feature, Model, CV, Train, Execution, Portfolio / 8个类型化配置类 |
__init__.py |
load_config(yaml_path) + default_config() / YAML加载 + 默认配置 |
| Module 模块 | Description 描述 |
|---|---|
engine.py |
Live bar ingestion → inference → simulated execution / 实时K线 → 推理 → 模拟执行 |
logger.py |
SQLite logger: signals, fills, equity snapshots / SQLite 日志 |
realtime_feed.py |
Binance WebSocket kline feed (replaces 6s REST) / WebSocket 实时K线 |
| Module 模块 | Description 描述 |
|---|---|
factor_analyzer.py |
Alphalens-style IC analysis across 1h/6h/24h/48h horizons / 因子IC分析 |
| Module 模块 | Description 描述 |
|---|---|
data/funding_fetcher.py |
Binance Futures real funding rate historical API / 真实资金费率 |
data/onchain_fetcher.py |
Coinmetrics community on-chain metrics (free) / 链上指标（免费） |
engine/numba_backtest.py |
Numba JIT backtest loop (~50x faster) / Numba加速回测 |
engine/adaptive_sizing.py |
RL-inspired Kelly with drawdown awareness / 自适应Kelly仓位 |
model/patch_tst.py |
PatchTST alternative (ICLR 2023) cross-asset variant / PatchTST模型 |
| Version 版本 | What Changed 改动 | Why 原因 |
|---|---|---|
| v1 | Single-asset Transformer + MSE / 单资产 Transformer + MSE | Starting point / 起点 |
| v2 | Fixed data leakage (global → rolling z-score) / 修复数据泄露 | v1 MSE = 10⁻⁶ was fake / v1 的 MSE 是假的 |
| v3 | Directional Focal Loss + OBI features / 方向性 Focal 损失 | MSE can't predict direction / MSE 无法预测方向 |
| v4 | Multi-asset 4D tensors + ListMLE ranking / 多资产 ListMLE | Ranking > absolute return prediction / 排序优于绝对收益预测 |
| v5 | Adverse selection execution / 逆向选择撮合 | v4 Sharpe 1.38 was "fill illusion" / v4 高夏普是"成交幻觉" |
| v6 | 1h + TWAP + 48h hold lock + 5% filter / 低频+TWAP+持仓锁 | v5 lost 48% to friction / v5 被摩擦吃掉 48% |
| v7 | Walk-Forward + GRU cross-asset attention / WFO+GRU跨资产注意力 | Static split leaks info / 静态划分泄露信息 |
| v8 | 1M+ bars, 60-fold WFO / 百万数据60折WFO | 720 bars not significant / 720条无统计意义 |
| v9 | Reversal diagnosis / 反转诊断 | Proved model > pure factors / 证明模型优于纯因子 |
| v10 | CPCV + config + factor plugins + paper trading + avro | WFO has boundary leakage; need industrial infra / WFO有边界泄露；需工业级基建 |
| v11 | 13 factors + d128 + 18-month data + daily paper trading | More data + alternative factors + production readiness / 更多数据+另类因子+生产就绪 |
| v11.1 | Checkpoint save/load + paper trading bug fix | Discovered paper trading was running on RANDOM weights for 12 days (41.7% win rate ≈ random) / 发现 paper trading 跑了12天随机权重模型 |
| v11.2 | 10-feature optimization sweep | Major upgrade: 21 factors, fold ensemble, Numba backtest, PatchTST, multi-TF, real funding rate, on-chain data, WS feed, adaptive Kelly, factor IC analyzer / 21因子+折集成+Numba回测+PatchTST等 |
Source / 数据源: Binance 5m klines (18 months, 3.15M rows) → aggregated to 1h
Binance 5分钟K线（18个月，315万行）→ 聚合为1小时
Assets / 资产: 20 crypto pairs / 20个加密货币交易对
Factors / 因子: 13 plugin factors (10 price-volume + 3 alternative)
13个插件因子（10个量价 + 3个另类）
Model / 模型: CrossAssetGRUAttention d_model=128, 617K params
Validation / 验证: CPCV (N=6, k=2) → 15 splits, purge=24, embargo=48
OOS Coverage / OOS覆盖: 117,672 bars (100% of samples) / 全部样本
Execution / 执行: TWAP 4-slice + adverse selection (65% adverse fill)
Avg Rank Corr / 平均排名相关: 0.062 (all 15 folds positive / 15折全部为正)
Total Return / 总收益: -56.1% (dominated by transaction costs / 交易成本主导)
Max Drawdown / 最大回撤: 60.6%
Rebalances / 换仓次数: 2,451
Transaction Cost / 交易成本: $369K (36.9% of capital / 占本金36.9%)
Avg Hold / 平均持仓: 48 hours / 48小时
| Metric 指标 | v8 (WFO) | v10 (CPCV 6m) | v11 (CPCV 18m) |
|---|---|---|---|
| Data / 数据 | 44K bars (6m) | 44K bars (6m) | 117K bars (18m) |
| Factors / 因子 | 10 | 10 | 13 |
| Model params / 模型参数 | 124K | 124K | 617K (d128) |
| Rank Correlation | 0.025 | 0.068 | 0.062 |
| OOS Coverage | 43,200 | 44,760 | 117,672 |
| Statistical confidence / 统计置信度 | Low / 低 | Medium / 中 | High / 高 |
| Bug: Lookahead | Yes | Fixed | Fixed |
| Bug: Boundary leak | Yes | Fixed | Fixed |
- CPCV >> WFO: 15-split CPCV with purge+embargo produces 2.7x better rank correlation than sequential WFO, because each fold trains on ~24K samples (vs 1.4K in WFO) CPCV 远优于 WFO：每个 fold 训练 24K 样本（WFO 仅 1.4K），排名相关性提升 2.7 倍
- Rank correlation is stable at 0.06: Consistent across 6-month and 18-month datasets, across d_model=64 and d_model=128, proving the signal is real and not an artifact of any specific configuration 排名相关性稳定在 0.06：跨 6 个月和 18 个月数据集、跨 d_model=64 和 128 均一致，证明信号真实
- 1h label >> 6h label: 6h cumulative return as training target degrades rank_corr from 0.068 to 0.039 — crypto 1h reversal signal is stronger at shorter horizons 1h 标签远优于 6h 标签：6h 标签将 rank_corr 从 0.068 降至 0.039，crypto 短期反转信号在更短周期更强
- Transaction costs dominate PnL: With 65% adverse fill rate and 2,451 rebalances, costs ($369K) exceed gross alpha — paper trading is the next validation step 交易成本主导 PnL：65% 逆向成交率 + 2451 次换仓，成本远超毛利 — 模拟盘是下一步验证
- Model > pure factors: v9 diagnosis proved GRU+Attention (rank_corr=0.025) beats pure factor reversal (-37%) and pure momentum (-25%) 模型优于纯因子：v9 诊断证明 GRU+Attention 优于纯因子反转和纯动量策略
After 12 days of paper trading (Mar 30 – Apr 24), reviewing accumulated data revealed:
- Win rate: 41.7% (5W / 7L) — close to random baseline
- Cumulative return: -0.74%
- Sharpe: -0.15
- Day-to-day volatility: 2.84% (way too high for market-neutral)
Root cause: run_paper_daily.py initialized the model with random weights at every run instead of loading the trained checkpoint. The model was effectively a random number generator — explaining why results matched random chance.
12天模拟盘数据的 review 揭露：胜率 41.7% 接近随机基准，累计 -0.74%，夏普 -0.15。
根因：run_paper_daily.py 每次初始化随机权重而非加载训练好的checkpoint，模型实质上是随机数生成器，所以结果等同于随机选择。
Fix in v11.1 / 修复方案:
run_v11_final.pynow trains a final production model on all data and saves checkpoint tocheckpoints/v11_production.ptrun_paper_daily.pynow loads that checkpoint at inference time- Old
paper_daily.dbarchived aspaper_daily_random_weights_backup.db - Need to retrain (run_v11_final.py once) before next paper trading session
torch>=2.0.0
ccxt
polars
pyarrow
websockets
aiohttp
pyyaml
dacite
fastavro
# Bulk download 6 months of 5m klines from Binance archive (886K rows, ~11s)
# 从 Binance 归档批量下载6个月K线（88.6万行，约11秒）
python data/archive_downloader.py
# Or fetch via CCXT (works in geo-restricted regions)
# 或通过 CCXT 获取（适用于网络受限地区）
python data/async_feed.py# v11: 13 factors + d128 + 18-month data + CPCV (~40 min on RTX 5090)
# v11：13因子 + d128 + 18个月数据 + CPCV（RTX 5090约40分钟）
python run_v11_final.py# Run once per day (~30 seconds): fetch bars → inference → log signal → reconcile
# 每天跑一次（约30秒）：拉K线 → 推理 → 记录信号 → 对账
python run_paper_daily.pypython run_v8_bigdata.py # WFO with bug fixes / 修复后的WFO
python run_v6_lowfreq.py # Low-freq TWAP / 低频TWAP
python hyperparam_search.py # Grid search / 网格搜索
python main.py # Single-asset synthetic / 单资产合成数据quant-infra/
├── config/ # Config system / 配置系统
│ ├── schema.py # 8 typed dataclasses / 8个类型化配置类
│ └── __init__.py # YAML loader / YAML加载器
├── configs/
│ └── v10_cpcv.yaml # Default CPCV config / 默认CPCV配置
├── engine/ # Backtest core / 回测核心
│ ├── cpcv.py # Combinatorial Purged CV / 组合净化交叉验证
│ ├── events.py # EventBus + 7 events / 事件总线
│ ├── order_book.py # LOB matching / 撮合引擎
│ ├── adverse_selection.py # Adverse selection / 逆向选择
│ ├── twap_executor.py # TWAP execution / TWAP执行
│ ├── execution.py # Kelly sizing / Kelly仓位
│ ├── portfolio.py # Portfolio / 组合管理
│ ├── risk.py # Risk manager / 风控
│ └── backtest.py # Event loop / 事件循环
├── factors/ # Plugin factor library / 插件化因子库
│ ├── base.py # BaseFactor + FactorRegistry / 基类+注册表
│ ├── log_return.py # Log returns
│ ├── sma_ratio.py # SMA5/SMA20 ratios
│ ├── ema_ratio.py # EMA10 ratio
│ ├── rsi.py # RSI
│ ├── macd.py # MACD
│ ├── bollinger.py # Bollinger position
│ ├── volume_zscore.py # Volume z-score
│ ├── trade_imbalance.py # Trade imbalance (OBI)
│ ├── price_impact.py # Amihud illiquidity
│ ├── funding_rate.py # Funding rate proxy / 资金费率代理
│ ├── btc_dominance.py # Relative strength / 相对强弱
│ └── volume_momentum.py # Volume acceleration / 量能加速
├── model/ # PyTorch models / 模型
│ ├── cross_asset_attention.py # GRU + cross-asset attention
│ ├── transformer.py # Encoder-Decoder Transformer
│ ├── cross_sectional.py # 4D CrossSectional + ListMLE
│ ├── features.py # Feature pipeline / 因子管线
│ ├── obi_features.py # OBI features
│ └── strategy.py # Signal generation / 信号生成
├── paper_trading/ # Paper trading / 模拟盘
│ ├── engine.py # Live inference engine / 实时推理引擎
│ └── logger.py # SQLite logger / SQLite日志
├── data/ # Data ingestion / 数据采集
│ ├── archive_downloader.py # Binance archive → Parquet
│ ├── async_feed.py # CCXT → SQLite
│ ├── avro_writer.py # Avro streaming / Avro流式写入
│ ├── ws_daemon.py # WebSocket daemon
│ ├── lake_loader.py # Parquet loader / 数据湖加载
│ └── synthetic_lob.py # Synthetic data / 合成数据
├── run_v11_final.py # v11 CPCV (13 factors, d128, 18m) / v11主管线
├── run_v10_cpcv.py # v10 CPCV pipeline / v10管线
├── run_paper.py # Paper trading entry / 模拟盘入口
├── run_paper_daily.py # Daily batch paper trading / 每日批处理模拟盘
├── run_v8_bigdata.py # v8 WFO (bug-fixed) / v8 WFO（已修复）
├── run_v6_lowfreq.py # v6 low-freq / v6低频
├── run_v7_wfo.py # v7 WFO
├── hyperparam_search.py # Grid search / 网格搜索
├── main.py # Single-asset / 单资产
└── requirements.txt
Developed and tested on / 开发和测试环境：
- CPU: AMD Ryzen 9 9950X3D
- GPU: NVIDIA GeForce RTX 5090 (32GB VRAM)
- RAM: 64GB DDR5
- Sentiment-Aware Stock Price Prediction with Transformer and LLM-Generated Formulaic Alpha (arXiv 2508.04975)
- From Attention to Profit: Quantitative Trading Strategy Based on Transformer (arXiv 2404.00424)
- Machine Learning Enhanced Multi-Factor Quantitative Trading (arXiv 2507.07107)
- A Controlled Comparison of Deep Learning for Multi-Horizon Financial Forecasting (arXiv 2603.16886)
- Exploring Microstructural Dynamics in Cryptocurrency LOBs (arXiv 2506.05764)
- TLOB: Transformer with Dual Attention for LOB Price Prediction (arXiv 2502.15757)
- Advances in Financial Machine Learning — Marcos Lopez de Prado (CPCV methodology)
MIT