SparkTrade Daily Equity Signals (Long-Short, All-Weather Unified Model)

## **SparkTrade Daily Equity Signals (Long-Short, All-Weather Unified Model)**

Daily 0-100 prediction ranks on **2 000+** U.S. equities, trained on full 1999-2025 market data, yielding a steady, low-turnover reference sleeve. Sharpe ≈ 0.83

## **Detailed Description**

*SparkTrade All-Weather Equity Signals* is the house benchmark for users who need a **single, volatility-agnostic feed**. The model ingests every trading day since January 1999, blending price/volume, point-in-time fundamentals, micro-structure metrics, options skew, ETF flow, and sentiment. It outputs regime-neutral long & short scores that behave consistently through bull, bear, calm, and crisis tapes—handy for benchmarking, factor stacking, or low-turnover strategies.

Unlike regime-segmented sleeves that specialize in low-vol or high-vol markets, this model is **trained on the entire market of US Equities 1999-2025** , capturing the full spectrum of bull runs, recessions, melt-ups, and “black-swan” shocks. The outcome is a **single, volatility-agnostic signal engine** that delivers dependable alpha without the need for external regime switches or human toggles.

## How It Works

**Multi-factor ensemble** Gradient-boosted trees and random forests ingest price/volume trends, point-in-time fundamentals, micro-structure metrics, options skew, ETF flow, and real-time sentiment.

**Purged k-fold cross-validation** Eliminates look-ahead bias and stabilizes out-of-sample results.

**Transparency** No RNNs or opaque deep nets—feature importance and rule paths are auditable for risk and compliance teams.

**Orthogonality** – Adjusted R² vs Fama-French 3-factor ≈ 0.11, adding fresh alpha vectors.

## **Coverage & logistics** 

## Universe

2000+ U.S. tickers (≥ $300 M float-adjusted cap).

## History

Jan-1999 → present (T-1 close).

## Delivery

Snowflake Secure Share posted **by 20:00 ET** each trading day.

## Data Dictionary

NAME TYPE DESCRIPTION

DATE DATE Date of prediction

TICKER STRING Stock ticker symbol (e.g., AAPL)

LONG_INDEX_SCORE INTEGER 0–100 scaled score (higher = stronger long)

SHORT_INDEX_SCORE INTEGER 0–100 scaled score (higher = stronger short)

LONG_RAW_SCORE FLOAT 0 to 1.0 unscaled signal before index conversion

SHORT_RAW_SCORE FLOAT 0 to 1.0 unscaled signal before index conversion

VOLATILITY_REGIME STRING "all weather"

MODEL_NAME STRING Source model name (e.g., "Low Vol", "AutoSwitch")

MODEL_VERSION STRING Version or release ID (for audit/debug)

## **Update Frequency**

Trading days, file posted **by 20:00 ET**

## **SparkTrade 每日股票多空全天候统一模型信号集**

针对2000余只美国上市股票生成0-100分预测排名，训练数据集覆盖1999年至2025年完整市场周期，可生成持续稳定、低换手率的参考投资组合，夏普比率（Sharpe Ratio）≈0.83。

## **详细说明**

*SparkTrade 全天候股票信号集* 是面向需要**单一、波动无关性输入源**用户的内部基准模型。该模型纳入1999年1月以来的全部交易日数据，融合量价、时点式基本面、微观结构指标、期权偏度、ETF（Exchange Traded Fund）资金流与市场情绪等多类因子，输出对市场状态中立的多空评分，在牛市、熊市、震荡市与危机行情中均保持一致表现，可用于基准测试、因子堆叠或低换手率策略开发。

不同于针对低波动或高波动市场细分场景的分状态组合，本模型**基于1999-2025年全市场美国股票数据训练**，完整覆盖了牛市周期、经济衰退、行情过热与"黑天鹅"冲击等各类市场环境。最终产出**单一、与波动无关的信号引擎**，无需外部状态切换或人工调整即可提供可靠的超额收益（Alpha）。

## **工作原理**

**多因子集成模型**：采用梯度提升树与随机森林，整合量价趋势、时点式基本面、微观结构指标、期权偏度、ETF资金流与实时市场情绪等因子。
**净化k折交叉验证（Purged k-fold cross-validation）**：消除前瞻偏差，提升样本外结果的稳定性。
**透明度**：未使用循环神经网络（RNN）或不透明的深度神经网络，特征重要性与规则路径可被审计，适配风控与合规团队需求。
**正交性**：与法玛-弗伦奇三因子模型（Fama-French 3-factor Model）的调整后决定系数（Adjusted R²）≈0.11，可提供全新的超额收益因子向量。

## **覆盖范围与交付细节**

### 标的池
2000余只美国上市股票（浮盈调整市值≥3亿美元）。
### 历史数据周期
1999年1月至今（数据基于前一交易日收盘价）。
### 交付方式
通过Snowflake安全共享（Snowflake Secure Share）于每个交易日**美国东部时间20:00前**发布。

## **数据字典**

| 字段名（NAME） | 数据类型（TYPE） | 描述（DESCRIPTION） |
| :--- | :--- | :--- |
| DATE | DATE | 预测日期 |
| TICKER | STRING | 股票代码（例如：AAPL） |
| LONG_INDEX_SCORE | INTEGER | 0–100分标准化评分（分值越高，做多信号越强） |
| SHORT_INDEX_SCORE | INTEGER | 0–100分标准化评分（分值越高，做空信号越强） |
| LONG_RAW_SCORE | FLOAT | 标准化前的原始做多信号，取值范围0至1.0 |
| SHORT_RAW_SCORE | FLOAT | 标准化前的原始做空信号，取值范围0至1.0 |
| VOLATILITY_REGIME | STRING | 固定为"all weather"（全天候） |
| MODEL_NAME | STRING | 源模型名称（例如："Low Vol"、"AutoSwitch"） |
| MODEL_VERSION | STRING | 版本或发布ID（用于审计与调试） |

## **更新频率**

每个交易日，文件于**美国东部时间20:00前**发布。