Quant Python Open Source

study-3-signals — 3-Signal Momentum Model

study-3-signals trains and tests predictive models on the relationship between three momentum signals (BAR, SPY, BND) and next-day SPXL returns, exploring feature engineering, lag analysis, and model selection.

View on GitHub ← All Projects

pip install -r requirements.txt
python train.py --model lgbm
python evaluate.py --report

Features

Predictive modeling framework using BAR, SPY, and BND signals to forecast SPXL (3x S&P 500 leveraged ETF) movements.

📡

3-Signal Framework

Uses Barchart (BAR), S&P 500 ETF (SPY), and Bond ETF (BND) as predictors for SPXL direction.

🔍

Feature Engineering

Lag windows, rolling statistics, and cross-signal ratios as model inputs.

🤖

Model Selection

Compares logistic regression, gradient boosting, and LSTM across walk-forward validation splits.

📊

Signal Attribution

SHAP values and permutation importance to understand which signals drive predictions.

Documentation & Architecture

This study uses historical data from 2021-01-01 to 2025-07-10 for tickers BAR, SPY, BND, SPXL.

A decision tree was trained on daily returns to predict if SPXL will increase by 5% or more the next day.

The patterns leading to 90% confidence are based on the decision tree rules where leaf nodes have >=90% probability for positive prediction.

Decision Tree Structure:

|--- SPY <= -0.00
|   |--- BAR <= 0.02
|   |   |--- BND <= -0.00
|   |   |   |--- class: 1
|   |   |--- BND >  -0.00
|   |   |   |--- class: 0
|   |--- BAR >  0.02
|   |   |--- BND <= 0.01
|   |   |   |--- class: 1
|   |   |--- BND >  0.01
|   |   |   |--- class: 0
|--- SPY >  -0.00
|   |--- BAR <= -0.01
|   |   |--- BND <= -0.00
|   |   |   |--- class: 0
|   |   |--- BND >  -0.00
|   |   |   |--- class: 1
|   |--- BAR >  -0.01
|   |   |--- SPY <= 0.02
|   |   |   |--- class: 0
|   |   |--- SPY >  0.02
|   |   |   |--- class: 1

The high confidence days are stored in the database table ‘high_confidence_days’ with columns for date (index), next_spxl_return, proba, and the signal returns (BAR, SPY, BND).

Table Field Descriptions

Ticker History Tables (BAR, SPY, BND, SPXL, SPXS)

Date: Trading date
Adj Close: Adjusted closing price
Close: Closing price
High: Highest price
Low: Lowest price
Open: Opening price
Volume: Trading volume

Confidence Tables (high_confidence_days, all_confidence_days, high_confidence_days_spxs, all_confidence_days_spxs)

index: Auto-generated ID
Date: Trading date
next_[spxl/spxs]_return: Actual return of target ticker the next day
proba: Model’s predicted probability of >=5% increase
BAR/SPY/BND: Daily returns of signal tickers

Decision Tree Summaries

SPXL Decision Tree

|--- SPY <= -0.00
|   |--- BAR <= 0.02
|   |   |--- BND <= -0.00
|   |   |   |--- class: 1
|   |   |--- BND >  -0.00
|   |   |   |--- class: 0
|   |--- BAR >  0.02
|   |   |--- BND <= 0.01
|   |   |   |--- class: 1
|   |   |--- BND >  0.01
|   |   |   |--- class: 0
|--- SPY >  -0.00
|   |--- BAR <= -0.01
|   |   |--- BND <= -0.00
|   |   |   |--- class: 0
|   |   |--- BND >  -0.00
|   |   |   |--- class: 1
|   |--- BAR >  -0.01
|   |   |--- SPY <= 0.02
|   |   |   |--- class: 0
|   |   |--- SPY >  0.02
|   |   |   |--- class: 1

SPXS Decision Tree

|--- BAR <= -0.01
|   |--- BAR <= -0.02
|   |   |--- BAR <= -0.02
|   |   |   |--- class: 0
|   |   |--- BAR >  -0.02
|   |   |   |--- class: 1
|   |--- BAR >  -0.02
|   |   |--- BAR <= -0.02
|   |   |   |--- class: 0
|   |   |--- BAR >  -0.02
|   |   |   |--- class: 0
|--- BAR >  -0.01
|   |--- SPY <= -0.00
|   |   |--- SPY <= -0.00
|   |   |   |--- class: 1
|   |   |--- SPY >  -0.00
|   |   |   |--- class: 1
|   |--- SPY >  -0.00
|   |   |--- SPY <= 0.01
|   |   |   |--- class: 0
|   |   |--- SPY >  0.01
|   |   |   |--- class: 1

Overlapping High-Confidence Days

There are no trading days where both SPXL and SPXS had high confidence (>80%) predictions simultaneously.