@Vu

July 2025

Executive Summary

This report covers a deep learning model designed to predict the opening gap percentage of the SPY ETF. The opening gap, or the price change between the previous day's close and the current day's open, is a significant source of daily volatility. By leveraging a hybrid CNN-LSTM architecture, the model analyzes pre-market and overnight 15-minute candlestick data to forecast the direction and magnitude of the gap.

The model was trained on data from February 2024 to June 2025. On the unseen test set, it achieved a Directional Accuracy of 57.1%, indicating a statistically significant edge in predicting whether the market will gap up or down. While the model's R-squared value is modest (0.13), its ability to correctly forecast the gap's direction presents a valuable signal for pre-market trading strategies.

1. Dataset Description

Data Sources and Timeframe

• Primary Dataset: SPY 15-minute candlestick data
• Training Period: February 1, 2024 to June 27, 2025.
• Total Observations: 22,441 hourly candles.
• Base Features: Datetime, Open, High, Low, Close, Volume.

Feature Engineering & Preprocessing

The model's predictive power is derived from a rich set of engineered features designed to capture market dynamics in the crucial pre-open hours.

• Target Variable: The Gap_Percentage, calculated as ((Current_Day_Open - Previous_Day_Close) / Previous_Day_Close) * 100.
• Input Feature Window: The model uses a variable-length sequence of 15-minute candles from the previous day's 12:45 PM PST close to the current day's 5:30 AM PST. This window is designed to capture overnight sentiment and early pre-market momentum.
• Technical Indicators: A comprehensive set of 84 technical indicators were computed for each 15-minute candle in the input window. These features provide signals on:
  • Momentum: RSI, Stochastic Oscillator, Williams %R.
  • Trend: MACD, ADX, Parabolic SAR.
  • Volume: On-Balance Volume (OBV), Chaikin Money Flow (CMF).
  • Volatility: Bollinger Bands, Average True Range (ATR).

2. Model Architecture and Design

Problem Formulation

The task is formulated as a many-to-one sequence-to-value regression problem. The model takes a sequence of feature vectors (one for each 15-minute candle) as input and outputs a single continuous value representing the predicted gap percentage.

Architecture Components