Vu Nguyen

Jun 2025

Executive Summary

This report presents a neural network approach for predicting SPY (S&P 500 ETF) closing prices using 1-hour candlestick data. The model employs an LSTM-Attention architecture to forecast the next 5 candles' closing prices, achieving an average MAE of 20.0 across prediction steps with MAPE ranging from 3.63% to 3.75%.

Dataset Description

Data Sources and Timeframe

• Primary Dataset: SPY 1-hour candlestick data
• Training Period: January 2, 2013 09:30 EST to February 18, 2025 15:30 EST
• Total Observations: 21,303 hourly candles
• Base Features: Date, Open, High, Low, Close, Volume
• Enhanced Features: Popular technical indicators (RSI, Moving Averages, MACD, etc.)
• See Appendix for dataset head.

Data Preprocessing

The dataset includes both raw OHLCV data and computed technical indicators to capture momentum, trend, and volatility signals that are commonly used in financial markets.

Model Architecture and Design

Problem Formulation

• Task Type: Many-to-many sequence prediction

source

• Input Sequence Length: 65 candles (approximately 10 trading days)