Forecasting S&P 500 Using Technical Analysis, Macro Indicators and Machine Learning. a Hybrid Approach

Abstract

This thesis investigates the application of machine learning techniques to predict directional movements in the SPY ETF, an exchange-traded fund tracking the S&P 500 index. The primary objective is to evaluate whether combining macroeconomic indicators with technical analysis features can improve the predictive performance and financial profitability of classificationbased trading models. Traditional models often rely on either technical or fundamental indicators in isolation, but recent research suggests that hybrid approaches may offer better robustness and generalization in volatile financial environments.

A comprehensive dataset was compiled covering the period from February 2003 to June 2025, incorporating over 230 features, including macroeconomic metrics such as interest rates, unemployment figures, inflation data, and monetary aggregates, alongside technical indicators like Bollinger Bands, MACD, RSI, and TSI. Principal Component Analysis (PCA) was used to reduce dimensionality, while various machine learning algorithms—including K-Nearest Neighbors (KNN), Random Forests, Gradient Boosting, and Logistic Regression—were tested for their classification accuracy.

Labels were generated based on future SPY returns over multiple time horizons (e.g., 3, 7, 30, and 90 days), and categorized into three trading signals: BUY, NEUTRAL, and SELL. In addition to evaluating classification accuracy, the thesis places significant emphasis on backtesting strategy performance using key metrics such as cumulative return and Sharpe ratio. The findings reveal that models using only macroeconomic or only technical indicators tend to underperform, while hybrid models substantially improve both prediction quality and trading outcomes.

The best-performing configuration was achieved using the KNN classifier with 20 selected PCA components and a 90-day prediction horizon, yielding a classification accuracy of approximately 85% and a Sharpe ratio exceeding 1.2. These results support the hypothesis that integrated feature sets combined with proper model selection and threshold tuning can enhance financial forecasting in complex market conditions.