George Shengelia

All projects

Project 01 — Machine learning · Forecasting

S&P 500 Return Prediction

Can a model built on price, technical and macroeconomic features beat buy-and-hold on a risk-adjusted basis? A forecasting project with walk-forward backtesting.

Role
[Solo / team project]
Timeline
[Term · Year]
Stack
Python · pandas · scikit-learn · XGBoost
Links

Problem

Can we predict short-term S&P 500 returns using historical data and macro/market indicators? The goal is to build a practical forecasting model and evaluate whether it adds value over simple benchmarks.

Data

Three groups of inputs, aligned to a daily trading calendar:

Group Features Source
Price Daily index close → returns [Source]
Technical Lagged returns, moving averages, RSI Derived
Macro VIX, PMI — forward-filled to daily [Source]
Placeholder for a data figure
Figure 1 — Images drop into the write-up like this, with a caption underneath.

Approach

  1. Construct a feature set from price, technical, and macroeconomic data.
  2. Train and compare several models — Ridge, Random Forest, and XGBoost.
  3. Evaluate with a walk-forward backtest and risk-adjusted metrics such as the Sharpe ratio.

Results

Over the walk-forward test period, the model outperformed a buy-and-hold position in the index on a risk-adjusted basis.

Line chart of cumulative returns, model versus S&P 500 buy-and-hold, 2020 to 2024
Figure 2 — Growth of $1, walk-forward test period 2020–2024.
Annual return14.2%
Volatility16.8%
Sharpe ratio0.85
Max drawdown−18.6%

Key takeaways

  • The model outperforms a buy-and-hold benchmark on a risk-adjusted basis.
  • Feature engineering and regularization are important for stable performance.

What I’d do next

[A short, reflective paragraph — e.g. adding transaction costs, testing regime-aware features — so the page reads like a lesson, not just a result.]